NVIDIA Virtual GPU Software v13.0 through 13.12
Version 13.12 Download PDF

Virtual GPU Software User Guide

Virtual GPU Software User Guide

Documentation for administrators that explains how to install and configure NVIDIA Virtual GPU manager, configure virtual GPU software in pass-through mode, and install drivers on guest operating systems.

NVIDIA vGPU software is a graphics virtualization platform that provides virtual machines (VMs) access to NVIDIA GPU technology.

1.1. How NVIDIA vGPU Software Is Used

NVIDIA vGPU software can be used in several ways.

1.1.1. NVIDIA vGPU

NVIDIA Virtual GPU (vGPU) enables multiple virtual machines (VMs) to have simultaneous, direct access to a single physical GPU, using the same NVIDIA graphics drivers that are deployed on non-virtualized operating systems. By doing this, NVIDIA vGPU provides VMs with unparalleled graphics performance, compute performance, and application compatibility, together with the cost-effectiveness and scalability brought about by sharing a GPU among multiple workloads.

For more information, see Installing and Configuring NVIDIA Virtual GPU Manager.

1.1.2. GPU Pass-Through

In GPU pass-through mode, an entire physical GPU is directly assigned to one VM, bypassing the NVIDIA Virtual GPU Manager. In this mode of operation, the GPU is accessed exclusively by the NVIDIA driver running in the VM to which it is assigned. The GPU is not shared among VMs.

For more information, see Using GPU Pass-Through.

1.1.3. Bare-Metal Deployment

In a bare-metal deployment, you can use NVIDIA vGPU software graphics drivers with vWS and vApps licenses to deliver remote virtual desktops and applications. If you intend to use Tesla boards without a hypervisor for this purpose, use NVIDIA vGPU software graphics drivers, not other NVIDIA drivers.

To use NVIDIA vGPU software drivers for a bare-metal deployment, complete these tasks:

  1. Install the driver on the physical host.

    For instructions, see Installing the NVIDIA vGPU Software Graphics Driver.

  2. License any NVIDIA vGPU software that you are using.

    For instructions, see Virtual GPU Client Licensing User Guide.

  3. Configure the platform for remote access.

    To use graphics features with Tesla GPUs, you must use a supported remoting solution, for example, RemoteFX, Citrix Virtual Apps and Desktops, VNC, or similar technology.

  4. Use the display settings feature of the host OS to configure the Tesla GPU as the primary display.

    NVIDIA Tesla generally operates as a secondary device on bare-metal platforms.

  5. If the system has multiple display adapters, disable display devices connected through adapters that are not from NVIDIA.

    You can use the display settings feature of the host OS or the remoting solution for this purpose. On NVIDIA GPUs, including Tesla GPUs, a default display device is enabled.

    Users can launch applications that require NVIDIA GPU technology for enhanced user experience only after displays that are driven by NVIDIA adapters are enabled.

1.2. Primary Display Adapter Requirements for NVIDIA vGPU Software Deployments

The GPU that is set as the primary display adapter cannot be used for NVIDIA vGPU deployments or GPU pass through deployments. The primary display is the boot display of the hypervisor host, which displays SBIOS console messages and then boot of the OS or hypervisor.

Any GPU that is being used for NVIDIA vGPU deployments or GPU pass through deployments must be set as a secondary display adapter.

Note:

Citrix Hypervisor provides a specific setting to allow the primary display adapter to be used for GPU pass through deployments.

citrix-hypervisor-gpu-settings-window.png

Only the following GPUs are supported as the primary display adapter:

  • Tesla M6
  • Quadro RTX 6000
  • Quadro RTX 8000

All other GPUs that support NVIDIA vGPU software cannot function as the primary display adapter because they are 3D controllers, not VGA devices.

If the hypervisor host does not have an extra graphics adapter, consider installing a low-end display adapter to be used as the primary display adapter. If necessary, ensure that the primary display adapter is set correctly in the BIOS options of the hypervisor host.

1.3. NVIDIA vGPU Software Features

NVIDIA vGPU software includes vWS, vCS, vPC, and vApps.

1.3.1. GPU Instance Support on NVIDIA vGPU Software

NVIDIA vGPU software supports GPU instances on GPUs that support the Multi-Instance GPU (MIG) feature in NVIDIA vGPU and GPU pass through deployments. MIG enables a physical GPU to be securely partitioned into multiple separate GPU instances, providing multiple users with separate GPU resources to accelerate their applications.

In addition to providing all the benefits of MIG, NVIDIA vGPU software adds virtual machine security and management for workloads. Single Root I/O Virtualization (SR-IOV) virtual functions enable full IOMMU protection for the virtual machines that are configured with vGPUs.

Figure 1 shows a GPU that is split into three GPU instances of different sizes, with each instance mapped to one vGPU. Although each GPU instance is managed by the hypervisor host and is mapped to one vGPU, each virtual machine can further subdivide the compute resources into smaller compute instances and run multiple containers on top of them in parallel, even within each vGPU.

Figure 1. GPU Instances Configured with NVIDIA vGPU

mig-with-vgpu-overview.png

NVIDIA vGPU software supports a single-slice MIG-backed vGPU with DEC, JPG, and OFA support. Only one MIG-backed vGPU with DEC, JPG, and OFA support can reside on a GPU. The instance can be placed identically to a single-slice instance without DEC, JPG, and OFA support.

Not all hypervisors support GPU instances in NVIDIA vGPU deployments. To determine if your chosen hypervisor supports GPU instances in NVIDIA vGPU deployments, consult the release notes for your hypervisor at NVIDIA Virtual GPU Software Documentation.

NVIDIA vGPU software supports GPU instances only with NVIDIA Virtual Compute Server and Linux guest operating systems.

To support GPU instances with NVIDIA vGPU, a GPU must be configured with MIG mode enabled and GPU instances must be created and configured on the physical GPU. For more information, see Configuring a GPU for MIG-Backed vGPUs. For general information about the MIG feature, see: NVIDIA Multi-Instance GPU User Guide.

1.3.2. API Support on NVIDIA vGPU

NVIDIA vGPU includes support for the following APIs:

  • Open Computing Language (OpenCL™ software) 3.0
  • OpenGL® 4.6
  • Vulkan® 1.2
  • DirectX 11
  • DirectX 12 (Windows 10)
  • Direct2D
  • DirectX Video Acceleration (DXVA)
  • NVIDIA® CUDA® 11.4
  • NVIDIA vGPU software SDK (remote graphics acceleration)
  • NVIDIA RTX (on GPUs based on the NVIDIA Volta graphic architecture and later architectures)

Note:

These APIs are backwards compatible. Older versions of the API are also supported.

1.3.3. NVIDIA CUDA Toolkit and OpenCL Support on NVIDIA vGPU Software

NVIDIA CUDA Toolkit and OpenCL are supported with NVIDIA vGPU only on a subset of vGPU types and supported GPUs.

For more information about NVIDIA CUDA Toolkit, see CUDA Toolkit 11.4 Documentation.

Note:

If you are using NVIDIA vGPU software with CUDA on Linux, avoid conflicting installation methods by installing CUDA from a distribution-independent runfile package. Do not install CUDA from a distribution-specific RPM or Deb package.

To ensure that the NVIDIA vGPU software graphics driver is not overwritten when CUDA is installed, deselect the CUDA driver when selecting the CUDA components to install.

For more information, see NVIDIA CUDA Installation Guide for Linux.

OpenCL and CUDA Application Support

OpenCL and CUDA applications are supported on the following NVIDIA vGPU types:

  • The 8Q vGPU type on Tesla M6, Tesla M10, and Tesla M60 GPUs
  • All Q-series vGPU types on the following GPUs:
    • NVIDIA A10
    • NVIDIA A16
    • NVIDIA A40
    • NVIDIA RTX A5000
    • NVIDIA RTX A6000
    • Tesla P4
    • Tesla P6
    • Tesla P40
    • Tesla P100 SXM2 16 GB
    • Tesla P100 PCIe 16 GB
    • Tesla P100 PCIe 12 GB
    • Tesla V100 SXM2
    • Tesla V100 SXM2 32GB
    • Tesla V100 PCIe
    • Tesla V100 PCIe 32GB
    • Tesla V100S PCIe 32GB
    • Tesla V100 FHHL
    • Tesla T4
    • Quadro RTX 6000
    • Quadro RTX 6000 passive
    • Quadro RTX 8000
    • Quadro RTX 8000 passive
  • All C-series vGPU types

NVIDIA CUDA Toolkit Development Tool Support

NVIDIA vGPU supports the following NVIDIA CUDA Toolkit development tools on some GPUs:

  • Debuggers:
    • CUDA-GDB
    • Compute Sanitizer
  • Profilers (supported on MIG-backed vGPUs since 13.1):
    • The Activity, Callback, and Profiling APIs of the CUDA Profiling Tools Interface (CUPTI)

      Other CUPTI APIs, such as the Event and Metric APIs, are not supported.

    • NVIDIA Nsight™ Compute
    • NVIDIA Nsight Systems
    • NVIDIA Nsight plugin
    • NVIDIA Nsight Visual Studio plugin

    Other CUDA profilers, such as nvprof and NVIDIA Visual Profiler, are not supported.

These tools are supported only in Linux guest VMs.

When unified memory is enabled for a VM, NVIDIA CUDA Toolkit profilers are disabled.

Note:

By default, NVIDIA CUDA Toolkit development tools are disabled on NVIDIA vGPU. If used, you must enable NVIDIA CUDA Toolkit development tools individually for each VM that requires them by setting vGPU plugin parameters. For instructions, see Enabling NVIDIA CUDA Toolkit Development Tools for NVIDIA vGPU.

The following table lists the GPUs on which NVIDIA vGPU supports these debuggers and profilers.
GPU vGPU Mode Debuggers Profilers
NVIDIA A10 Time-sliced

13.0 only: -

Since 13.1:

NVIDIA A16 Time-sliced

13.0 only: -

Since 13.1:

NVIDIA A30 Time-sliced
MIG-backed

13.0 only: -

Since 13.1:

NVIDIA A40 Time-sliced

13.0 only: -

Since 13.1:

NVIDIA A100 HGX 40GB Time-sliced
MIG-backed

13.0 only: -

Since 13.1:

NVIDIA A100 PCIe 40GB Time-sliced
MIG-backed

13.0 only: -

Since 13.1:

NVIDIA A100 HGX 80GB Time-sliced
MIG-backed

13.0 only: -

Since 13.1:

NVIDIA A100 PCIe 80GB Time-sliced
MIG-backed

13.0 only: -

Since 13.1:

NVIDIA RTX A5000 Time-sliced

13.0 only: -

Since 13.1:

NVIDIA RTX A6000 Time-sliced

13.0 only: -

Since 13.1:

Tesla T4 Time-sliced
Quadro RTX 6000 Time-sliced
Quadro RTX 6000 passive Time-sliced
Quadro RTX 8000 Time-sliced
Quadro RTX 8000 passive Time-sliced
Tesla V100 SXM2 Time-sliced
Tesla V100 SXM2 32GB Time-sliced
Tesla V100 PCIe Time-sliced
Tesla V100 PCIe 32GB Time-sliced
Tesla V100S PCIe 32GB Time-sliced
Tesla V100 FHHL Time-sliced

✓ Feature is supported

- Feature is not supported

Supported NVIDIA CUDA Toolkit Features

NVIDIA vGPU supports the following NVIDIA CUDA Toolkit features if the vGPU type, physical GPU, and the hypervisor software version support the feature:

  • Error-correcting code (ECC) memory
  • Peer-to-peer CUDA transfers over NVLink
    Note:

    To determine the NVLink topology between physical GPUs in a host or vGPUs assigned to a VM, run the following command from the host or VM:

    Copy
    Copied!
                

    $ nvidia-smi topo -m

  • GPUDirect® technology remote direct memory access (RDMA)
  • Unified Memory
    Note:

    Unified memory is disabled by default. If used, you must enable unified memory individually for each vGPU that requires it by setting a vGPU plugin parameter. For instructions, see Enabling Unified Memory for a vGPU.

Dynamic page retirement is supported for all vGPU types on physical GPUs that support ECC memory, even if ECC memory is disabled on the physical GPU.

NVIDIA CUDA Toolkit Features Not Supported by NVIDIA vGPU

NVIDIA vGPU does not support the following NVIDIA CUDA Toolkit features:

  • GPUDirect technology storage
  • GPU context switch trace
  • NVIDIA Nsight Graphics

Note:

These features, except GPUDirect technology storage, are supported in GPU pass-through mode and in bare-metal deployments.

1.3.4. Additional vWS Features

In addition to the features of vPC and vApps, vWS provides the following features:

  • Workstation-specific graphics features and accelerations
  • Certified drivers for professional applications
  • GPU pass through for workstation or professional 3D graphics

    In pass-through mode, vWS supports multiple virtual display heads at resolutions up to 8K and flexible virtual display resolutions based on the number of available pixels. For details, see Display Resolutions for Physical GPUs.

  • 10-bit color for Windows users. (HDR/10-bit color is not currently supported on Linux, NvFBC capture is supported but deprecated.)

1.3.5. NVIDIA GPU Cloud (NGC) Containers Support on NVIDIA vGPU Software

NVIDIA vGPU software supports NGC containers in NVIDIA vGPU and GPU pass-through deployments on all supported hypervisors.

In NVIDIA vGPU deployments, the following vGPU types are supported only on GPUs based on NVIDIA GPU architectures after the Maxwell architecture:

  • All Q-series vGPU types
  • All C-series vGPU types

In GPU pass-through deployments, all GPUs based on NVIDIA GPU architectures after the NVIDIA Maxwell™ architecture that support NVIDIA vGPU software are supported.

NVIDIA vGPU software supports NGC containers on any guest operating system listed in Supported Platforms - NVIDIA Container Toolkit that is also supported by NVIDIA vGPU software.

For more information about setting up NVIDIA vGPU software for use with NGC containers, see Using NGC with NVIDIA Virtual GPU Software Setup Guide.

1.3.6. NVIDIA GPU Operator Support

NVIDIA GPU Operator simplifies the deployment of NVIDIA vGPU software on software container platforms that are managed by the Kubernetes container orchestration engine. It automates the installation and update of NVIDIA vGPU software graphics drivers for container platforms running in guest VMs that are configured with NVIDIA vGPU.

NVIDIA GPU Operator uses a driver catalog published with the NVIDIA vGPU software graphics drivers to determine automatically the NVIDIA vGPU software graphics driver version that is compatible with a platform's Virtual GPU Manager.

Any drivers to be installed by NVIDIA GPU Operator must be downloaded from the NVIDIA Licensing Portal to a local computer. Automated access to the NVIDIA Licensing Portal by NVIDIA GPU Operator is not supported.

NVIDIA GPU Operator supports automated configuration of NVIDIA vGPU software and provides telemetry support through DCGM Exporter running in a guest VM.

NVIDIA GPU Operator is supported only on specific combinations of hypervisor software release, container platform, vGPU type, and guest OS release. To determine if your configuration supports NVIDIA GPU Operator with NVIDIA vGPU deployments, consult the release notes for your chosen hypervisor at NVIDIA Virtual GPU Software Documentation.

For more information, see NVIDIA GPU Operator Overview on the NVIDIA documentation portal.

1.4. How this Guide Is Organized

Virtual GPU Software User Guide is organized as follows:

The process for installing and configuring NVIDIA Virtual GPU Manager depends on the hypervisor that you are using. After you complete this process, you can install the display drivers for your guest OS and license any NVIDIA vGPU software licensed products that you are using.

2.1. About NVIDIA Virtual GPUs

2.1.1. NVIDIA vGPU Architecture

The high-level architecture of NVIDIA vGPU is illustrated in Figure 2. Under the control of the NVIDIA Virtual GPU Manager running under the hypervisor, NVIDIA physical GPUs are capable of supporting multiple virtual GPU devices (vGPUs) that can be assigned directly to guest VMs.

Guest VMs use NVIDIA vGPUs in the same manner as a physical GPU that has been passed through by the hypervisor: an NVIDIA driver loaded in the guest VM provides direct access to the GPU for performance-critical fast paths, and a paravirtualized interface to the NVIDIA Virtual GPU Manager is used for non-performant management operations.

Figure 2. NVIDIA vGPU System Architecture

architecture-grid-vgpu-overview.png

Each NVIDIA vGPU is analogous to a conventional GPU, having a fixed amount of GPU framebuffer, and one or more virtual display outputs or "heads". The vGPU’s framebuffer is allocated out of the physical GPU’s framebuffer at the time the vGPU is created, and the vGPU retains exclusive use of that framebuffer until it is destroyed.

Depending on the physical GPU, different types of vGPU can be created on the vGPU:

  • On all GPUs that support NVIDIA vGPU software, time-sliced vGPUs can be created.
  • Additionally, on GPUs that support the Multi-Instance GPU (MIG) feature, MIG-backed vGPUs can be created. The MIG feature is introduced on GPUs that are based on the NVIDIA Ampere GPU architecture.

2.1.1.1. Time-Sliced NVIDIA vGPU Internal Architecture

A time-sliced vGPU is a vGPU that resides on a physical GPU that is not partitioned into multiple GPU instances. All time-sliced vGPUs resident on a GPU share access to the GPU’s engines including the graphics (3D), video decode, and video encode engines.

In a time-sliced vGPU, processes that run on the vGPU are scheduled to run in series. Each vGPU waits while other processes run on other vGPUs. While processes are running on a vGPU, the vGPU has exclusive use of the GPU's engines. You can change the default scheduling behavior as explained in Changing Scheduling Behavior for Time-Sliced vGPUs.

Figure 3. Time-Sliced NVIDIA vGPU Internal Architecture

architecture-grid-vgpu-internal.png

2.1.1.2. MIG-Backed NVIDIA vGPU Internal Architecture

A MIG-backed vGPU is a vGPU that resides on a GPU instance in a MIG-capable physical GPU. Each MIG-backed vGPU resident on a GPU has exclusive access to the GPU instance’s engines, including the compute and video decode engines.

In a MIG-backed vGPU, processes that run on the vGPU run in parallel with processes running on other vGPUs on the GPU. Process run on all vGPUs resident on a physical GPU simultaneously.

Figure 4. MIG-Backed NVIDIA vGPU Internal Architecture

architecture-grid-vgpu-mig-backed-internal.png

2.1.2. About Virtual GPU Types

The number of physical GPUs that a board has depends on the board. Each physical GPU can support several different types of virtual GPU (vGPU). vGPU types have a fixed amount of frame buffer, number of supported display heads, and maximum resolutions1. They are grouped into different series according to the different classes of workload for which they are optimized. Each series is identified by the last letter of the vGPU type name.

Series Optimal Workload
Q-series Virtual workstations for creative and technical professionals who require the performance and features of Quadro technology
C-series Compute-intensive server workloads, such as artificial intelligence (AI), deep learning, or high-performance computing (HPC)2, 3
B-series Virtual desktops for business professionals and knowledge workers
A-series App streaming or session-based solutions for virtual applications users6

The number after the board type in the vGPU type name denotes the amount of frame buffer that is allocated to a vGPU of that type. For example, a vGPU of type A16-4C is allocated 4096 Mbytes of frame buffer on an NVIDIA A16 board.

Due to their differing resource requirements, the maximum number of vGPUs that can be created simultaneously on a physical GPU varies according to the vGPU type. For example, an NVDIA A16 board can support up to 4 A16-4C vGPUs on each of its two physical GPUs, for a total of 16 vGPUs, but only 2 A16-8C vGPUs, for a total of 8 vGPUs. When enabled, the frame-rate limiter (FRL) limits the maximum frame rate in frames per second (FPS) for a vGPU as follows:

  • For B-series vGPUs, the maximum frame rate is 45 FPS.
  • For Q-series, C-series, and A-series vGPUs, the maximum frame rate is 60 FPS.

By default, the FRL is enabled for all GPUs. The FRL is disabled when the vGPU scheduling behavior is changed from the default best-effort scheduler on GPUs that support alternative vGPU schedulers. For details, see Changing Scheduling Behavior for Time-Sliced vGPUs. On vGPUs that use the best-effort scheduler, the FRL can be disabled as explained in the release notes for your chosen hypervisor at NVIDIA Virtual GPU Software Documentation.

Note:

NVIDIA vGPU is a licensed product on all supported GPU boards. A software license is required to enable all vGPU features within the guest VM. The type of license required depends on the vGPU type.

  • Q-series vGPU types require a vWS license.
  • C-series vGPU types require an NVIDIA Virtual Compute Server (vCS) license but can also be used with a vWS license.
  • B-series vGPU types require a vPC license but can also be used with a vWS license.
  • A-series vGPU types require a vApps license.

For details of the virtual GPU types available from each supported GPU, see Virtual GPU Types for Supported GPUs.

2.1.3. Virtual Display Resolutions for Q-series and B-series vGPUs

Instead of a fixed maximum resolution per display, Q-series and B-series vGPUs support a maximum combined resolution based on the number of available pixels, which is determined by their frame buffer size. You can choose between using a small number of high resolution displays or a larger number of lower resolution displays with these vGPUs.

The number of virtual displays that you can use depends on a combination of the following factors:

  • Virtual GPU series
  • GPU architecture
  • vGPU frame buffer size
  • Display resolution
Note:

You cannot use more than the maximum number of displays that a vGPU supports even if the combined resolution of the displays is less than the number of available pixels from the vGPU. For example, because -0Q and -0B vGPUs support a maximum of only two displays, you cannot use four 1280×1024 displays with these vGPUs even though the combined resolution of the displays (6220800) is less than the number of available pixels from these vGPUs (8192000).

Various factors affect the consumption of the GPU frame buffer, which can impact the user experience. These factors include and are not limited to the number of displays, display resolution, workload and applications deployed, remoting solution, and guest OS. The ability of a vGPU to drive a certain combination of displays does not guarantee that enough frame buffer remains free for all applications to run. If applications run out of frame buffer, consider changing your setup in one of the following ways:

  • Switching to a vGPU type with more frame buffer
  • Using fewer displays
  • Using lower resolution displays

The maximum number of displays per vGPU listed in Virtual GPU Types for Supported GPUs is based on a configuration in which all displays have the same resolution. For examples of configurations with a mixture of display resolutions, see Mixed Display Configurations for B-Series and Q-Series vGPUs.

2.1.4. Valid Virtual GPU Configurations on a Single GPU

Valid vGPU configurations on a single GPU depend on whether the vGPUs are time sliced or, on GPUs that support MIG, are MIG-backed.

2.1.4.1. Valid Time-Sliced Virtual GPU Configurations on a Single GPU

This release of NVIDIA vGPU supports only homogeneous time-sliced virtual GPUs. At any given time, the virtual GPUs resident on a single physical GPU must be all of the same type.

However, this restriction doesn’t extend across physical GPUs on the same card. Different physical GPUs on the same card may host different types of virtual GPU at the same time, provided that the vGPU types on any one physical GPU are the same. For example, a Tesla M60 card has two physical GPUs, and can support several types of virtual GPU. Figure 5 shows the following examples of valid and invalid virtual GPU configurations on Tesla M60:

  • A valid configuration with M60-2Q vGPUs on GPU 0 and M60-4Q vGPUs on GPU 1
  • A valid configuration with M60-1B vGPUs on GPU 0 and M60-2Q vGPUs on GPU 1
  • An invalid configuration with mixed vGPU types on GPU 0

Figure 5. Example Time-Sliced vGPU Configurations on Tesla M60

sample-vgpu-configurations-grid-2gpus-on-card.png

2.1.4.2. Valid MIG-Backed Virtual GPU Configurations on a Single GPU

This release of NVIDIA vGPU supports both homogeneous and mixed MIG-backed virtual GPUs based on the underlying GPU instance configuration.

For example, an NVIDIA A100 PCIe 40GB card has one physical GPU, and can support several types of virtual GPU. Figure 6 shows the following examples of valid homogeneous and mixed MIG-backed virtual GPU configurations on NVIDIA A100 PCIe 40GB.

  • A valid homogeneous configuration with 3 A100-2-10C vGPUs on 3 MIG.2g.10b GPU instances
  • A valid homogeneous configuration with 2 A100-3-20C vGPUs on 3 MIG.3g.20b GPU instances
  • A valid mixed configuration with 1 A100-4-20C vGPU on a MIG.4g.20b GPU instance, 1 A100-2-10C vGPU on a MIG.2.10b GPU instance, and 1 A100-1-5C vGPU on a MIG.1g.5b instance

Figure 6. Example MIG-Backed vGPU Configurations on NVIDIA A100 PCIe 40GB

mixed-vgpu-configurations.png

2.1.5. Guest VM Support

NVIDIA vGPU supports Windows and Linux guest VM operating systems. The supported vGPU types depend on the guest VM OS.

For details of the supported releases of Windows and Linux, and for further information on supported configurations, see the driver release notes for your hypervisor at NVIDIA Virtual GPU Software Documentation.

2.1.5.1. Windows Guest VM Support

Windows guest VMs are supported only on Q-series, B-series, and A-series NVIDIA vGPU types. They are not supported on C-series NVIDIA vGPU types.

2.1.5.2. Linux Guest VM support

64-bit Linux guest VMs are supported only on Q-series, C-series, and B-series NVIDIA vGPU types. They are not supported on A-series NVIDIA vGPU types.

2.2. Prerequisites for Using NVIDIA vGPU

Before proceeding, ensure that these prerequisites are met:

  • You have a server platform that is capable of hosting your chosen hypervisor and NVIDIA GPUs that support NVIDIA vGPU software.
  • One or more NVIDIA GPUs that support NVIDIA vGPU software is installed in your server platform.
  • If you are using GPUs based on the NVIDIA Ampere architecture, the following BIOS settings are enabled on your server platform:

    • VT-D/IOMMU
    • SR-IOV
    • Alternative Routing ID Interpretation (ARI)
  • You have downloaded the NVIDIA vGPU software package for your chosen hypervisor, which consists of the following software:
    • NVIDIA Virtual GPU Manager for your hypervisor
    • NVIDIA vGPU software graphics drivers for supported guest operating systems
  • The following software is installed according to the instructions in the software vendor's documentation:
    • Your chosen hypervisor, for example, Citrix Hypervisor, Red Hat Enterprise Linux KVM, Red Hat Virtualization (RHV), or VMware vSphere Hypervisor (ESXi)
    • The software for managing your chosen hypervisor, for example, Citrix XenCenter management GUI, or VMware vCenter Server
    • The virtual desktop software that you will use with virtual machines (VMs) running NVIDIA Virtual GPU, for example, Citrix Virtual Apps and Desktops, or VMware Horizon
    Note:

    If you are using VMware vSphere Hypervisor (ESXi), ensure that the ESXi host on which you will configure a VM with NVIDIA vGPU is not a member of a fully automated VMware Distributed Resource Scheduler (DRS) cluster. For more information, see Installing and Configuring the NVIDIA Virtual GPU Manager for VMware vSphere.

  • A VM to be enabled with vGPU is created.
    Note:

    If the VM uses UEFI boot and you plan to install a Linux guest OS in the VM, ensure that secure boot is disabled.

    All hypervisors covered in this guide support multiple vGPUs in a VM.

  • Your chosen guest OS is installed in the VM.

For information about supported hardware and software, and any known issues for this release of NVIDIA vGPU software, refer to the Release Notes for your chosen hypervisor:

2.3. Switching the Mode of a GPU that Supports Multiple Display Modes

Some GPUs support displayless and display-enabled modes but must be used in NVIDIA vGPU software deployments in displayless mode.

The GPUs listed in the following table support multiple display modes. As shown in the table, some GPUs are supplied from the factory in displayless mode, but other GPUs are supplied in a display-enabled mode.

GPU Mode as Supplied from the Factory
NVIDIA A40 Displayless
NVIDIA RTX A5000 Display enabled
NVIDIA RTX A6000 Display enabled

A GPU that is supplied from the factory in displayless mode, such as the NVIDIA A40 GPU, might be in a display-enabled mode if its mode has previously been changed.

To change the mode of a GPU that supports multiple display modes, use the displaymodeselector tool, which you can request from the NVIDIA Display Mode Selector Tool page on the NVIDIA Developer website.

Note:

Only the following GPUs support the displaymodeselector tool:

  • NVIDIA A40
  • NVIDIA RTX A5000
  • NVIDIA RTX A6000

Other GPUs that support NVIDIA vGPU software do not support the displaymodeselector tool and, unless otherwise stated, do not require display mode switching.

2.4. Switching the Mode of a Tesla M60 or M6 GPU

Tesla M60 and M6 GPUs support compute mode and graphics mode. NVIDIA vGPU requires GPUs that support both modes to operate in graphics mode.

Recent Tesla M60 GPUs and M6 GPUs are supplied in graphics mode. However, your GPU might be in compute mode if it is an older Tesla M60 GPU or M6 GPU or if its mode has previously been changed.

To configure the mode of Tesla M60 and M6 GPUs, use the gpumodeswitch tool provided with NVIDIA vGPU software releases. If you are unsure which mode your GPU is in, use the gpumodeswitch tool to find out the mode.

Note:

Only Tesla M60 and M6 GPUs support the gpumodeswitch tool. Other GPUs that support NVIDIA vGPU do not support the gpumodeswitch tool and, except as stated in Switching the Mode of a GPU that Supports Multiple Display Modes, do not require mode switching.

Even in compute mode, Tesla M60 and M6 GPUs do not support NVIDIA Virtual Compute Server vGPU types. Furthermore, vCS is not supported on any GPU on Citrix Hypervisor.

For more information, refer to gpumodeswitch User Guide.

2.5. Installing and Configuring the NVIDIA Virtual GPU Manager for Citrix Hypervisor

The following topics step you through the process of setting up a single Citrix Hypervisor VM to use NVIDIA vGPU. After the process is complete, you can install the graphics driver for your guest OS and license any NVIDIA vGPU software licensed products that you are using.

These setup steps assume familiarity with the Citrix Hypervisor skills covered in Citrix Hypervisor Basics.

2.5.1. Installing and Updating the NVIDIA Virtual GPU Manager for Citrix Hypervisor

The NVIDIA Virtual GPU Manager runs in the Citrix Hypervisor dom0 domain. The NVIDIA Virtual GPU Manager for Citrix Hypervisor is supplied as an RPM file and as a Supplemental Pack.

CAUTION:

NVIDIA Virtual GPU Manager and guest VM drivers must be compatible. If you update vGPU Manager to a release that is incompatible with the guest VM drivers, guest VMs will boot with vGPU disabled until their guest vGPU driver is updated to a compatible version. Consult Virtual GPU Software for Citrix Hypervisor Release Notes for further details.


2.5.1.1. Installing the RPM package for Citrix Hypervisor

The RPM file must be copied to the Citrix Hypervisor dom0 domain prior to installation (see Copying files to dom0).

  1. Use the rpm command to install the package:
    Copy
    Copied!
                

    [root@xenserver ~]# rpm -iv NVIDIA-vGPU-NVIDIA-vGPU-CitrixHypervisor-8.2-470.256.02.x86_64.rpm Preparing packages for installation... NVIDIA-vGPU-NVIDIA-vGPU-CitrixHypervisor-8.2-470.256.02 [root@xenserver ~]#

  2. Reboot the Citrix Hypervisor platform:
    Copy
    Copied!
                

    [root@xenserver ~]# shutdown –r now Broadcast message from root (pts/1) (Fri Jul 12 14:24:11 2024): The system is going down for reboot NOW! [root@xenserver ~]#

2.5.1.2. Updating the RPM Package for Citrix Hypervisor

If an existing NVIDIA Virtual GPU Manager is already installed on the system and you want to upgrade, follow these steps:

  1. Shut down any VMs that are using NVIDIA vGPU.
  2. Install the new package using the –U option to the rpm command, to upgrade from the previously installed package:
    Copy
    Copied!
                

    [root@xenserver ~]# rpm -Uv NVIDIA-vGPU-NVIDIA-vGPU-CitrixHypervisor-8.2-470.256.02.x86_64.rpm Preparing packages for installation... NVIDIA-vGPU-NVIDIA-vGPU-CitrixHypervisor-8.2-470.256.02 [root@xenserver ~]#

    Note:

    You can query the version of the current NVIDIA Virtual GPU Manager package using the rpm –q command:

    Copy
    Copied!
                

    [root@xenserver ~]# rpm –q NVIDIA-vGPU-NVIDIA-vGPU-CitrixHypervisor-8.2-470.256.02 [root@xenserver ~]# If an existing NVIDIA GRID package is already installed and you don’t select the upgrade (-U) option when installing a newer GRID package, the rpm command will return many conflict errors. Preparing packages for installation... file /usr/bin/nvidia-smi from install of NVIDIA-vGPU-NVIDIA-vGPU-CitrixHypervisor-8.2-470.256.02.x86_64 conflicts with file from package NVIDIA-vGPU-xenserver-8.2-470.223.02.x86_64 file /usr/lib/libnvidia-ml.so from install of NVIDIA-vGPU-NVIDIA-vGPU-CitrixHypervisor-8.2-470.256.02.x86_64 conflicts with file from package NVIDIA-vGPU-xenserver-8.2-470.223.02.x86_64 ...

  3. Reboot the Citrix Hypervisor platform:
    Copy
    Copied!
                

    [root@xenserver ~]# shutdown –r now Broadcast message from root (pts/1) (Fri Jul 12 14:24:11 2024): The system is going down for reboot NOW! [root@xenserver ~]#

2.5.1.3. Installing or Updating the Supplemental Pack for Citrix Hypervisor

XenCenter can be used to install or update Supplemental Packs on Citrix Hypervisor hosts. The NVIDIA Virtual GPU Manager supplemental pack is provided as an ISO.

  1. Select Install Update from the Tools menu.
  2. Click Next after going through the instructions on the Before You Start section.
  3. Click Select update or supplemental pack from disk on the Select Update section and open NVIDIA’s Citrix Hypervisor Supplemental Pack ISO.

    Figure 7. NVIDIA vGPU Manager supplemental pack selected in XenCenter

    supplemental-pack-vgu-selected-xencenter.png

  4. Click Next on the Select Update section.
  5. In the Select Servers section select all the Citrix Hypervisor hosts on which the Supplemental Pack should be installed on and click Next.
  6. Click Next on the Upload section once the Supplemental Pack has been uploaded to all the Citrix Hypervisor hosts.
  7. Click Next on the Prechecks section.
  8. Click Install Update on the Update Mode section.
  9. Click Finish on the Install Update section.

Figure 8. Successful installation of NVIDIA vGPU Manager supplemental pack

supplemental-pack-vgu-installed-xencenter.png

2.5.1.4. Verifying the Installation of the NVIDIA vGPU Software for Citrix Hypervisor Package

After the Citrix Hypervisor platform has rebooted, verify the installation of the NVIDIA vGPU software package for Citrix Hypervisor.

  1. Verify that the NVIDIA vGPU software package is installed and loaded correctly by checking for the NVIDIA kernel driver in the list of kernel loaded modules.
    Copy
    Copied!
                

    [root@xenserver ~]# lsmod | grep nvidia nvidia 9522927 0 i2c_core 20294 2 nvidia,i2c_i801 [root@xenserver ~]#

  2. Verify that the NVIDIA kernel driver can successfully communicate with the NVIDIA physical GPUs in your system by running the nvidia-smi command. The nvidia-smi command is described in more detail in NVIDIA System Management Interface nvidia-smi.

Running the nvidia-smi command should produce a listing of the GPUs in your platform.

Copy
Copied!
            

[root@xenserver ~]# nvidia-smi Fri Jul 12 18:46:50 2024 +------------------------------------------------------+ | NVIDIA-SMI 470.256.02 Driver Version: 470.256.02 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 Tesla M60 On | 00000000:05:00.0 Off | Off | | N/A 25C P8 24W / 150W | 13MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ | 1 Tesla M60 On | 00000000:06:00.0 Off | Off | | N/A 24C P8 24W / 150W | 13MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ | 2 Tesla M60 On | 00000000:86:00.0 Off | Off | | N/A 25C P8 25W / 150W | 13MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ | 3 Tesla M60 On | 00000000:87:00.0 Off | Off | | N/A 28C P8 24W / 150W | 13MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | No running processes found | +-----------------------------------------------------------------------------+ [root@xenserver ~]#


If nvidia-smi fails to run or doesn’t produce the expected output for all the NVIDIA GPUs in your system, see Troubleshooting for troubleshooting steps.

2.5.2. Configuring a Citrix Hypervisor VM with Virtual GPU

To support applications and workloads that are compute or graphics intensive, you can add multiple vGPUs to a single VM.

For details about which Citrix Hypervisor versions and NVIDIA vGPUs support the assignment of multiple vGPUs to a VM, see Virtual GPU Software for Citrix Hypervisor Release Notes.

Citrix Hypervisor supports configuration and management of virtual GPUs using XenCenter, or the xe command line tool that is run in a Citrix Hypervisor dom0 shell. Basic configuration using XenCenter is described in the following sections. Command line management using xe is described in Citrix Hypervisor vGPU Management.

Note:

If you are using Citrix Hypervisor 8.1 or later and need to assign plugin configuration parameters, create vGPUs using the xe command as explained in Creating a vGPU Using xe.


  1. Ensure the VM is powered off.
  2. Right-click the VM in XenCenter, select Properties to open the VM’s properties, and select the GPU property. The available GPU types are listed in the GPU type drop-down list:

    Figure 9. Using Citrix XenCenter to configure a VM with a vGPU

    xencenter-configure-vm-with-gpu.png

After you have configured a Citrix Hypervisor VM with a vGPU, start the VM, either from XenCenter or by using xe vm-start in a dom0 shell. You can view the VM’s console in XenCenter.

After the VM has booted, install the NVIDIA vGPU software graphics driver as explained in Installing the NVIDIA vGPU Software Graphics Driver.

2.5.3. Setting vGPU Plugin Parameters on Citrix Hypervisor

Plugin parameters for a vGPU control the behavior of the vGPU, such as the frame rate limiter (FRL) configuration in frames per second or whether console virtual network computing (VNC) for the vGPU is enabled. The VM to which the vGPU is assigned is started with these parameters. If parameters are set for multiple vGPUs assigned to the same VM, the VM is started with the parameters assigned to each vGPU.

For each vGPU for which you want to set plugin parameters, perform this task in a command shell in the Citrix Hypervisor dom0 domain.

  1. Get the UUIDs of all VMs on the hypervisor host and use the output from the command to identify the VM to which the vGPU is assigned.
    Copy
    Copied!
                

    [root@xenserver ~] xe vm-list ... uuid ( RO) : 7f6c855d-5635-2d57-9fbc-b1200172162f name-label ( RW): RHEL8.3 power-state ( RO): running ...

  2. Get the UUIDs of all vGPUs on the hypervisor host and from the UUID of the VM to which the vGPU is assigned, determine the UUID of the vGPU.
    Copy
    Copied!
                

    [root@xenserver ~] xe vgpu-list ... uuid ( RO) : d15083f8-5c59-7474-d0cb-fbc3f7284f1b vm-uuid ( RO): 7f6c855d-5635-2d57-9fbc-b1200172162f device ( RO): 0 gpu-group-uuid ( RO): 3a2fbc36-827d-a078-0b2f-9e869ae6fd93 ...

  3. Use the xe command to set each vGPU plugin parameter that you want to set.
    Copy
    Copied!
                

    [root@xenserver ~] xe vgpu-param-set uuid=vgpu-uuid extra_args='parameter=value'

    vgpu-uuid
    The UUID of the vGPU, which you obtained in the previous step.
    parameter
    The name of the vGPU plugin parameter that you want to set.
    value
    The value to which you want to set the vGPU plugin parameter.

    This example sets the enable_uvm vGPU plugin parameter to 1 for the vGPU that has the UUID d15083f8-5c59-7474-d0cb-fbc3f7284f1b. This parameter setting enables unified memory for the vGPU.

    Copy
    Copied!
                

    [root@xenserver ~] xe vgpu-param-set uuid=d15083f8-5c59-7474-d0cb-fbc3f7284f1b extra_args='enable_uvm=1'

2.6. Installing the Virtual GPU Manager Package for Linux KVM

NVIDIA vGPU software for Linux Kernel-based Virtual Machine (KVM) (Linux KVM) is intended only for use with supported versions of Linux KVM hypervisors. For details about which Linux KVM hypervisor versions are supported, see Virtual GPU Software for Generic Linux with KVM Release Notes.

Note:

If you are using Red Hat Enterprise Linux KVM, follow the instructions in Installing and Configuring the NVIDIA Virtual GPU Manager for Red Hat Enterprise Linux KVM or RHV.


Before installing the Virtual GPU Manager package for Linux KVM, ensure that the following prerequisites are met:

  • The following packages are installed on the Linux KVM server:

    • The x86_64 build of the GNU Compiler Collection (GCC)
    • Linux kernel headers
  • The package file is copied to a directory in the file system of the Linux KVM server.

If the Nouveau driver for NVIDIA graphics cards is present, disable it before installing the package.

  1. Change to the directory on the Linux KVM server that contains the package file.
    Copy
    Copied!
                

    # cd package-file-directory

    package-file-directory
    The path to the directory that contains the package file.
  2. Make the package file executable.
    Copy
    Copied!
                

    # chmod +x package-file-name

    package-file-name
    The name of the file that contains the Virtual GPU Manager package for Linux KVM, for example NVIDIA-Linux-x86_64-390.42-vgpu-kvm.run.
  3. Run the package file as the root user.
    Copy
    Copied!
                

    # sudo sh./package-file-name

    The package file should launch and display the license agreement.
  4. Accept the license agreement to continue with the installation.
  5. When installation has completed, select OK to exit the installer.
  6. Reboot the Linux KVM server.
    Copy
    Copied!
                

    # systemctl reboot

2.7. Installing and Configuring the NVIDIA Virtual GPU Manager for Red Hat Enterprise Linux KVM or RHV

The following topics step you through the process of setting up a single Red Hat Enterprise Linux Kernel-based Virtual Machine (KVM) or Red Hat Virtualization (RHV) VM to use NVIDIA vGPU.

Red Hat Enterprise Linux KVM and RHV use the same Virtual GPU Manager package, but are configured with NVIDIA vGPU in different ways.

CAUTION:

Output from the VM console is not available for VMs that are running vGPU. Make sure that you have installed an alternate means of accessing the VM (such as a VNC server) before you configure vGPU.

For RHV, follow this sequence of instructions:

  1. Installing the NVIDIA Virtual GPU Manager for Red Hat Enterprise Linux KVM or RHV
  2. MIG-backed vGPUs only: Configuring a GPU for MIG-Backed vGPUs
  3. Creating an NVIDIA vGPU on a Linux with KVM Hypervisor
  4. Adding a vGPU to a Red Hat Virtualization (RHV) VM

For Red Hat Enterprise Linux KVM, follow this sequence of instructions:

  1. Installing the NVIDIA Virtual GPU Manager for Red Hat Enterprise Linux KVM or RHV
  2. MIG-backed vGPUs only: Configuring a GPU for MIG-Backed vGPUs
  3. Getting the BDF and Domain of a GPU on a Linux with KVM Hypervisor
  4. Creating an NVIDIA vGPU on a Linux with KVM Hypervisor
  5. Adding One or More vGPUs to a Linux with KVM Hypervisor VM
  6. Setting vGPU Plugin Parameters on a Linux with KVM Hypervisor

After the process is complete, you can install the graphics driver for your guest OS and license any NVIDIA vGPU software licensed products that you are using.

Note:

If you are using a generic Linux KVM hypervisor, follow the instructions in Installing the Virtual GPU Manager Package for Linux KVM.


2.7.1. Installing the NVIDIA Virtual GPU Manager for Red Hat Enterprise Linux KVM or RHV

The NVIDIA Virtual GPU Manager for Red Hat Enterprise Linux KVM and Red Hat Virtualization (RHV) is provided as a .rpm file.

CAUTION:

NVIDIA Virtual GPU Manager and guest VM drivers must be compatible. If you update vGPU Manager to a release that is incompatible with the guest VM drivers, guest VMs will boot with vGPU disabled until their guest vGPU driver is updated to a compatible version. Consult Virtual GPU Software for Red Hat Enterprise Linux with KVM Release Notes for further details.


2.7.1.1. Installing the Virtual GPU Manager Package for Red Hat Enterprise Linux KVM or RHV

Before installing the RPM package for Red Hat Enterprise Linux KVM or RHV, ensure that the sshd service on the Red Hat Enterprise Linux KVM or RHV server is configured to permit root login. If the Nouveau driver for NVIDIA graphics cards is present, disable it before installing the package. For instructions, see How to disable the Nouveau driver and install the Nvidia driver in RHEL 7 (Red Hat subscription required).

Some versions of Red Hat Enterprise Linux KVM have z-stream updates that break Kernel Application Binary Interface (kABI) compatibility with the previous kernel or the GA kernel. For these versions of Red Hat Enterprise Linux KVM, the following Virtual GPU Manager RPM packages are supplied:

  • A package for the GA Linux KVM kernel
  • A package for the updated z-stream kernel

To differentiate these packages, the name of each RPM package includes the kernel version. Ensure that you install the RPM package that is compatible with your Linux KVM kernel version.

  1. Securely copy the RPM file from the system where you downloaded the file to the Red Hat Enterprise Linux KVM or RHV server.
    • From a Windows system, use a secure copy client such as WinSCP.
    • From a Linux system, use the scp command.
  2. Use secure shell (SSH) to log in as root to the Red Hat Enterprise Linux KVM or RHV server.
    Copy
    Copied!
                

    # ssh root@kvm-server

    kvm-server
    The host name or IP address of the Red Hat Enterprise Linux KVM or RHV server.
  3. Change to the directory on the Red Hat Enterprise Linux KVM or RHV server to which you copied the RPM file.
    Copy
    Copied!
                

    # cd rpm-file-directory

    rpm-file-directory
    The path to the directory to which you copied the RPM file.
  4. Use the rpm command to install the package.
    Copy
    Copied!
                

    # rpm -iv NVIDIA-vGPU-rhel-8.8-470.256.02.x86_64.rpm Preparing packages for installation... NVIDIA-vGPU-rhel-8.8-470.256.02 #

  5. Reboot the Red Hat Enterprise Linux KVM or RHV server.
    Copy
    Copied!
                

    # systemctl reboot

2.7.1.2. Verifying the Installation of the NVIDIA vGPU Software for Red Hat Enterprise Linux KVM or RHV

After the Red Hat Enterprise Linux KVM or RHV server has rebooted, verify the installation of the NVIDIA vGPU software package for Red Hat Enterprise Linux KVM or RHV.

  1. Verify that the NVIDIA vGPU software package is installed and loaded correctly by checking for the VFIO drivers in the list of kernel loaded modules.
    Copy
    Copied!
                

    # lsmod | grep vfio nvidia_vgpu_vfio 27099 0 nvidia 12316924 1 nvidia_vgpu_vfio vfio_mdev 12841 0 mdev 20414 2 vfio_mdev,nvidia_vgpu_vfio vfio_iommu_type1 22342 0 vfio 32331 3 vfio_mdev,nvidia_vgpu_vfio,vfio_iommu_type1 #

  2. Verify that the libvirtd service is active and running.
    Copy
    Copied!
                

    # service libvirtd status

  3. Verify that the NVIDIA kernel driver can successfully communicate with the NVIDIA physical GPUs in your system by running the nvidia-smi command. The nvidia-smi command is described in more detail in NVIDIA System Management Interface nvidia-smi.

Running the nvidia-smi command should produce a listing of the GPUs in your platform.

Copy
Copied!
            

# nvidia-smi Fri Jul 12 18:46:50 2024 +------------------------------------------------------+ | NVIDIA-SMI 470.256.02 Driver Version: 470.256.02 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 Tesla M60 On | 0000:85:00.0 Off | Off | | N/A 23C P8 23W / 150W | 13MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ | 1 Tesla M60 On | 0000:86:00.0 Off | Off | | N/A 29C P8 23W / 150W | 13MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ | 2 Tesla P40 On | 0000:87:00.0 Off | Off | | N/A 21C P8 18W / 250W | 53MiB / 24575MiB | 0% Default | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | No running processes found | +-----------------------------------------------------------------------------+ #


If nvidia-smi fails to run or doesn’t produce the expected output for all the NVIDIA GPUs in your system, see Troubleshooting for troubleshooting steps.

2.7.2. Adding a vGPU to a Red Hat Virtualization (RHV) VM

Ensure that the VM to which you want to add the vGPU is shut down.

  1. Determine the mediated device type (mdev_type) identifiers of the vGPU types available on the RHV host.
    Copy
    Copied!
                

    # vdsm-client Host hostdevListByCaps ... "mdev": { "nvidia-155": { "name": "GRID M10-2B", "available_instances": "4" }, "nvidia-36": { "name": "GRID M10-0Q", "available_instances": "16" }, ...

    The preceding example shows the mdev_type identifiers of the following vGPU types:
    • For the GRID M10-2B vGPU type, the mdev_type identifier is nvidia-155.
    • For the GRID M10-0Q vGPU type, the mdev_type identifier is nvidia-36.
  2. Note the mdev_type identifier of the vGPU type that you want to add.
  3. Log in to the RHV Administration Portal.
  4. From the Main Navigation Menu, choose Compute > Virtual Machines > virtual-machine-name.
    virtual-machine-name
    The name of the virtual machine to which you want to add the vGPU.
  5. Click Edit.
  6. In the Edit Virtual Machine window that opens, click Show Advanced Options and in the list of options, select Custom Properties.
  7. From the drop-down list, select mdev_type.
  8. In the text field, type the mdev_type identifier of the vGPU type that you want to add and click OK.

    rhv-assign-vgpu.png

After adding a vGPU to an RHV VM, start the VM.

After the VM has booted, install the NVIDIA vGPU software graphics driver as explained in Installing the NVIDIA vGPU Software Graphics Driver.

2.8. Since 13.1: Installing and Configuring the NVIDIA Virtual GPU Manager for Ubuntu

Follow this sequence of instructions to set up a single Ubuntu VM to use NVIDIA vGPU.

  1. Installing the NVIDIA Virtual GPU Manager for Ubuntu
  2. MIG-backed vGPUs only: Configuring a GPU for MIG-Backed vGPUs
  3. Getting the BDF and Domain of a GPU on a Linux with KVM Hypervisor
  4. Creating an NVIDIA vGPU on a Linux with KVM Hypervisor
  5. Adding One or More vGPUs to a Linux with KVM Hypervisor VM
  6. Setting vGPU Plugin Parameters on a Linux with KVM Hypervisor
CAUTION:

Output from the VM console is not available for VMs that are running vGPU. Make sure that you have installed an alternate means of accessing the VM (such as a VNC server) before you configure vGPU.

After the process is complete, you can install the graphics driver for your guest OS and license any NVIDIA vGPU software licensed products that you are using.

2.8.1. Installing the NVIDIA Virtual GPU Manager for Ubuntu

The NVIDIA Virtual GPU Manager for Ubuntu is provided as a Debian package (.deb) file.

CAUTION:

NVIDIA Virtual GPU Manager and guest VM drivers must be compatible. If you update vGPU Manager to a release that is incompatible with the guest VM drivers, guest VMs will boot with vGPU disabled until their guest vGPU driver is updated to a compatible version. Consult Virtual GPU Software for Ubuntu Release Notes for further details.


2.8.1.1. Installing the Virtual GPU Manager Package for Ubuntu

Before installing the Debian package for Ubuntu, ensure that the sshd service on the Ubuntu server is configured to permit root login. If the Nouveau driver for NVIDIA graphics cards is present, disable it before installing the package.

  1. Securely copy the Debian package file from the system where you downloaded the file to the Ubuntu server.
    • From a Windows system, use a secure copy client such as WinSCP.
    • From a Linux system, use the scp command.
  2. Use secure shell (SSH) to log in as root to the Ubuntu server.
    Copy
    Copied!
                

    # ssh root@ubuntu-server

    ubuntu-server
    The host name or IP address of the Ubuntu server.
  3. Change to the directory on the Ubuntu server to which you copied the Debian package file.
    Copy
    Copied!
                

    # cd deb-file-directory

    deb-file-directory
    The path to the directory to which you copied the Debian package file.
  4. Use the apt command to install the package.
    Copy
    Copied!
                

    # apt install ./nvidia-vgpu-ubuntu-470.256.02_amd64.deb

  5. Reboot the Ubuntu server.
    Copy
    Copied!
                

    # systemctl reboot

2.8.1.2. Verifying the Installation of the NVIDIA vGPU Software for Ubuntu

After the Ubuntu server has rebooted, verify the installation of the NVIDIA vGPU software package for Red Hat Enterprise Linux KVM or RHV.

  1. Verify that the NVIDIA vGPU software package is installed and loaded correctly by checking for the VFIO drivers in the list of kernel loaded modules.
    Copy
    Copied!
                

    # lsmod | grep vfio nvidia_vgpu_vfio 27099 0 nvidia 12316924 1 nvidia_vgpu_vfio vfio_mdev 12841 0 mdev 20414 2 vfio_mdev,nvidia_vgpu_vfio vfio_iommu_type1 22342 0 vfio 32331 3 vfio_mdev,nvidia_vgpu_vfio,vfio_iommu_type1 #

  2. Verify that the libvirtd service is active and running.
    Copy
    Copied!
                

    # service libvirtd status

  3. Verify that the NVIDIA kernel driver can successfully communicate with the NVIDIA physical GPUs in your system by running the nvidia-smi command. The nvidia-smi command is described in more detail in NVIDIA System Management Interface nvidia-smi.

Running the nvidia-smi command should produce a listing of the GPUs in your platform.

Copy
Copied!
            

# nvidia-smi Fri Jul 12 18:46:50 2024 +------------------------------------------------------+ | NVIDIA-SMI 470.256.02 Driver Version: 470.256.02 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 Tesla M60 On | 0000:85:00.0 Off | Off | | N/A 23C P8 23W / 150W | 13MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ | 1 Tesla M60 On | 0000:86:00.0 Off | Off | | N/A 29C P8 23W / 150W | 13MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ | 2 Tesla P40 On | 0000:87:00.0 Off | Off | | N/A 21C P8 18W / 250W | 53MiB / 24575MiB | 0% Default | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | No running processes found | +-----------------------------------------------------------------------------+ #


If nvidia-smi fails to run or doesn’t produce the expected output for all the NVIDIA GPUs in your system, see Troubleshooting for troubleshooting steps.

2.9. Installing and Configuring the NVIDIA Virtual GPU Manager for VMware vSphere

You can use the NVIDIA Virtual GPU Manager for VMware vSphere to set up a VMware vSphere VM to use NVIDIA vGPU or VMware vSGA. The vGPU Manager vSphere Installation Bundles (VIBs) for VMware vSphere 6.5 and later provide vSGA and vGPU functionality in a single VIB. For VMware vSphere 6.0, vSGA and vGPU functionality are provided in separate vGPU Manager VIBs.

Note:

Some servers, for example, the Dell R740, do not configure SR-IOV capability if the SR-IOV SBIOS setting is disabled on the server. If you are using the Tesla T4 GPU with VMware vSphere on such a server, you must ensure that the SR-IOV SBIOS setting is enabled on the server.

However, with any server hardware, do not enable SR-IOV in VMware vCenter Server for the Tesla T4 GPU. If SR-IOV is enabled in VMware vCenter Server for T4, VMware vCenter Server lists the status of the GPU as needing a reboot. You can ignore this status message.

NVIDIA vGPU Instructions

Note:

As of VMware vSphere 7.0 Update 1, the Xorg service is no longer required for graphics devices in NVIDIA vGPU mode. For more information, see Installing and Updating the NVIDIA Virtual GPU Manager for vSphere.

For NVIDIA vGPU, follow this sequence of instructions:

  1. Installing and Updating the NVIDIA Virtual GPU Manager for vSphere
  2. Configuring VMware vMotion with vGPU for VMware vSphere
  3. Changing the Default Graphics Type in VMware vSphere 6.5 and Later
  4. MIG-backed vGPUs only: Configuring a GPU for MIG-Backed vGPUs
  5. Configuring a vSphere VM with NVIDIA vGPU
  6. Optional: Setting vGPU Plugin Parameters on VMware vSphere

After configuring a vSphere VM to use NVIDIA vGPU, you can install the NVIDIA vGPU software graphics driver for your guest OS and license any NVIDIA vGPU software licensed products that you are using.

VMware vSGA Instructions

For VMware vSGA, follow this sequence of instructions:

  1. Installing and Updating the NVIDIA Virtual GPU Manager for vSphere
  2. Configuring a vSphere VM with VMware vSGA

Installation of the NVIDIA vGPU software graphics driver for the guest OS is not required for vSGA.

Requirements for Configuring NVIDIA vGPU in a DRS Cluster

You can configure a VM with NVIDIA vGPU on an ESXi host in a VMware Distributed Resource Scheduler (DRS) cluster. However, you must ensure that the automation level of the cluster supports VMs configured with NVIDIA vGPU:

  • For any supported VMware vSphere release, set the automation level to Manual.
  • For VMware vSphere 6.7 Update 1 or later, set the automation level to Partially Automated or Manual.

For more information about these settings, see Edit Cluster Settings in the VMware documentation.

2.9.1. Installing and Updating the NVIDIA Virtual GPU Manager for vSphere

The NVIDIA Virtual GPU Manager runs on the ESXi host. How the NVIDIA Virtual GPU Manager package is distributed depends on the release of VMware vSphere.

  • For all supported VMware vSphere releases, the NVIDIA Virtual GPU Manager package is distributed as a software component in a ZIP archive, which you can install in one of the following ways:

  • For supported releases before VMware vSphere 7.0, the NVIDIA Virtual GPU Manager package is also distributed as a vSphere Installation Bundle (VIB) file, which you must copy to the ESXi host and then install as explained in Installing the NVIDIA Virtual GPU Manager Package for vSphere.
CAUTION:

NVIDIA Virtual GPU Manager and guest VM drivers must be compatible. If you update vGPU Manager to a release that is incompatible with the guest VM drivers, guest VMs will boot with vGPU disabled until their guest vGPU driver is updated to a compatible version. Consult Virtual GPU Software for VMware vSphere Release Notes for further details.


2.9.1.1. Installing the NVIDIA Virtual GPU Manager Package for vSphere

To install the vGPU Manager package you need to access the ESXi host via the ESXi Shell or SSH. Refer to VMware’s documentation on how to enable ESXi Shell or SSH for an ESXi host.
Before you begin, ensure that the following prerequisites are met:

  • The ZIP archive that contains NVIDIA vGPU software has been downloaded from the NVIDIA Licensing Portal.
  • The NVIDIA Virtual GPU Manager package has been extracted from the downloaded ZIP archive.
  1. Copy the NVIDIA Virtual GPU Manager package file to the ESXi host.
  2. Put the ESXi host into maintenance mode.
    Copy
    Copied!
                

    $ esxcli system maintenanceMode set –-enable true

  3. Use the esxcli command to install the vGPU Manager package. For more information about the esxcli command, see esxcli software Commands in the VMware vSphere documentation.
    • For a software component, run the following command:

      Copy
      Copied!
                  

      [root@esxi:~] esxcli software vib install -d /vmfs/volumes/datastore/software-component.zip

      datastore
      The name of the VMFS datastore to which you copied the software component.
      software-component
      The name of the file that contains the NVIDIA Virtual GPU Manager package in the form of a software component. Ensure that you specify the file that was extracted from the downloaded ZIP archive. For example, for VMware vSphere 7.0, software-component is NVD.NVIDIA_bootbank_NVIDIA-VMware_470.256.02-1OEM.700.0.0.8169922-offline_bundle-build-number.
    • For a VIB file, run the following command:

      Copy
      Copied!
                  

      [root@esxi:~] esxcli software vib install -v directory/NVIDIA-vGPU-VMware_ESXi_7.0_Host_Driver_470.256.02-1OEM.700.0.0.8169922.vib Installation Result Message: Operation finished successfully. Reboot Required: false VIBs Installed: NVIDIA-vGPU-VMware_ESXi_7.0_Host_Driver_470.256.02-1OEM.700.0.0.8169922 VIBs Removed: VIBs Skipped:

      directory
      The absolute path to the directory to which you copied the VIB file. You must specify the absolute path even if the VIB file is in the current working directory.
  4. Exit maintenance mode.
    Copy
    Copied!
                

    $ esxcli system maintenanceMode set –-enable false

  5. Reboot the ESXi host.
    Copy
    Copied!
                

    $ reboot

2.9.1.2. Updating the NVIDIA Virtual GPU Manager Package for vSphere

Update the vGPU Manager VIB package if you want to install a new version of NVIDIA Virtual GPU Manager on a system where an existing version is already installed.

CAUTION:

Do not perform this task on a system where an existing version isn't already installed. If you perform this task on a system where an existing version isn't already installed, the Xorg service (when required) fails to start after the NVIDIA vGPU software driver is installed. Instead, install the vGPU Manager VIB package as explained in Installing the NVIDIA Virtual GPU Manager Package for vSphere.

To update the vGPU Manager VIB you need to access the ESXi host via the ESXi Shell or SSH. Refer to VMware’s documentation on how to enable ESXi Shell or SSH for an ESXi host.

Note:

Before proceeding with the vGPU Manager update, make sure that all VMs are powered off and the ESXi host is placed in maintenance mode. Refer to VMware’s documentation on how to place an ESXi host in maintenance mode


  1. Use the esxcli command to update the vGPU Manager package:
    Copy
    Copied!
                

    [root@esxi:~] esxcli software vib update -v directory/NVIDIA-vGPU-VMware_ESXi_7.0_Host_Driver_470.256.02-1OEM.700.0.0.8169922.vib Installation Result    Message: Operation finished successfully.    Reboot Required: false    VIBs Installed: NVIDIA-vGPU-VMware_ESXi_7.0_Host_Driver_470.256.02-1OEM.700.0.0.8169922    VIBs Removed: NVIDIA-vGPU-VMware_ESXi_7.0_Host_Driver_470.223.02-1OEM.700.0.0.8169922    VIBs Skipped:

    directory is the path to the directory that contains the VIB file.

  2. Reboot the ESXi host and remove it from maintenance mode.

2.9.1.3. Verifying the Installation of the NVIDIA vGPU Software Package for vSphere

After the ESXi host has rebooted, verify the installation of the NVIDIA vGPU software package for vSphere.

  1. Verify that the NVIDIA vGPU software package installed and loaded correctly by checking for the NVIDIA kernel driver in the list of kernel loaded modules.
    Copy
    Copied!
                

    [root@esxi:~] vmkload_mod -l | grep nvidia nvidia 5 8420

  2. If the NVIDIA driver is not listed in the output, check dmesg for any load-time errors reported by the driver.
  3. Verify that the NVIDIA kernel driver can successfully communicate with the NVIDIA physical GPUs in your system by running the nvidia-smi command. The nvidia-smi command is described in more detail in NVIDIA System Management Interface nvidia-smi.

Running the nvidia-smi command should produce a listing of the GPUs in your platform.

Copy
Copied!
            

[root@esxi:~] nvidia-smi Fri Jul 12 17:56:22 2024 +------------------------------------------------------+ | NVIDIA-SMI 470.256.02 Driver Version: 470.256.02 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 Tesla M60 On | 00000000:05:00.0 Off | Off | | N/A 25C P8 24W / 150W | 13MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ | 1 Tesla M60 On | 00000000:06:00.0 Off | Off | | N/A 24C P8 24W / 150W | 13MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ | 2 Tesla M60 On | 00000000:86:00.0 Off | Off | | N/A 25C P8 25W / 150W | 13MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ | 3 Tesla M60 On | 00000000:87:00.0 Off | Off | | N/A 28C P8 24W / 150W | 13MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | No running processes found | +-----------------------------------------------------------------------------+


If nvidia-smi fails to report the expected output for all the NVIDIA GPUs in your system, see Troubleshooting for troubleshooting steps.

2.9.2. Configuring VMware vMotion with vGPU for VMware vSphere

NVIDIA vGPU software supports vGPU migration, which includes VMware vMotion and suspend-resume, for VMs that are configured with vGPU. To enable VMware vMotion with vGPU, an advanced vCenter Server setting must be enabled. However, suspend-resume for VMs that are configured with vGPU is enabled by default.

For details about which VMware vSphere versions, NVIDIA GPUs, and guest OS releases support vGPU migration, see Virtual GPU Software for VMware vSphere Release Notes.
Before configuring VMware vMotion with vGPU for an ESXi host, ensure that the current NVIDIA Virtual GPU Manager for VMware vSphere package is installed on the host.

  1. Log in to vCenter Server by using the vSphere Web Client.
  2. In the Hosts and Clusters view, select the vCenter Server instance.
    Note:

    Ensure that you select the vCenter Server instance, not the vCenter Server VM.

  3. Click the Configure tab.
  4. In the Settings section, select Advanced Settings and click Edit.
  5. In the Edit Advanced vCenter Server Settings window that opens, type vGPU in the search field.
  6. When the vgpu.hotmigrate.enabled setting appears, set the Enabled option and click OK.

    vmware-vcenter-edit-advanced-server-settings-window.png

2.9.3. Changing the Default Graphics Type in VMware vSphere 6.5 and Later

The vGPU Manager VIBs for VMware vSphere 6.5 and later provide vSGA and vGPU functionality in a single VIB. After this VIB is installed, the default graphics type is Shared, which provides vSGA functionality. To enable vGPU support for VMs in VMware vSphere 6.5, you must change the default graphics type to Shared Direct. If you do not change the default graphics type, VMs to which a vGPU is assigned fail to start and the following error message is displayed:

Copy
Copied!
            

The amount of graphics resource available in the parent resource pool is insufficient for the operation.

Note:

If you are using a supported version of VMware vSphere earlier than 6.5, or are configuring a VM to use vSGA, omit this task.

Change the default graphics type before configuring vGPU. Output from the VM console in the VMware vSphere Web Client is not available for VMs that are running vGPU.


Before changing the default graphics type, ensure that the ESXi host is running and that all VMs on the host are powered off.

  1. Log in to vCenter Server by using the vSphere Web Client.
  2. In the navigation tree, select your ESXi host and click the Configure tab.
  3. From the menu, choose Graphics and then click the Host Graphics tab.
  4. On the Host Graphics tab, click Edit.

    Figure 10. Shared default graphics type

    configure-host-graphics-tab-shared-mode-vmware-vsphere.png

  5. In the Edit Host Graphics Settings dialog box that opens, select Shared Direct and click OK.

    Figure 11. Host graphics settings for vGPU

    shared-to-shared-direct-mode-vmware-vsphere.png

    Note:

    In this dialog box, you can also change the allocation scheme for vGPU-enabled VMs. For more information, see Modifying GPU Allocation Policy on VMware vSphere.

    After you click OK, the default graphics type changes to Shared Direct.

  6. Click the Graphics Devices tab to verify the configured type of each physical GPU on which you want to configure vGPU. The configured type of each physical GPU must be Shared Direct. For any physical GPU for which the configured type is Shared, change the configured type as follows:
    1. On the Graphics Devices tab, select the physical GPU and click the Edit icon.

      Figure 12. Shared graphics type

      graphics-devices-shared-vmware-vsphere.png

    2. In the Edit Graphics Device Settings dialog box that opens, select Shared Direct and click OK.

      Figure 13. Graphics device settings for a physical GPU

      shared-to-shared-direct-mode-pgpu-vmware-vsphere.png

  7. Restart the ESXi host or stop and restart the Xorg service if necessary and nv-hostengine on the ESXi host.

    To stop and restart the Xorg service and nv-hostengine, perform these steps:

    1. VMware vSphere releases before 7.0 Update 1 only: Stop the Xorg service.

      As of VMware vSphere 7.0 Update 1, the Xorg service is no longer required for graphics devices in NVIDIA vGPU mode.

    2. Stop nv-hostengine.
      Copy
      Copied!
                  

      [root@esxi:~] nv-hostengine -t

    3. Wait for 1 second to allow nv-hostengine to stop.
    4. Start nv-hostengine.
      Copy
      Copied!
                  

      [root@esxi:~] nv-hostengine -d

    5. VMware vSphere releases before 7.0 Update 1 only: Start the Xorg service.

      As of VMware vSphere 7.0 Update 1, the Xorg service is no longer required for graphics devices in NVIDIA vGPU mode.

      Copy
      Copied!
                  

      [root@esxi:~] /etc/init.d/xorg start

  8. In the Graphics Devices tab of the VMware vCenter Web UI, confirm that the active type and the configured type of each physical GPU are Shared Direct.

    Figure 14. Shared direct graphics type

    graphics-devices-shared-direct-vmware-vsphere.png

After changing the default graphics type, configure vGPU as explained in Configuring a vSphere VM with NVIDIA vGPU.

See also the following topics in the VMware vSphere documentation:

2.9.4. Configuring a vSphere VM with NVIDIA vGPU

To support applications and workloads that are compute or graphics intensive, you can add multiple vGPUs to a single VM.

For details about which VMware vSphere versions and NVIDIA vGPUs support the assignment of multiple vGPUs to a VM, see Virtual GPU Software for VMware vSphere Release Notes.

If you upgraded to VMware vSphere 6.7 Update 3 from an earlier version and are using VMs that were created with that version, change the VM compatibility to vSphere 6.7 Update 2 and later. For details, see Virtual Machine Compatibility in the VMware documentation.

If you are adding multiple vGPUs to a single VM, perform this task for each vGPU that you want to add to the VM.

CAUTION:

Output from the VM console in the VMware vSphere Web Client is not available for VMs that are running vGPU. Make sure that you have installed an alternate means of accessing the VM (such as VMware Horizon or a VNC server) before you configure vGPU.

VM console in vSphere Web Client will become active again once the vGPU parameters are removed from the VM’s configuration.

Note:

If you are configuring a VM to use VMware vSGA, omit this task.


  1. Open the vCenter Web UI.
  2. In the vCenter Web UI, right-click the VM and choose Edit Settings.
  3. Click the Virtual Hardware tab.
  4. In the New device list, select Shared PCI Device and click Add. The PCI device field should be auto-populated with NVIDIA GRID vGPU.

    Figure 15. VM settings for vGPU

    vgpu-settings-vmware-vsphere.png

  5. From the GPU Profile drop-down menu, choose the type of vGPU you want to configure and click OK.
    Note:

    VMware vSphere does not support vCS. Therefore, C-series vGPU types are not available for selection from the GPU Profile drop-down menu.

  6. Ensure that VMs running vGPU have all their memory reserved:
    1. Select Edit virtual machine settings from the vCenter Web UI.
    2. Expand the Memory section and click Reserve all guest memory (All locked).

After you have configured a vSphere VM with a vGPU, start the VM. VM console in vSphere Web Client is not supported in this vGPU release. Therefore, use VMware Horizon or VNC to access the VM’s desktop.

After the VM has booted, install the NVIDIA vGPU software graphics driver as explained in Installing the NVIDIA vGPU Software Graphics Driver.

2.9.5. Setting vGPU Plugin Parameters on VMware vSphere

Plugin parameters for a vGPU control the behavior of the vGPU, such as the frame rate limiter (FRL) configuration in frames per second or whether console virtual network computing (VNC) for the vGPU is enabled. The VM to which the vGPU is assigned is started with these parameters. If parameters are set for multiple vGPUs assigned to the same VM, the VM is started with the parameters assigned to each vGPU.

Ensure that the VM to which the vGPU is assigned is powered off.
For each vGPU for which you want to set plugin parameters, perform this task in the vSphere Client. vGPU plugin parameters are PCI pass through configuration parameters in advanced VM attributes.

  1. In the vSphere Client, browse to the VM to which the vGPU is assigned.
  2. Context-click the VM and choose Edit Settings.
  3. In the Edit Settings window, click the VM Options tab.
  4. From the Advanced drop-down list, select Edit Configuration.
  5. In the Configuration Parameters dialog box, click Add Row.
  6. In the Name field, type the parameter name pciPassthruvgpu-id.cfg.parameter, in the Value field type the parameter value, and click OK.
    vgpu-id
    A positive integer that identifies the vGPU assigned to a VM. For the first vGPU assigned to a VM, vgpu-id is 0. For example, if two vGPUs are assigned to a VM and you are setting a plugin parameter for both vGPUs, set the following parameters:
    • pciPassthru0.cfg.parameter
    • pciPassthru1.cfg.parameter
    parameter
    The name of the vGPU plugin parameter that you want to set. For example, the name of the vGPU plugin parameter for enabling unified memory is enable_uvm.

    To enable unified memory for two vGPUs that are assigned to a VM, set pciPassthru0.cfg.enable_uvm and pciPassthru1.cfg.enable_uvm to 1.

2.9.6. Configuring a vSphere VM with VMware vSGA

Virtual Shared Graphics Acceleration (vSGA) is a feature of VMware vSphere that enables multiple virtual machines to share the physical GPUs on ESXi hosts.

Note:

If you are configuring a VM to use NVIDIA vGPU, omit this task.


Before configuring a vSphere VM with vSGA, ensure that these prerequisites are met:

  • VMware tools are installed on the VM.
  • The VM is powered off.
  • The NVIDIA Virtual GPU Manager package for vSphere is installed.
  1. Open the vCenter Web UI.
  2. In the vCenter Web UI, right-click the VM and choose Edit Settings.
  3. Click the Virtual Hardware tab.
  4. In the device list, expand the Video card node and set the following options:
    1. Select the Enable 3D support option.
    2. Set the 3D Renderer to Hardware.

    For more information, see Configure 3D Graphics and Video Cards in the VMware Horizon documentation.

  5. Start the VM.
  6. After the VM has booted, verify that the VM has been configured correctly with vSGA.
    1. Under the Display Adapter section of Device Manager, confirm that VMware SVGA 3D is listed.
    2. Verify that the virtual machine is using the GPU card.
      Copy
      Copied!
                  

      # gpuvm

      The output from the command is similar to the following example for a VM named samplevm1:

      Copy
      Copied!
                  

      Xserver unix:0, GPU maximum memory 4173824KB pid 21859, VM samplevm1, reserved 131072KB of GPU memory. GPU memory left 4042752KB.

      The memory reserved for the VM and the GPU maximum memory depend on the GPU installed in the host and the 3D memory allocated to the virtual machine.

Installation of the NVIDIA vGPU software graphics driver for the guest OS is not required for vSGA.

2.10. Configuring the vGPU Manager for a Linux with KVM Hypervisor

NVIDIA vGPU software supports the following Linux with KVM hypervisors: Red Hat Enterprise Linux with KVM and Ubuntu.

2.10.1. Getting the BDF and Domain of a GPU on a Linux with KVM Hypervisor

Sometimes when configuring a physical GPU for use with NVIDIA vGPU software, you must find out which directory in the sysfs file system represents the GPU. This directory is identified by the domain, bus, slot, and function of the GPU.

For more information about the directory in the sysfs file system that represents a physical GPU, see NVIDIA vGPU Information in the sysfs File System.

  1. Obtain the PCI device bus/device/function (BDF) of the physical GPU.
    Copy
    Copied!
                

    # lspci | grep NVIDIA

    The NVIDIA GPUs listed in this example have the PCI device BDFs 06:00.0 and 07:00.0.

    Copy
    Copied!
                

    # lspci | grep NVIDIA 06:00.0 VGA compatible controller: NVIDIA Corporation GM204GL [Tesla M10] (rev a1) 07:00.0 VGA compatible controller: NVIDIA Corporation GM204GL [Tesla M10] (rev a1)

  2. Obtain the full identifier of the GPU from its PCI device BDF.
    Copy
    Copied!
                

    # virsh nodedev-list --cap pci| grep transformed-bdf

    transformed-bdf
    The PCI device BDF of the GPU with the colon and the period replaced with underscores, for example, 06_00_0.

    This example obtains the full identifier of the GPU with the PCI device BDF 06:00.0.

    Copy
    Copied!
                

    # virsh nodedev-list --cap pci| grep 06_00_0 pci_0000_06_00_0

  3. Obtain the domain, bus, slot, and function of the GPU from the full identifier of the GPU.
    Copy
    Copied!
                

    virsh nodedev-dumpxml full-identifier| egrep 'domain|bus|slot|function'

    full-identifier
    The full identifier of the GPU that you obtained in the previous step, for example, pci_0000_06_00_0.

    This example obtains the domain, bus, slot, and function of the GPU with the PCI device BDF 06:00.0.

    Copy
    Copied!
                

    # virsh nodedev-dumpxml pci_0000_06_00_0| egrep 'domain|bus|slot|function' <domain>0x0000</domain> <bus>0x06</bus> <slot>0x00</slot> <function>0x0</function> <address domain='0x0000' bus='0x06' slot='0x00' function='0x0'/>

2.10.2. Creating an NVIDIA vGPU on a Linux with KVM Hypervisor

For each vGPU that you want to create, perform this task in a Linux command shell on the a Linux with KVM hypervisor host.

Before you begin, ensure that you have the domain, bus, slot, and function of the GPU on which you are creating the vGPU. For instructions, see Getting the BDF and Domain of a GPU on a Linux with KVM Hypervisor.

How to create an NVIDIA vGPU on a Linux with KVM hypervisor depends on whether the NVIDIA vGPU supports single root I/O virtualization (SR-IOV). For details, refer to:

2.10.2.1. Creating a Legacy NVIDIA vGPU on a Linux with KVM Hypervisor

A legacy NVIDIA vGPU does not support SR-IOV.

  1. Change to the mdev_supported_types directory for the physical GPU.
    Copy
    Copied!
                

    # cd /sys/class/mdev_bus/domain\:bus\:slot.function/mdev_supported_types/

    domain
    bus
    slot
    function
    The domain, bus, slot, and function of the GPU, without the 0x prefix.

    This example changes to the mdev_supported_types directory for the GPU with the domain 0000 and PCI device BDF 06:00.0.

    Copy
    Copied!
                

    # cd /sys/bus/pci/devices/0000\:06\:00.0/mdev_supported_types/

  2. Find out which subdirectory of mdev_supported_types contains registration information for the vGPU type that you want to create.
    Copy
    Copied!
                

    # grep -l "vgpu-type" nvidia-*/name

    vgpu-type
    The vGPU type, for example, M10-2Q.

    This example shows that the registration information for the M10-2Q vGPU type is contained in the nvidia-41 subdirectory of mdev_supported_types.

    Copy
    Copied!
                

    # grep -l "M10-2Q" nvidia-*/name nvidia-41/name

  3. Confirm that you can create an instance of the vGPU type on the physical GPU.
    Copy
    Copied!
                

    # cat subdirectory/available_instances

    subdirectory
    The subdirectory that you found in the previous step, for example, nvidia-41.

    The number of available instances must be at least 1. If the number is 0, either an instance of another vGPU type already exists on the physical GPU, or the maximum number of allowed instances has already been created.

    This example shows that four more instances of the M10-2Q vGPU type can be created on the physical GPU.

    Copy
    Copied!
                

    # cat nvidia-41/available_instances 4

  4. Generate a correctly formatted universally unique identifier (UUID) for the vGPU.
    Copy
    Copied!
                

    # uuidgen aa618089-8b16-4d01-a136-25a0f3c73123

  5. Write the UUID that you obtained in the previous step to the create file in the registration information directory for the vGPU type that you want to create.
    Copy
    Copied!
                

    # echo "uuid"> subdirectory/create

    uuid
    The UUID that you generated in the previous step, which will become the UUID of the vGPU that you want to create.
    subdirectory
    The registration information directory for the vGPU type that you want to create, for example, nvidia-41.

    This example creates an instance of the M10-2Q vGPU type with the UUID aa618089-8b16-4d01-a136-25a0f3c73123.

    Copy
    Copied!
                

    # echo "aa618089-8b16-4d01-a136-25a0f3c73123" > nvidia-41/create

    An mdev device file for the vGPU is added to the parent physical device directory of the vGPU. The vGPU is identified by its UUID.

    The /sys/bus/mdev/devices/ directory contains a symbolic link to the mdev device file.

  6. Make the mdev device file that you created to represent the vGPU persistent.
    Copy
    Copied!
                

    # mdevctl define --auto --uuid uuid

    uuid
    The UUID that you specified in the previous step for the vGPU that you are creating.
    Note:

    Not all Linux with KVM hypervisor releases include the mdevctl command. If your release does not include the mdevctl command, you can use standard features of the operating system to automate the re-creation of this device file when the host is booted. For example, you can write a custom script that is executed when the host is rebooted.

  7. Confirm that the vGPU was created.
    1. Confirm that the /sys/bus/mdev/devices/ directory contains the mdev device file for the vGPU.
      Copy
      Copied!
                  

      # ls -l /sys/bus/mdev/devices/ total 0 lrwxrwxrwx. 1 root root 0 Nov 24 13:33 aa618089-8b16-4d01-a136-25a0f3c73123 -> ../../../devices/pci0000:00/0000:00:03.0/0000:03:00.0/0000:04:09.0/0000:06:00.0/aa618089-8b16-4d01-a136-25a0f3c73123

    2. If your release includes the mdevctl command, list the active mediated devices on the hypervisor host.
      Copy
      Copied!
                  

      # mdevctl list aa618089-8b16-4d01-a136-25a0f3c73123 0000:06:00.0 nvidia-41

2.10.2.2. Creating an NVIDIA vGPU that Supports SR-IOV on a Linux with KVM Hypervisor

An NVIDIA vGPU that supports SR-IOV resides on a physical GPU that supports SR-IOV, such as a GPU based on the NVIDIA Ampere architecture.

  1. Enable the virtual functions for the physical GPU in the sysfs file system.
    Note:
    • Before performing this step, ensure that the GPU is not being used by any other processes, such as CUDA applications, monitoring applications, or the nvidia-smi command.
    • The virtual functions for the physical GPU in the sysfs file system are disabled after the hypervisor host is rebooted or if the driver is reloaded or upgraded.

    Use only the custom script sriov-manage provided by NVIDIA vGPU software for this purpose. Do not try to enable the virtual function for the GPU by any other means.

    Copy
    Copied!
                

    # /usr/lib/nvidia/sriov-manage -e domain:bus:slot.function

    domain
    bus
    slot
    function
    The domain, bus, slot, and function of the GPU, without the 0x prefix.
    Note:

    Only one mdev device file can be created on a virtual function.

    This example enables the virtual functions for the GPU with the domain 00, bus 41, slot 0000, and function 0.

    Copy
    Copied!
                

    # /usr/lib/nvidia/sriov-manage -e 00:41:0000.0

  2. Obtain the domain, bus, slot, and function of the available virtual functions on the GPU.
    Copy
    Copied!
                

    # ls -l /sys/bus/pci/devices/domain\:bus\:slot.function/ | grep virtfn

    domain
    bus
    slot
    function
    The domain, bus, slot, and function of the GPU, without the 0x prefix.

    This example shows the output of this command for a physical GPU with slot 00, bus 41, domain 0000, and function 0.

    Copy
    Copied!
                

    # ls -l /sys/bus/pci/devices/0000:41:00.0/ | grep virtfn lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn0 -> ../0000:41:00.4 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn1 -> ../0000:41:00.5 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn10 -> ../0000:41:01.6 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn11 -> ../0000:41:01.7 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn12 -> ../0000:41:02.0 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn13 -> ../0000:41:02.1 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn14 -> ../0000:41:02.2 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn15 -> ../0000:41:02.3 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn16 -> ../0000:41:02.4 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn17 -> ../0000:41:02.5 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn18 -> ../0000:41:02.6 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn19 -> ../0000:41:02.7 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn2 -> ../0000:41:00.6 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn20 -> ../0000:41:03.0 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn21 -> ../0000:41:03.1 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn22 -> ../0000:41:03.2 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn23 -> ../0000:41:03.3 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn24 -> ../0000:41:03.4 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn25 -> ../0000:41:03.5 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn26 -> ../0000:41:03.6 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn27 -> ../0000:41:03.7 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn28 -> ../0000:41:04.0 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn29 -> ../0000:41:04.1 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn3 -> ../0000:41:00.7 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn30 -> ../0000:41:04.2 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn31 -> ../0000:41:04.3 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn4 -> ../0000:41:01.0 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn5 -> ../0000:41:01.1 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn6 -> ../0000:41:01.2 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn7 -> ../0000:41:01.3 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn8 -> ../0000:41:01.4 lrwxrwxrwx. 1 root root 0 Jul 16 04:42 virtfn9 -> ../0000:41:01.5

  3. Choose the available virtual function on which you want to create the vGPU and note its domain, bus, slot, and function.
  4. Change to the mdev_supported_types directory for the virtual function on which you want to create the vGPU.
    Copy
    Copied!
                

    # cd /sys/class/mdev_bus/domain\:bus\:vf-slot.v-function/mdev_supported_types/

    domain
    bus
    The domain and bus of the GPU, without the 0x prefix.
    vf-slot
    v-function
    The slot and function of the virtual function.

    This example changes to the mdev_supported_types directory for the first virtual function (virtfn0) for the GPU with the domain 0000 and bus 41. The first virtual function (virtfn0) has slot 00 and function 4.

    Copy
    Copied!
                

    # cd /sys/class/mdev_bus/0000\:41\:00.4/mdev_supported_types

  5. Find out which subdirectory of mdev_supported_types contains registration information for the vGPU type that you want to create.
    Copy
    Copied!
                

    # grep -l "vgpu-type" nvidia-*/name

    vgpu-type
    The vGPU type, for example, A40-2Q.

    This example shows that the registration information for the A40-2Q vGPU type is contained in the nvidia-558 subdirectory of mdev_supported_types.

    Copy
    Copied!
                

    # grep -l "A40-2Q" nvidia-*/name nvidia-558/name

  6. Confirm that you can create an instance of the vGPU type on the virtual function.
    Copy
    Copied!
                

    # cat subdirectory/available_instances

    subdirectory
    The subdirectory that you found in the previous step, for example, nvidia-558.

    The number of available instances must be 1. If the number is 0, a vGPU has already been created on the virtual function. Only one instance of any vGPU type can be created on a virtual function.

    This example shows that an instance of the A40-2Q vGPU type can be created on the virtual function.

    Copy
    Copied!
                

    # cat nvidia-558/available_instances 1

  7. Generate a correctly formatted universally unique identifier (UUID) for the vGPU.
    Copy
    Copied!
                

    # uuidgen aa618089-8b16-4d01-a136-25a0f3c73123

  8. Write the UUID that you obtained in the previous step to the create file in the registration information directory for the vGPU type that you want to create.
    Copy
    Copied!
                

    # echo "uuid"> subdirectory/create

    uuid
    The UUID that you generated in the previous step, which will become the UUID of the vGPU that you want to create.
    subdirectory
    The registration information directory for the vGPU type that you want to create, for example, nvidia-558.

    This example creates an instance of the A40-2Q vGPU type with the UUID aa618089-8b16-4d01-a136-25a0f3c73123.

    Copy
    Copied!
                

    # echo "aa618089-8b16-4d01-a136-25a0f3c73123" > nvidia-558/create

    An mdev device file for the vGPU is added to the parent virtual function directory of the vGPU. The vGPU is identified by its UUID.

  9. Time-sliced vGPUs only: Make the mdev device file that you created to represent the vGPU persistent.
    Copy
    Copied!
                

    # mdevctl define --auto --uuid uuid

    uuid
    The UUID that you specified in the previous step for the vGPU that you are creating.
    Note:
    • If you are using a GPU that supports SR-IOV, the mdev device file persists after a host reboot only if you perform Step 1 before rebooting any VM that is configured with a vGPU on the GPU.
    • You cannot use the mdevctl command to make the mdev device file for a MIG-backed vGPU persistent. The mdev device file for a MIG-backed vGPU is not retained after the host is rebooted because MIG instances are no longer available.
    • Not all Linux with KVM hypervisor releases include the mdevctl command. If your release does not include the mdevctl command, you can use standard features of the operating system to automate the re-creation of this device file when the host is booted. For example, you can write a custom script that is executed when the host is rebooted.

  10. Confirm that the vGPU was created.
    1. Confirm that the /sys/bus/mdev/devices/ directory contains a symbolic link to the mdev device file.
      Copy
      Copied!
                  

      # ls -l /sys/bus/mdev/devices/ total 0 lrwxrwxrwx. 1 root root 0 Jul 16 05:57 aa618089-8b16-4d01-a136-25a0f3c73123 -> ../../../devices/pci0000:40/0000:40:01.1/0000:41:00.4/aa618089-8b16-4d01-a136-25a0f3c73123

    2. If your release includes the mdevctl command, list the active mediated devices on the hypervisor host.
      Copy
      Copied!
                  

      # mdevctl list aa618089-8b16-4d01-a136-25a0f3c73123 0000:06:00.0 nvidia-558

2.10.3. Adding One or More vGPUs to a Linux with KVM Hypervisor VM

To support applications and workloads that are compute or graphics intensive, you can add multiple vGPUs to a single VM.

For details about which hypervisor versions and NVIDIA vGPUs support the assignment of multiple vGPUs to a VM, see Virtual GPU Software for Red Hat Enterprise Linux with KVM Release Notes and Virtual GPU Software for Ubuntu Release Notes.
Ensure that the following prerequisites are met:

You can add vGPUs to a Linux with KVM hypervisor VM by using any of the following tools:

  • The virsh command
  • The QEMU command line

After adding vGPUs to a Linux with KVM hypervisor VM, start the VM.

Copy
Copied!
            

# virsh start vm-name

vm-name
The name of the VM that you added the vGPUs to.

After the VM has booted, install the NVIDIA vGPU software graphics driver as explained in Installing the NVIDIA vGPU Software Graphics Driver.

2.10.3.1. Adding One or More vGPUs to a Linux with KVM Hypervisor VM by Using virsh

  1. In virsh, open for editing the XML file of the VM that you want to add the vGPU to.
    Copy
    Copied!
                

    # virsh edit vm-name

    vm-name
    The name of the VM to that you want to add the vGPUs to.
  2. For each vGPU that you want to add to the VM, add a device entry in the form of an address element inside the source element to add the vGPU to the guest VM.
    Copy
    Copied!
                

    <device> ... <hostdev mode='subsystem' type='mdev' model='vfio-pci'> <source> <address uuid='uuid'/> </source> </hostdev> </device>

    uuid
    The UUID that was assigned to the vGPU when the vGPU was created.

    This example adds a device entry for the vGPU with the UUID a618089-8b16-4d01-a136-25a0f3c73123.

    Copy
    Copied!
                

    <device> ... <hostdev mode='subsystem' type='mdev' model='vfio-pci'> <source> <address uuid='a618089-8b16-4d01-a136-25a0f3c73123'/> </source> </hostdev> </device>

    This example adds device entries for two vGPUs with the following UUIDs:

    • c73f1fa6-489e-4834-9476-d70dabd98c40
    • 3b356d38-854e-48be-b376-00c72c7d119c
    Copy
    Copied!
                

    <device> ... <hostdev mode='subsystem' type='mdev' model='vfio-pci'> <source> <address uuid='c73f1fa6-489e-4834-9476-d70dabd98c40'/> </source> </hostdev> <hostdev mode='subsystem' type='mdev' model='vfio-pci'> <source> <address uuid='3b356d38-854e-48be-b376-00c72c7d119c'/> </source> </hostdev> </device>

  3. Optional: Add a video element that contains a model element in which the type attribute is set to none.
    Copy
    Copied!
                

    <video> <model type='none'/> </video>

    Adding this video element prevents the default video device that libvirt adds from being loaded into the VM. If you don't add this video element, you must configure the Xorg server or your remoting solution to load only the vGPU devices you added and not the default video device.

2.10.3.2. Adding One or More vGPUs to a Linux with KVM Hypervisor VM by Using the QEMU Command Line

Add the following options to the QEMU command line:

  • For each vGPU that you want to add to the VM, add one -device option in the following format:

    Copy
    Copied!
                

    -device vfio-pci,sysfsdev=/sys/bus/mdev/devices/vgpu-uuid

    vgpu-uuid
    The UUID that was assigned to the vGPU when the vGPU was created.
  • Add a -uuid option to specify the VM as follows:

    Copy
    Copied!
                

    -uuid vm-uuid

    vm-uuid
    The UUID that was assigned to the VM when the VM was created.

This example adds the vGPU with the UUID aa618089-8b16-4d01-a136-25a0f3c73123 to the VM with the UUID ebb10a6e-7ac9-49aa-af92-f56bb8c65893.

Copy
Copied!
            

-device vfio-pci,sysfsdev=/sys/bus/mdev/devices/aa618089-8b16-4d01-a136-25a0f3c73123 \ -uuid ebb10a6e-7ac9-49aa-af92-f56bb8c65893

This example adds device entries for two vGPUs with the following UUIDs:

  • 676428a0-2445-499f-9bfd-65cd4a9bd18f
  • 6c5954b8-5bc1-4769-b820-8099fe50aaba

The entries are added to the VM with the UUID ec5e8ee0-657c-4db6-8775-da70e332c67e.

Copy
Copied!
            

-device vfio-pci,sysfsdev=/sys/bus/mdev/devices/676428a0-2445-499f-9bfd-65cd4a9bd18f \ -device vfio-pci,sysfsdev=/sys/bus/mdev/devices/6c5954b8-5bc1-4769-b820-8099fe50aaba \ -uuid ec5e8ee0-657c-4db6-8775-da70e332c67e

2.10.4. Setting vGPU Plugin Parameters on a Linux with KVM Hypervisor

Plugin parameters for a vGPU control the behavior of the vGPU, such as the frame rate limiter (FRL) configuration in frames per second or whether console virtual network computing (VNC) for the vGPU is enabled. The VM to which the vGPU is assigned is started with these parameters. If parameters are set for multiple vGPUs assigned to the same VM, the VM is started with the parameters assigned to each vGPU.

For each vGPU for which you want to set plugin parameters, perform this task in a Linux command shell on the Linux with KVM hypervisor host.

  1. Change to the nvidia subdirectory of the mdev device directory that represents the vGPU.
    Copy
    Copied!
                

    # cd /sys/bus/mdev/devices/uuid/nvidia

    uuid
    The UUID of the vGPU, for example, aa618089-8b16-4d01-a136-25a0f3c73123.
  2. Write the plugin parameters that you want to set to the vgpu_params file in the directory that you changed to in the previous step.
    Copy
    Copied!
                

    # echo "plugin-config-params" > vgpu_params

    plugin-config-params
    A comma-separated list of parameter-value pairs, where each pair is of the form parameter-name=value.

    This example disables frame rate limiting and console VNC for a vGPU.

    Copy
    Copied!
                

    # echo "frame_rate_limiter=0, disable_vnc=1" > vgpu_params

    This example enables unified memory for a vGPU.

    Copy
    Copied!
                

    # echo "enable_uvm=1" > vgpu_params

    This example enables NVIDIA CUDA Toolkit debuggers for a vGPU.

    Copy
    Copied!
                

    # echo "enable_debugging=1" > vgpu_params

    This example enables NVIDIA CUDA Toolkit profilers for a vGPU.

    Copy
    Copied!
                

    # echo "enable_profiling=1" > vgpu_params

To clear any vGPU plugin parameters that were set previously, write a space to the vgpu_params file for the vGPU.

Copy
Copied!
            

# echo " " > vgpu_params

2.10.5. Deleting a vGPU on a Linux with KVM Hypervisor

For each vGPU that you want to delete, perform this task in a Linux command shell on the Linux with KVM hypervisor host.

Before you begin, ensure that the following prerequisites are met:

  1. Change to the mdev_supported_types directory for the physical GPU.
    Copy
    Copied!
                

    # cd /sys/class/mdev_bus/domain\:bus\:slot.function/mdev_supported_types/

    domain
    bus
    slot
    function
    The domain, bus, slot, and function of the GPU, without the 0x prefix.

    This example changes to the mdev_supported_types directory for the GPU with the PCI device BDF 06:00.0.

    Copy
    Copied!
                

    # cd /sys/bus/pci/devices/0000\:06\:00.0/mdev_supported_types/

  2. Change to the subdirectory of mdev_supported_types that contains registration information for the vGPU.
    Copy
    Copied!
                

    # cd `find . -type d -name uuid`

    uuid
    The UUID of the vGPU, for example, aa618089-8b16-4d01-a136-25a0f3c73123.
  3. Write the value 1 to the remove file in the registration information directory for the vGPU that you want to delete.
    Copy
    Copied!
                

    # echo "1" > remove

    Note:

    On the Red Hat Virtualization (RHV) kernel, if you try to remove a vGPU device while its VM is running, the vGPU device might not be removed even if the remove file has been written to successfully. To confirm that the vGPU device is removed, confirm that the UUID of the vGPU is not found in the sysfs file system.

2.10.6. Preparing a GPU Configured for Pass-Through for Use with vGPU

The mode in which a physical GPU is being used determines the Linux kernel module to which the GPU is bound. If you want to switch the mode in which a GPU is being used, you must unbind the GPU from its current kernel module and bind it to the kernel module for the new mode. After binding the GPU to the correct kernel module, you can then configure it for vGPU.

A physical GPU that is passed through to a VM is bound to the vfio-pci kernel module. A physical GPU that is bound to the vfio-pci kernel module can be used only for pass-through. To enable the GPU to be used for vGPU, the GPU must be unbound from vfio-pci kernel module and bound to the nvidia kernel module.
Before you begin, ensure that you have the domain, bus, slot, and function of the GPU that you are preparing for use with vGPU. For instructions, see Getting the BDF and Domain of a GPU on a Linux with KVM Hypervisor.

  1. Determine the kernel module to which the GPU is bound by running the lspci command with the -k option on the NVIDIA GPUs on your host.
    Copy
    Copied!
                

    # lspci -d 10de: -k

    The Kernel driver in use: field indicates the kernel module to which the GPU is bound.

    The following example shows that the NVIDIA Tesla M60 GPU with BDF06:00.0 is bound to the vfio-pci kernel module and is being used for GPU pass through.

    Copy
    Copied!
                

    06:00.0 VGA compatible controller: NVIDIA Corporation GM204GL [Tesla M60] (rev a1) Subsystem: NVIDIA Corporation Device 115e Kernel driver in use: vfio-pci

  2. Unbind the GPU from vfio-pci kernel module.
    1. Change to the sysfs directory that represents the vfio-pci kernel module.
      Copy
      Copied!
                  

      # cd /sys/bus/pci/drivers/vfio-pci

    2. Write the domain, bus, slot, and function of the GPU to the unbind file in this directory.
      Copy
      Copied!
                  

      # echo domain:bus:slot.function > unbind

      domain
      bus
      slot
      function
      The domain, bus, slot, and function of the GPU, without a 0x prefix.

      This example writes the domain, bus, slot, and function of the GPU with the domain 0000 and PCI device BDF 06:00.0.

      Copy
      Copied!
                  

      # echo 0000:06:00.0 > unbind

  3. Bind the GPU to the nvidia kernel module.
    1. Change to the sysfs directory that contains the PCI device information for the physical GPU.
      Copy
      Copied!
                  

      # cd /sys/bus/pci/devices/domain\:bus\:slot.function

      domain
      bus
      slot
      function
      The domain, bus, slot, and function of the GPU, without a 0x prefix.

      This example changes to the sysfs directory that contains the PCI device information for the GPU with the domain 0000 and PCI device BDF 06:00.0.

      Copy
      Copied!
                  

      # cd /sys/bus/pci/devices/0000\:06\:00.0

    2. Write the kernel module name nvidia to the driver_override file in this directory.
      Copy
      Copied!
                  

      # echo nvidia > driver_override

    3. Change to the sysfs directory that represents the nvidia kernel module.
      Copy
      Copied!
                  

      # cd /sys/bus/pci/drivers/nvidia

    4. Write the domain, bus, slot, and function of the GPU to the bind file in this directory.
      Copy
      Copied!
                  

      # echo domain:bus:slot.function > bind

      domain
      bus
      slot
      function
      The domain, bus, slot, and function of the GPU, without a 0x prefix.

      This example writes the domain, bus, slot, and function of the GPU with the domain 0000 and PCI device BDF 06:00.0.

      Copy
      Copied!
                  

      # echo 0000:06:00.0 > bind

You can now configure the GPU with vGPU as explained in Installing and Configuring the NVIDIA Virtual GPU Manager for Red Hat Enterprise Linux KVM or RHV.

2.10.7. NVIDIA vGPU Information in the sysfs File System

Information about the NVIDIA vGPU types supported by each physical GPU in a Linux with KVM hypervisor host is stored in the sysfs file system.

All physical GPUs on the host are registered with the mdev kernel module. Information about the physical GPUs and the vGPU types that can be created on each physical GPU is stored in directories and files under the /sys/class/mdev_bus/ directory.

The sysfs directory for each physical GPU is at the following locations:

  • /sys/bus/pci/devices/
  • /sys/class/mdev_bus/

Both directories are a symbolic link to the real directory for PCI devices in the sysfs file system.

The organization the sysfs directory for each physical GPU is as follows:

Copy
Copied!
            

/sys/class/mdev_bus/ |-parent-physical-device |-mdev_supported_types |-nvidia-vgputype-id |-available_instances |-create |-description |-device_api |-devices |-name


parent-physical-device

Each physical GPU on the host is represented by a subdirectory of the /sys/class/mdev_bus/ directory.

The name of each subdirectory is as follows:

domain\:bus\:slot.function

domain, bus, slot, function are the domain, bus, slot, and function of the GPU, for example, 0000\:06\:00.0.

Each directory is a symbolic link to the real directory for PCI devices in the sysfs file system. For example:

Copy
Copied!
            

# ll /sys/class/mdev_bus/ total 0 lrwxrwxrwx. 1 root root 0 Dec 12 03:20 0000:05:00.0 -> ../../devices/pci0000:00/0000:00:03.0/0000:03:00.0/0000:04:08.0/0000:05:00.0 lrwxrwxrwx. 1 root root 0 Dec 12 03:20 0000:06:00.0 -> ../../devices/pci0000:00/0000:00:03.0/0000:03:00.0/0000:04:09.0/0000:06:00.0 lrwxrwxrwx. 1 root root 0 Dec 12 03:20 0000:07:00.0 -> ../../devices/pci0000:00/0000:00:03.0/0000:03:00.0/0000:04:10.0/0000:07:00.0 lrwxrwxrwx. 1 root root 0 Dec 12 03:20 0000:08:00.0 -> ../../devices/pci0000:00/0000:00:03.0/0000:03:00.0/0000:04:11.0/0000:08:00.0

mdev_supported_types
A directory named mdev_supported_types is required under the sysfs directory for each physical GPU that will be configured with NVIDIA vGPU. How this directory is created for a GPU depends on whether the GPU supports SR-IOV.
  • For a GPU that does not support SR-IOV, this directory is created automatically after the Virtual GPU Manager is installed on the host and the host has been rebooted.
  • For a GPU that supports SR-IOV, such as a GPU based on the NVIDIA Ampere architecture, you must create this directory by enabling the virtual function for the GPU as explained in Creating an NVIDIA vGPU on a Linux with KVM Hypervisor. The mdev_supported_types directory itself is never visible on the physical function.

The mdev_supported_types directory contains a subdirectory for each vGPU type that the physical GPU supports. The name of each subdirectory is nvidia-vgputype-id, where vgputype-id is an unsigned integer serial number. For example:

Copy
Copied!
            

# ll mdev_supported_types/ total 0 drwxr-xr-x 3 root root 0 Dec 6 01:37 nvidia-35 drwxr-xr-x 3 root root 0 Dec 5 10:43 nvidia-36 drwxr-xr-x 3 root root 0 Dec 5 10:43 nvidia-37 drwxr-xr-x 3 root root 0 Dec 5 10:43 nvidia-38 drwxr-xr-x 3 root root 0 Dec 5 10:43 nvidia-39 drwxr-xr-x 3 root root 0 Dec 5 10:43 nvidia-40 drwxr-xr-x 3 root root 0 Dec 5 10:43 nvidia-41 drwxr-xr-x 3 root root 0 Dec 5 10:43 nvidia-42 drwxr-xr-x 3 root root 0 Dec 5 10:43 nvidia-43 drwxr-xr-x 3 root root 0 Dec 5 10:43 nvidia-44 drwxr-xr-x 3 root root 0 Dec 5 10:43 nvidia-45

nvidia-vgputype-id
Each directory represents an individual vGPU type and contains the following files and directories:
available_instances
This file contains the number of instances of this vGPU type that can still be created. This file is updated any time a vGPU of this type is created on or removed from the physical GPU.
Note:

When a time-sliced vGPU is created, the content of the available_instances for all other time-sliced vGPU types on the physical GPU is set to 0. This behavior enforces the requirement that all time-sliced vGPUs on a physical GPU must be of the same type. However, this requirement does not apply to MIG-backed vGPUs. Therefore, when a MIG-backed vGPU is created, available_instances for all other MIG-backed vGPU types on the physical GPU is not set to 0

create
This file is used for creating a vGPU instance. A vGPU instance is created by writing the UUID of the vGPU to this file. The file is write only.
description
This file contains the following details of the vGPU type:
  • The maximum number of virtual display heads that the vGPU type supports
  • The frame rate limiter (FRL) configuration in frames per second
  • The frame buffer size in Mbytes
  • The maximum resolution per display head
  • The maximum number of vGPU instances per physical GPU

For example:

Copy
Copied!
            

# cat description num_heads=4, frl_config=60, framebuffer=2048M, max_resolution=4096x2160, max_instance=4

device_api
This file contains the string vfio_pci to indicate that a vGPU is a PCI device.
devices
This directory contains all the mdev devices that are created for the vGPU type. For example:
Copy
Copied!
            

# ll devices total 0 lrwxrwxrwx 1 root root 0 Dec 6 01:52 aa618089-8b16-4d01-a136-25a0f3c73123 -> ../../../aa618089-8b16-4d01-a136-25a0f3c73123

name
This file contains the name of the vGPU type. For example:
Copy
Copied!
            

# cat name GRID M10-2Q

2.11. Configuring a GPU for MIG-Backed vGPUs

To support GPU instances with NVIDIA vGPU, a GPU must be configured with MIG mode enabled and GPU instances must be created and configured on the physical GPU. Optionally, you can create compute instances within the GPU instances. If you don't create compute instances within the GPU instances, they can be added later for individual vGPUs from within the guest VMs.

Ensure that the following prerequisites are met:

  • The NVIDIA Virtual GPU Manager is installed on the hypervisor host.
  • You have root user privileges on your hypervisor host machine.
  • You have determined which GPU instances correspond to the vGPU types of the MIG-backed vGPUs that you will create.

    To get this information, consult the table of MIG-backed vGPUs for your GPU in Virtual GPU Types for Supported GPUs.

  • The GPU is not being used by any other processes, such as CUDA applications, monitoring applications, or the nvidia-smi command.

To configure a GPU for MIG-backed vGPUs, follow these instructions:

  1. Enabling MIG Mode for a GPU

    Note:

    For VMware vSphere, only enabling MIG mode is required because VMware vSphere creates the GPU instances and, after the VM is booted and guest driver is installed, one compute instance is automatically created in the VM.

  2. Creating GPU Instances on a MIG-Enabled GPU

  3. Optional: Creating Compute Instances in a GPU instance

After configuring a GPU for MIG-backed vGPUs, create the vGPUs that you need and add them to their VMs.

2.11.1. Enabling MIG Mode for a GPU

Perform this task in your hypervisor command shell.

  1. Open a command shell as the root user on your hypervisor host machine. On all supported hypervisors, you can use secure shell (SSH) for this purpose. Individual hypervisors may provide additional means for logging in. For details, refer to the documentation for your hypervisor.
  2. Determine whether MIG mode is enabled. Use the nvidia-smi command for this purpose. By default, MIG mode is disabled.

    This example shows that MIG mode is disabled on GPU 0.

    Note:

    In the output from nvidia-smi, the NVIDIA A100 HGX 40GB GPU is referred to as A100-SXM4-40GB.

    Copy
    Copied!
                

    $ nvidia-smi -i 0 +-----------------------------------------------------------------------------+ | NVIDIA-SMI 470.256.02 Driver Version: 470.256.02 CUDA Version: 11.4 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |===============================+======================+======================| | 0 A100-SXM4-40GB On | 00000000:36:00.0 Off | 0 | | N/A 29C P0 62W / 400W | 0MiB / 40537MiB | 6% Default | | | | Disabled | +-------------------------------+----------------------+----------------------+

  3. If MIG mode is disabled, enable it.
    Copy
    Copied!
                

    $ nvidia-smi -i [gpu-ids] -mig 1

    gpu-ids
    A comma-separated list of GPU indexes, PCI bus IDs or UUIDs that specifies the GPUs on which you want to enable MIG mode. If gpu-ids is omitted, MIG mode is enabled on all GPUs on the system.

    This example enables MIG mode on GPU 0.

    Copy
    Copied!
                

    $ nvidia-smi -i 0 -mig 1 Enabled MIG Mode for GPU 00000000:36:00.0 All done.

    Note:

    If the GPU is being used by another process, this command fails and displays a warning message that MIG mode for the GPU is in the pending enable state. In this situation, stop all processes that are using the GPU and retry the command.

  4. VMware vSphere ESXi only: Reboot the hypervisor host.
  5. Query the GPUs on which you enabled MIG mode to confirm that MIG mode is enabled.

    This example queries GPU 0 for the PCI bus ID and MIG mode in comma-separated values (CSV) format.

    Copy
    Copied!
                

    $ nvidia-smi -i 0 --query-gpu=pci.bus_id,mig.mode.current --format=csv pci.bus_id, mig.mode.current 00000000:36:00.0, Enabled

2.11.2. Creating GPU Instances on a MIG-Enabled GPU

Note:

If you are using VMware vSphere, omit this task. VMware vSphere creates the GPU instances automatically.

Perform this task in your hypervisor command shell.

  1. If necessary, open a command shell as the root user on your hypervisor host machine.
  2. List the GPU instance profiles that are available on your GPU.

    You will need to specify the profiles by their IDs, not their names, when you create them.

    Copy
    Copied!
                

    $ nvidia-smi mig -lgip +--------------------------------------------------------------------------+ | GPU instance profiles: | | GPU Name ID Instances Memory P2P SM DEC ENC | | Free/Total GiB CE JPEG OFA | |==========================================================================| | 0 MIG 1g.5gb 19 7/7 4.95 No 14 0 0 | | 1 0 0 | +--------------------------------------------------------------------------+ | 0 MIG 2g.10gb 14 3/3 9.90 No 28 1 0 | | 2 0 0 | +--------------------------------------------------------------------------+ | 0 MIG 3g.20gb 9 2/2 19.79 No 42 2 0 | | 3 0 0 | +--------------------------------------------------------------------------+ | 0 MIG 4g.20gb 5 1/1 19.79 No 56 2 0 | | 4 0 0 | +--------------------------------------------------------------------------+ | 0 MIG 7g.40gb 0 1/1 39.59 No 98 5 0 | | 7 1 1 | +--------------------------------------------------------------------------+

  3. Create the GPU instances that correspond to the vGPU types of the MIG-backed vGPUs that you will create.
    Copy
    Copied!
                

    $ nvidia-smi mig -cgi gpu-instance-profile-ids

    gpu-instance-profile-ids
    A comma-separated list of GPU instance profile IDs that specifies the GPU instances that you want to create.

    This example creates two GPU instances of type 2g.10gb, which has profile ID 14.

    Copy
    Copied!
                

    $ nvidia-smi mig -cgi 14,14 Successfully created GPU instance ID 5 on GPU 2 using profile MIG 2g.10gb (ID 14) Successfully created GPU instance ID 3 on GPU 2 using profile MIG 2g.10gb (ID 14)

2.11.3. Optional: Creating Compute Instances in a GPU instance

Creating compute instances within GPU instances is optional. If you don't create compute instances within the GPU instances, they can be added later for individual vGPUs from within the guest VMs.

Note:

If you are using VMware vSphere, omit this task. After the VM is booted and guest driver is installed, one compute instance is automatically created in the VM.

Perform this task in your hypervisor command shell.

  1. If necessary, open a command shell as the root user on your hypervisor host machine.
  2. List the available GPU instances.
    Copy
    Copied!
                

    $ nvidia-smi mig -lgi +----------------------------------------------------+ | GPU instances: | | GPU Name Profile Instance Placement | | ID ID Start:Size | |====================================================| | 2 MIG 2g.10gb 14 3 0:2 | +----------------------------------------------------+ | 2 MIG 2g.10gb 14 5 4:2 | +----------------------------------------------------+

  3. Create the compute instances that you need within each GPU instance.
    Copy
    Copied!
                

    $ nvidia-smi mig -cci -gi gpu-instance-ids

    gpu-instance-ids
    A comma-separated list of GPU instance IDs that specifies the GPU instances within which you want to create the compute instances.
    CAUTION:

    To avoid an inconsistent state between a guest VM and the hypervisor host, do not create compute instances from the hypervisor on a GPU instance on which an active guest VM is running. Instead, create the compute instances from within the guest VM as explained in Modifying a MIG-Backed vGPU's Configuration.

    This example creates a compute instance on each of GPU instances 3 and 5.

    Copy
    Copied!
                

    $ nvidia-smi mig -cci -gi 3,5 Successfully created compute instance on GPU 0 GPU instance ID 1 using profile ID 2 Successfully created compute instance on GPU 0 GPU instance ID 2 using profile ID 2

  4. Verify that the compute instances were created within each GPU instance.
    Copy
    Copied!
                

    $ nvidia-smi +-----------------------------------------------------------------------------+ | MIG devices: | +------------------+----------------------+-----------+-----------------------+ | GPU GI CI MIG | Memory-Usage | Vol| Shared | | ID ID Dev | BAR1-Usage | SM Unc| CE ENC DEC OFA JPG| | | | ECC| | |==================+======================+===========+=======================| | 2 3 0 0 | 0MiB / 9984MiB | 28 0 | 2 0 1 0 0 | | | 0MiB / 16383MiB | | | +------------------+----------------------+-----------+-----------------------+ | 2 5 0 1 | 0MiB / 9984MiB | 28 0 | 2 0 1 0 0 | | | 0MiB / 16383MiB | | | +------------------+----------------------+-----------+-----------------------+ +-----------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=============================================================================|

    Note:

    Additional compute instances that have been created in a VM are destroyed when the VM is shut down or rebooted. After the shutdown or reboot, only one compute instance remains in the VM. This compute instance is created automatically after the NVIDIA vGPU software graphics driver is installed.

2.12. Disabling MIG Mode for One or More GPUs

If a GPU that you want to use for time-sliced vGPUs or GPU pass through has previously been configured for MIG-backed vGPUs, disable MIG mode on the GPU.

Ensure that the following prerequisites are met:

  • The NVIDIA Virtual GPU Manager is installed on the hypervisor host.
  • You have root user privileges on your hypervisor host machine.
  • The GPU is not being used by any other processes, such as CUDA applications, monitoring applications, or the nvidia-smi command.

Perform this task in your hypervisor command shell.

  1. Open a command shell as the root user on your hypervisor host machine. You can use secure shell (SSH) for this purpose.
  2. Determine whether MIG mode is disabled. Use the nvidia-smi command for this purpose. By default, MIG mode is disabled, but might have previously been enabled.

    This example shows that MIG mode is enabled on GPU 0.

    Note:

    In the output from output from nvidia-smi, the NVIDIA A100 HGX 40GB GPU is referred to as A100-SXM4-40GB.

    Copy
    Copied!
                

    $ nvidia-smi -i 0 +-----------------------------------------------------------------------------+ | NVIDIA-SMI 470.256.02 Driver Version: 470.256.02 CUDA Version: 11.4 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |===============================+======================+======================| | 0 A100-SXM4-40GB Off | 00000000:36:00.0 Off | 0 | | N/A 29C P0 62W / 400W | 0MiB / 40537MiB | 6% Default | | | | Enabled | +-------------------------------+----------------------+----------------------+

  3. If MIG mode is enabled, disable it.
    Copy
    Copied!
                

    $ nvidia-smi -i [gpu-ids] -mig 0

    gpu-ids
    A comma-separated list of GPU indexes, PCI bus IDs or UUIDs that specifies the GPUs on which you want to disable MIG mode. If gpu-ids is omitted, MIG mode is disabled on all GPUs on the system.

    This example disables MIG mode on GPU 0.

    Copy
    Copied!
                

    $ sudo nvidia-smi -i 0 -mig 0 Disabled MIG Mode for GPU 00000000:36:00.0 All done.

  4. Confirm that MIG mode was disabled. Use the nvidia-smi command for this purpose.

    This example shows that MIG mode is disabled on GPU 0.

    Copy
    Copied!
                

    $ nvidia-smi -i 0 +-----------------------------------------------------------------------------+ | NVIDIA-SMI 470.256.02 Driver Version: 470.256.02 CUDA Version: 11.4 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |===============================+======================+======================| | 0 A100-SXM4-40GB Off | 00000000:36:00.0 Off | 0 | | N/A 29C P0 62W / 400W | 0MiB / 40537MiB | 6% Default | | | | Disabled | +-------------------------------+----------------------+----------------------+

2.13. Disabling and Enabling ECC Memory

Some GPUs that support NVIDIA vGPU software support error correcting code (ECC) memory with NVIDIA vGPU. ECC memory improves data integrity by detecting and handling double-bit errors. However, not all GPUs, vGPU types, and hypervisor software versions support ECC memory with NVIDIA vGPU.

On GPUs that support ECC memory with NVIDIA vGPU, ECC memory is supported with C-series and Q-series vGPUs, but not with A-series and B-series vGPUs. Although A-series and B-series vGPUs start on physical GPUs on which ECC memory is enabled, enabling ECC with vGPUs that do not support it might incur some costs.

On physical GPUs that do not have HBM2 memory, the amount of frame buffer that is usable by vGPUs is reduced. All types of vGPU are affected, not just vGPUs that support ECC memory.

The effects of enabling ECC memory on a physical GPU are as follows:

  • ECC memory is exposed as a feature on all supported vGPUs on the physical GPU.
  • In VMs that support ECC memory, ECC memory is enabled, with the option to disable ECC in the VM.
  • ECC memory can be enabled or disabled for individual VMs. Enabling or disabling ECC memory in a VM does not affect the amount of frame buffer that is usable by vGPUs.

GPUs based on the Pascal GPU architecture and later GPU architectures support ECC memory with NVIDIA vGPU. To determine whether ECC memory is enabled for a GPU, run nvidia-smi -q for the GPU.

Tesla M60 and M6 GPUs support ECC memory when used without GPU virtualization, but NVIDIA vGPU does not support ECC memory with these GPUs. In graphics mode, these GPUs are supplied with ECC memory disabled by default.

Some hypervisor software versions do not support ECC memory with NVIDIA vGPU.

If you are using a hypervisor software version or GPU that does not support ECC memory with NVIDIA vGPU and ECC memory is enabled, NVIDIA vGPU fails to start. In this situation, you must ensure that ECC memory is disabled on all GPUs if you are using NVIDIA vGPU.

2.13.1. Disabling ECC Memory

If ECC memory is unsuitable for your workloads but is enabled on your GPUs, disable it. You must also ensure that ECC memory is disabled on all GPUs if you are using NVIDIA vGPU with a hypervisor software version or a GPU that does not support ECC memory with NVIDIA vGPU. If your hypervisor software version or GPU does not support ECC memory and ECC memory is enabled, NVIDIA vGPU fails to start.

Where to perform this task depends on whether you are changing ECC memory settings for a physical GPU or a vGPU.

  • For a physical GPU, perform this task from the hypervisor host.
  • For a vGPU, perform this task from the VM to which the vGPU is assigned.
    Note:

    ECC memory must be enabled on the physical GPU on which the vGPUs reside.

Before you begin, ensure that NVIDIA Virtual GPU Manager is installed on your hypervisor. If you are changing ECC memory settings for a vGPU, also ensure that the NVIDIA vGPU software graphics driver is installed in the VM to which the vGPU is assigned.

  1. Use nvidia-smi to list the status of all physical GPUs or vGPUs, and check for ECC noted as enabled.
    Copy
    Copied!
                

    # nvidia-smi -q ==============NVSMI LOG============== Timestamp : Mon Jul 15 18:36:45 2024 Driver Version : 470.256.02 Attached GPUs : 1 GPU 0000:02:00.0 [...] Ecc Mode Current : Enabled Pending : Enabled [...]

  2. Change the ECC status to off for each GPU for which ECC is enabled.
    • If you want to change the ECC status to off for all GPUs on your host machine or vGPUs assigned to the VM, run this command:
      Copy
      Copied!
                  

      # nvidia-smi -e 0

    • If you want to change the ECC status to off for a specific GPU or vGPU, run this command:
      Copy
      Copied!
                  

      # nvidia-smi -i id -e 0

      id is the index of the GPU or vGPU as reported by nvidia-smi.

      This example disables ECC for the GPU with index 0000:02:00.0.

      Copy
      Copied!
                  

      # nvidia-smi -i 0000:02:00.0 -e 0

  3. Reboot the host or restart the VM.
  4. Confirm that ECC is now disabled for the GPU or vGPU.
    Copy
    Copied!
                

    # nvidia—smi —q ==============NVSMI LOG============== Timestamp : Mon Jul 15 18:37:53 2024 Driver Version : 470.256.02 Attached GPUs : 1 GPU 0000:02:00.0 [...] Ecc Mode Current : Disabled Pending : Disabled [...]

If you later need to enable ECC on your GPUs or vGPUs, follow the instructions in Enabling ECC Memory.

2.13.2. Enabling ECC Memory

If ECC memory is suitable for your workloads and is supported by your hypervisor software and GPUs, but is disabled on your GPUs or vGPUs, enable it.

Where to perform this task depends on whether you are changing ECC memory settings for a physical GPU or a vGPU.

  • For a physical GPU, perform this task from the hypervisor host.
  • For a vGPU, perform this task from the VM to which the vGPU is assigned.
    Note:

    ECC memory must be enabled on the physical GPU on which the vGPUs reside.

Before you begin, ensure that NVIDIA Virtual GPU Manager is installed on your hypervisor. If you are changing ECC memory settings for a vGPU, also ensure that the NVIDIA vGPU software graphics driver is installed in the VM to which the vGPU is assigned.

  1. Use nvidia-smi to list the status of all physical GPUs or vGPUs, and check for ECC noted as disabled.
    Copy
    Copied!
                

    # nvidia-smi -q ==============NVSMI LOG============== Timestamp : Mon Jul 15 18:36:45 2024 Driver Version : 470.256.02 Attached GPUs : 1 GPU 0000:02:00.0 [...] Ecc Mode Current : Disabled Pending : Disabled [...]

  2. Change the ECC status to on for each GPU or vGPU for which ECC is enabled.
    • If you want to change the ECC status to on for all GPUs on your host machine or vGPUs assigned to the VM, run this command:
      Copy
      Copied!
                  

      # nvidia-smi -e 1

    • If you want to change the ECC status to on for a specific GPU or vGPU, run this command:
      Copy
      Copied!
                  

      # nvidia-smi -i id -e 1

      id is the index of the GPU or vGPU as reported by nvidia-smi.

      This example enables ECC for the GPU with index 0000:02:00.0.

      Copy
      Copied!
                  

      # nvidia-smi -i 0000:02:00.0 -e 1

  3. Reboot the host or restart the VM.
  4. Confirm that ECC is now enabled for the GPU or vGPU.
    Copy
    Copied!
                

    # nvidia—smi —q ==============NVSMI LOG============== Timestamp : Mon Jul 15 18:37:53 2024 Driver Version : 470.256.02 Attached GPUs : 1 GPU 0000:02:00.0 [...] Ecc Mode Current : Enabled Pending : Enabled [...]

If you later need to disable ECC on your GPUs or vGPUs, follow the instructions in Disabling ECC Memory.

GPU pass-through is used to directly assign an entire physical GPU to one VM, bypassing the NVIDIA Virtual GPU Manager. In this mode of operation, the GPU is accessed exclusively by the NVIDIA driver running in the VM to which it is assigned; the GPU is not shared among VMs.

In pass-through mode, GPUs based on NVIDIA GPU architectures after the Maxwell architecture support error-correcting code (ECC).

GPU pass-through can be used in a server platform alongside NVIDIA vGPU, with some restrictions:

  • A physical GPU can host NVIDIA vGPUs, or can be used for pass-through, but cannot do both at the same time. Some hypervisors, for example VMware vSphere ESXi, require a host reboot to change a GPU from pass-through mode to vGPU mode.
  • A single VM cannot be configured for both vGPU and GPU pass-through at the same time.
  • The performance of a physical GPU passed through to a VM can be monitored only from within the VM itself. Such a GPU cannot be monitored by tools that operate through the hypervisor, such as XenCenter or nvidia-smi (see Monitoring GPU Performance).
  • The following BIOS settings must be enabled on your server platform:

    • VT-D/IOMMU
    • SR-IOV in Advanced Options
  • All GPUs directly connected to each other through NVLink must be assigned to the same VM.

You can assign multiple physical GPUs to one VM. The maximum number of physical GPUs that you can assign to a VM depends on the maximum number of PCIe pass-through devices per VM that your chosen hypervisor can support. For more information, refer to the documentation for your hypervisor, for example:

Note:

If you intend to configure all GPUs in your server platform for pass-through, you do not need to install the NVIDIA Virtual GPU Manager.


3.1. Display Resolutions for Physical GPUs

The display resolutions supported by a physical GPU depend on the NVIDIA GPU architecture and the NVIDIA vGPU software license that is applied to the GPU.

vWS Physical GPU Resolutions

GPUs that are licensed with a vWS license support a maximum combined resolution based on the number of available pixels, which is determined by the NVIDIA GPU architecture. You can choose between using a small number of high resolution displays or a larger number of lower resolution displays with these GPUs.

The following table lists the maximum number of displays per GPU at each supported display resolution for configurations in which all displays have the same resolution.

NVIDIA GPU Architecture Available Pixels Display Resolution Displays per GPU
Pascal and later 66355200 7680×4320 2
5120×2880 or lower 4
Maxwell 35389440 5120×2880 2
4096×2160 or lower 4

The following table provides examples of configurations with a mixture of display resolutions.

NVIDIA GPU Architecture Available Pixels Available Pixel Basis Maximum Displays Sample Mixed Display Configurations
Pascal and later 66355200 2 7680×4320 displays 4 1 7680×4320 display plus 2 5120×2880 displays
1 7680×4320 display plus 3 4096×2160 displays
Maxwell 35389440 4 4096×2160 displays 4 1 5120×2880 display plus 2 4096×2160 displays

Note:

You cannot use more than four displays even if the combined resolution of the displays is less than the number of available pixels from the GPU. For example, you cannot use five 4096×2160 displays with a GPU based on the NVIDIA Pascal architecture even though the combined resolution of the displays (44236800) is less than the number of available pixels from the GPU (66355200).


vApps or vCS Physical GPU Resolutions

GPUs that are licensed with a vApps or a vCS license support a single display with a fixed maximum resolution. The maximum resolution depends on the following factors:

  • NVIDIA GPU architecture
  • The NVIDIA vGPU Software license that is applied to the GPU
  • The operating system that is running in the on the system to which the GPU is assigned

License NVIDIA GPU Architecture Operating System Maximum Display Resolution Displays per GPU
vApps Pascal or later Linux 2560×1600 1
Pascal or later Windows 1280×1024 1
Maxwell Windows and Linux 2560×1600 1
vCS Pascal or later Linux 4096×2160 1

3.2. Using GPU Pass-Through on Citrix Hypervisor

You can configure a GPU for pass-through on Citrix Hypervisor by using XenCenter or by using the xe command.

The following additional restrictions apply when GPU pass-through is used in a server platform alongside NVIDIA vGPU:

  • The performance of a physical GPU passed through to a VM cannot be monitored through XenCenter.
  • nvidia-smi in dom0 no longer has access to the GPU.
  • Pass-through GPUs do not provide console output through XenCenter’s VM Console tab. Use a remote graphics connection directly into the VM to access the VM’s OS.

3.2.1. Configuring a VM for GPU Pass Through by Using XenCenter

Select the Pass-through whole GPU option as the GPU type in the VM’s Properties:

Figure 16. Using XenCenter to configure a pass-through GPU

using-gpu-pass-through-citrix-xenserver-xencenter-screen.png

After configuring a Citrix Hypervisor VM for GPU pass through, install the NVIDIA graphics driver in the guest OS on the VM as explained in Installing the NVIDIA vGPU Software Graphics Driver.

3.2.2. Configuring a VM for GPU Pass Through by Using xe

Create a vgpu object with the passthrough vGPU type:

Copy
Copied!
            

[root@xenserver ~]# xe vgpu-type-list model-name="passthrough" uuid ( RO) : fa50b0f0-9705-6c59-689e-ea62a3d35237 vendor-name ( RO): model-name ( RO): passthrough framebuffer-size ( RO): 0 [root@xenserver ~]# xe vgpu-create vm-uuid=753e77a9-e10d-7679-f674-65c078abb2eb vgpu-type-uuid=fa50b0f0-9705-6c59-689e-ea62a3d35237 gpu-group-uuid=585877ef-5a6c-66af-fc56-7bd525bdc2f6 6aa530ec-8f27-86bd-b8e4-fe4fde8f08f9 [root@xenserver ~]#

CAUTION:

Do not assign pass-through GPUs using the legacy other-config:pci parameter setting. This mechanism is not supported alongside the XenCenter UI and xe vgpu mechanisms, and attempts to use it may lead to undefined results.


After configuring a Citrix Hypervisor VM for GPU pass through, install the NVIDIA graphics driver in the guest OS on the VM as explained in Installing the NVIDIA vGPU Software Graphics Driver.

3.3. Using GPU Pass-Through on Red Hat Enterprise Linux KVM or Ubuntu

You can configure a GPU for pass-through on Red Hat Enterprise Linux Kernel-based Virtual Machine (KVM) or Ubuntu by using any of the following tools:

  • The Virtual Machine Manager (virt-manager) graphical tool
  • The virsh command
  • The QEMU command line

Before configuring a GPU for pass-through on Red Hat Enterprise Linux KVM or Ubuntu, ensure that the following prerequisites are met:

  • Red Hat Enterprise Linux KVM or Ubuntu is installed.
  • A virtual disk has been created.
    Note:

    Do not create any virtual disks in /root.

  • A virtual machine has been created.

3.3.1. Configuring a VM for GPU Pass-Through by Using Virtual Machine Manager (virt-manager)

For more information about using Virtual Machine Manager, see the following topics in the documentation for Red Hat Enterprise Linux 7:

  1. Start virt-manager.
  2. In the virt-manager main window, select the VM that you want to configure for pass-through.
  3. From the Edit menu, choose Virtual Machine Details.
  4. In the virtual machine hardware information window that opens, click Add Hardware.
  5. In the Add New Virtual Hardware dialog box that opens, in the hardware list on the left, select PCI Host Device.
  6. From the Host Device list that appears, select the GPU that you want to assign to the VM and click Finish.

If you want to remove a GPU from the VM to which it is assigned, in the virtual machine hardware information window, select the GPU and click Remove.

After configuring the VM for GPU pass through, install the NVIDIA graphics driver in the guest OS on the VM as explained in Installing the NVIDIA vGPU Software Graphics Driver.

3.3.2. Configuring a VM for GPU Pass-Through by Using virsh

For more information about using virsh, see the following topics in the documentation for Red Hat Enterprise Linux 7:

  1. Verify that the vfio-pci module is loaded.
    Copy
    Copied!
                

    # lsmod | grep vfio-pci

  2. Obtain the PCI device bus/device/function (BDF) of the GPU that you want to assign in pass-through mode to a VM.
    Copy
    Copied!
                

    # lspci | grep NVIDIA

    The NVIDIA GPUs listed in this example have the PCI device BDFs 85:00.0 and 86:00.0.

    Copy
    Copied!
                

    # lspci | grep NVIDIA 85:00.0 VGA compatible controller: NVIDIA Corporation GM204GL [Tesla M60] (rev a1) 86:00.0 VGA compatible controller: NVIDIA Corporation GM204GL [Tesla M60] (rev a1)

  3. Obtain the full identifier of the GPU from its PCI device BDF.
    Copy
    Copied!
                

    # virsh nodedev-list --cap pci| grep transformed-bdf

    transformed-bdf
    The PCI device BDF of the GPU with the colon and the period replaced with underscores, for example, 85_00_0.

    This example obtains the full identifier of the GPU with the PCI device BDF 85:00.0.

    Copy
    Copied!
                

    # virsh nodedev-list --cap pci| grep 85_00_0 pci_0000_85_00_0

  4. Obtain the domain, bus, slot, and function of the GPU.
    Copy
    Copied!
                

    virsh nodedev-dumpxml full-identifier| egrep 'domain|bus|slot|function'

    full-identifier
    The full identifier of the GPU that you obtained in the previous step, for example, pci_0000_85_00_0.

    This example obtains the domain, bus, slot, and function of the GPU with the PCI device BDF 85:00.0.

    Copy
    Copied!
                

    # virsh nodedev-dumpxml pci_0000_85_00_0| egrep 'domain|bus|slot|function' <domain>0x0000</domain> <bus>0x85</bus> <slot>0x00</slot> <function>0x0</function> <address domain='0x0000' bus='0x85' slot='0x00' function='0x0'/>

  5. In virsh, open for editing the XML file of the VM that you want to assign the GPU to.
    Copy
    Copied!
                

    # virsh edit vm-name

    vm-name
    The name of the VM to that you want to assign the GPU to.
  6. Add a device entry in the form of an address element inside the source element to assign the GPU to the guest VM. You can optionally add a second address element after the source element to set a fixed PCI device BDF for the GPU in the guest operating system.
    Copy
    Copied!
                

    <hostdev mode='subsystem' type='pci' managed='yes'> <source> <address domain='domain' bus='bus' slot='slot' function='function'/> </source> <address type='pci' domain='0x0000' bus='0x00' slot='0x05' function='0x0'/> </hostdev>

    domain
    bus
    slot
    function
    The domain, bus, slot, and function of the GPU, which you obtained in the previous step.

    This example adds a device entry for the GPU with the PCI device BDF 85:00.0 and fixes the BDF for the GPU in the guest operating system.

    Copy
    Copied!
                

    <hostdev mode='subsystem' type='pci' managed='yes'> <source> <address domain='0x0000' bus='0x85' slot='0x00' function='0x0'/> </source> <address type='pci' domain='0x0000' bus='0x00' slot='0x05' function='0x0'/> </hostdev>

  7. Start the VM that you assigned the GPU to.
    Copy
    Copied!
                

    # virsh start vm-name

    vm-name
    The name of the VM that you assigned the GPU to.

After configuring the VM for GPU pass through, install the NVIDIA graphics driver in the guest OS on the VM as explained in Installing the NVIDIA vGPU Software Graphics Driver.

3.3.3. Configuring a VM for GPU Pass-Through by Using the QEMU Command Line

  1. Obtain the PCI device bus/device/function (BDF) of the GPU that you want to assign in pass-through mode to a VM.
    Copy
    Copied!
                

    # lspci | grep NVIDIA

    The NVIDIA GPUs listed in this example have the PCI device BDFs 85:00.0 and 86:00.0.

    Copy
    Copied!
                

    # lspci | grep NVIDIA 85:00.0 VGA compatible controller: NVIDIA Corporation GM204GL [Tesla M60] (rev a1) 86:00.0 VGA compatible controller: NVIDIA Corporation GM204GL [Tesla M60] (rev a1)

  2. Add the following option to the QEMU command line:
    Copy
    Copied!
                

    -device vfio-pci,host=bdf

    bdf
    The PCI device BDF of the GPU that you want to assign in pass-through mode to a VM, for example, 85:00.0.

    This example assigns the GPU with the PCI device BDF 85:00.0 in pass-through mode to a VM.

    Copy
    Copied!
                

    -device vfio-pci,host=85:00.0

After configuring the VM for GPU pass through, install the NVIDIA graphics driver in the guest OS on the VM as explained in Installing the NVIDIA vGPU Software Graphics Driver.

3.3.4. Preparing a GPU Configured for vGPU for Use in Pass-Through Mode

The mode in which a physical GPU is being used determines the Linux kernel module to which the GPU is bound. If you want to switch the mode in which a GPU is being used, you must unbind the GPU from its current kernel module and bind it to the kernel module for the new mode. After binding the GPU to the correct kernel module, you can then configure it for pass-through.

When the Virtual GPU Manager is installed on a Red Hat Enterprise Linux KVM or Ubuntu host, the physical GPUs on the host are bound to the nvidia kernel module. A physical GPU that is bound to the nvidia kernel module can be used only for vGPU. To enable the GPU to be passed through to a VM, the GPU must be unbound from nvidia kernel module and bound to the vfio-pci kernel module.
Before you begin, ensure that you have the domain, bus, slot, and function of the GPU that you are preparing for use in pass-through mode. For instructions, see Getting the BDF and Domain of a GPU on a Linux with KVM Hypervisor.

  1. If you are using a GPU that supports SR-IOV, such as a GPU based on the NVIDIA Ampere architecture, disable the virtual function for the GPU in the sysfs file system.

    If your GPU does not support SR-IOV, omit this step.

    Note:

    Before performing this step, ensure that the GPU is not being used by any other processes, such as CUDA applications, monitoring applications, or the nvidia-smi command.

    Use the custom script sriov-manage provided by NVIDIA vGPU software for this purpose.

    Copy
    Copied!
                

    # /usr/lib/nvidia/sriov-manage -d domain:bus:slot.function

    domain
    bus
    slot
    function
    The domain, bus, slot, and function of the GPU, without the 0x prefix.

    This example disables the virtual function for the GPU with the domain 00, bus 06, slot 0000, and function 0.

    Copy
    Copied!
                

    # /usr/lib/nvidia/sriov-manage -d 00:06:0000.0

  2. Determine the kernel module to which the GPU is bound by running the lspci command with the -k option on the NVIDIA GPUs on your host.
    Copy
    Copied!
                

    # lspci -d 10de: -k

    The Kernel driver in use: field indicates the kernel module to which the GPU is bound.

    The following example shows that the NVIDIA Tesla M60 GPU with BDF06:00.0 is bound to the nvidia kernel module and is being used for vGPU.

    Copy
    Copied!
                

    06:00.0 VGA compatible controller: NVIDIA Corporation GM204GL [Tesla M60] (rev a1) Subsystem: NVIDIA Corporation Device 115e Kernel driver in use: nvidia

  3. To ensure that no clients are using the GPU, acquire the unbind lock of the GPU.
    1. Ensure that no VM is running to which a vGPU on the physical GPU is assigned and that no process running on the host is using that GPU. Processes on the host that use the GPU include the nvidia-smi command and all processes based on the NVIDIA Management Library (NVML).
    2. Change to the directory in the proc file system that represents the GPU.
      Copy
      Copied!
                  

      # cd /proc/driver/nvidia/gpus/domain\:bus\:slot.function

      domain
      bus
      slot
      function
      The domain, bus, slot, and function of the GPU, without a 0x prefix.

      This example changes to the directory in the proc file system that represents the GPU with the domain 0000 and PCI device BDF 06:00.0.

      Copy
      Copied!
                  

      # cd /proc/driver/nvidia/gpus/0000\:06\:00.0

    3. Write the value 1 to the unbindLock file in this directory.
      Copy
      Copied!
                  

      # echo 1 > unbindLock

    4. Confirm that the unbindLock file now contains the value 1.
      Copy
      Copied!
                  

      # cat unbindLock 1

      If the unbindLock file contains the value 0, the unbind lock could not be acquired because a process or client is using the GPU.

  4. Unbind the GPU from nvidia kernel module.
    1. Change to the sysfs directory that represents the nvidia kernel module.
      Copy
      Copied!
                  

      # cd /sys/bus/pci/drivers/nvidia

    2. Write the domain, bus, slot, and function of the GPU to the unbind file in this directory.
      Copy
      Copied!
                  

      # echo domain:bus:slot.function > unbind

      domain
      bus
      slot
      function
      The domain, bus, slot, and function of the GPU, without a 0x prefix.

      This example writes the domain, bus, slot, and function of the GPU with the domain 0000 and PCI device BDF 06:00.0.

      Copy
      Copied!
                  

      # echo 0000:06:00.0 > unbind

  5. Bind the GPU to the vfio-pci kernel module.
    1. Change to the sysfs directory that contains the PCI device information for the physical GPU.
      Copy
      Copied!
                  

      # cd /sys/bus/pci/devices/domain\:bus\:slot.function

      domain
      bus
      slot
      function
      The domain, bus, slot, and function of the GPU, without a 0x prefix.

      This example changes to the sysfs directory that contains the PCI device information for the GPU with the domain 0000 and PCI device BDF 06:00.0.

      Copy
      Copied!
                  

      # cd /sys/bus/pci/devices/0000\:06\:00.0

    2. Write the kernel module name vfio-pci to the driver_override file in this directory.
      Copy
      Copied!
                  

      # echo vfio-pci > driver_override

    3. Change to the sysfs directory that represents the nvidia kernel module.
      Copy
      Copied!
                  

      # cd /sys/bus/pci/drivers/vfio-pci

    4. Write the domain, bus, slot, and function of the GPU to the bind file in this directory.
      Copy
      Copied!
                  

      # echo domain:bus:slot.function > bind

      domain
      bus
      slot
      function
      The domain, bus, slot, and function of the GPU, without a 0x prefix.

      This example writes the domain, bus, slot, and function of the GPU with the domain 0000 and PCI device BDF 06:00.0.

      Copy
      Copied!
                  

      # echo 0000:06:00.0 > bind

    5. Change back to the sysfs directory that contains the PCI device information for the physical GPU.
      Copy
      Copied!
                  

      # cd /sys/bus/pci/devices/domain\:bus\:slot.function

    6. Clear the content of the driver_override file in this directory.
      Copy
      Copied!
                  

      # echo > driver_override

You can now configure the GPU for use in pass-through mode as explained in Using GPU Pass-Through on Red Hat Enterprise Linux KVM or Ubuntu.

3.4. Using GPU Pass-Through on Microsoft Windows Server

On supported versons of Microsoft Windows Server with Hyper-V role, you can use Discrete Device Assignment (DDA) to enable a VM to access a GPU directly.

3.4.1. Assigning a GPU to a VM on Microsoft Windows Server with Hyper-V

Perform this task in Windows PowerShell. If you do not know the location path of the GPU that you want to assign to a VM, use Device Manager to obtain it.

If you are using an actively cooled NVIDIA Quadro graphics card such as the RTX 8000 or RTX 6000, you must also pass through the audio device on the graphics card.

Ensure that the following prerequisites are met:

  1. Obtain the location path of the GPU that you want to assign to a VM.
    1. In the device manager, context-click the GPU and from the menu that pops up, choose Properties.
    2. In the Properties window that opens, click the Details tab and in the Properties drop-down list, select Location paths.

    An example location path is as follows:

    Copy
    Copied!
                

    PCIROOT(80)#PCI(0200)#PCI(0000)#PCI(1000)#PCI(0000)

  2. If you are using an actively cooled NVIDIA Quadro graphics card, obtain the location path of the audio device on the graphics card and disable the device.
    1. In the device manager, from the View menu, choose Devices by connection.
    2. Navigate to ACPI x64-based PC > Microsoft ACPI-Compliant System > PCI Express Root Complex > PCI-to-PCI Bridge.
    3. Context-click High Definition Audio Controller and from the menu that pops up, choose Properties.
    4. In the Properties window that opens, click the Details tab and in the Properties drop-down list, select Location paths.
    5. Context-click High Definition Audio Controller again and from the menu that pops up, choose Disable device.
  3. Dismount the GPU and, if present, the audio device from host to make them unavailable to the host so that they can be used solely by the VM.

    For each device that you are dismounting, type the following command:

    Copy
    Copied!
                

    Dismount-VMHostAssignableDevice -LocationPath gpu-device-location -force

    gpu-device-location
    The location path of the GPU or the audio device that you obtained previously.

    This example dismounts the GPU at the location path PCIROOT(80)#PCI(0200)#PCI(0000)#PCI(1000)#PCI(0000).

    Copy
    Copied!
                

    Dismount-VMHostAssignableDevice -LocationPath "PCIROOT(80)#PCI(0200)#PCI(0000)#PCI(1000)#PCI(0000)" -force

  4. Assign the GPU and, if present, the audio device that you dismounted in the previous step to the VM.

    For each device that you are assigning, type the following command:

    Copy
    Copied!
                

    Add-VMAssignableDevice -LocationPath gpu-device-location -VMName vm-name

    gpu-device-location
    The location path of the GPU or the audio device that you dismounted in the previous step.
    vm-name
    The name of the VM to which you are attaching the GPU or the audio device.
    Note:

    You can assign a pass-through GPU and, if present, its audio device to only one virtual machine at a time.

    This example assigns the GPU at the location path PCIROOT(80)#PCI(0200)#PCI(0000)#PCI(1000)#PCI(0000) to the VM VM1.

    Copy
    Copied!
                

    Add-VMAssignableDevice -LocationPath "PCIROOT(80)#PCI(0200)#PCI(0000)#PCI(1000)#PCI(0000)" -VMName VM1

  5. Power on the VM. The guest OS should now be able to use the GPU and, if present, the audio device.

After assigning a GPU to a VM, install the NVIDIA graphics driver in the guest OS on the VM as explained in Installing the NVIDIA vGPU Software Graphics Driver.

3.4.2. Returning a GPU to the Host OS from a VM on Windows Server with Hyper-V

Perform this task in the Windows PowerShell.

If you are using an actively cooled NVIDIA Quadro graphics card such as the RTX 8000 or RTX 6000, you must also return the audio device on the graphics card.

  1. List the GPUs and, if present, the audio devices that are currently assigned to the virtual machine (VM).
    Copy
    Copied!
                

    Get-VMAssignableDevice -VMName vm-name

    vm-name
    The name of the VM whose assigned GPUs and audio devices you want to list.
  2. Shut down the VM to which the GPU and any audio devices are assigned.
  3. Remove the GPU and, if present, the audio device from the VM to which they are assigned.

    For each device that you are removing, type the following command:

    Copy
    Copied!
                

    Remove-VMAssignableDevice –LocationPath gpu-device-location -VMName vm-name

    gpu-device-location
    The location path of the GPU or the audio device that you are removing, which you obtained previously.
    vm-name
    The name of the VM from which you are removing the GPU or the audio device.

    This example removes the GPU at the location path PCIROOT(80)#PCI(0200)#PCI(0000)#PCI(1000)#PCI(0000) from the VM VM1.

    Copy
    Copied!
                

    Remove-VMAssignableDevice –LocationPath "PCIROOT(80)#PCI(0200)#PCI(0000)#PCI(1000)#PCI(0000)" -VMName VM1

    After the GPU and, if present, its audio device are removed from the VM, they are unavailable to the host operating system (OS) until you remount them on the host OS.

  4. Remount the GPU and, if present, its audio device on the host OS.

    For each device that you are remounting, type the following command:

    Copy
    Copied!
                

    Mount-VMHostAssignableDevice –LocationPath gpu-device-location

    gpu-device-location
    The location path of the GPU or the audio device that you are remounting, which you specified in the previous step to remove the GPU or the audio device from the VM.

    This example remounts the GPU at the location path PCIROOT(80)#PCI(0200)#PCI(0000)#PCI(1000)#PCI(0000) on the host OS.

    Copy
    Copied!
                

    Mount-VMHostAssignableDevice -LocationPath "PCIROOT(80)#PCI(0200)#PCI(0000)#PCI(1000)#PCI(0000)"

    The host OS should now be able to use the GPU and, if present, its audio device.

3.5. Using GPU Pass-Through on VMware vSphere

On VMware vSphere, you can use Virtual Dedicated Graphics Acceleration (vDGA) to enable a VM to access a GPU directly. vDGA is a feature of VMware vSphere that dedicates a single physical GPU on an ESXi host to a single virtual machine.
Before configuring a vSphere VM with vDGA, ensure that these prerequisites are met

  1. Open the vCenter Web UI.
  2. In the vCenter Web UI, right-click the ESXi host and choose Configure.
  3. From the Hardware menu, choose PCI Devices.
  4. On the PCI Devices page that opens, click ALL PCI DEVICES and in the table of devices, select the GPU.
    Note:

    When selecting the GPU to pass through, you must select only the physical device. To list only NVIDIA physical devices, set the filter on the Vendor Name field to NVIDIA and filter out any virtual function devices of the GPU by setting the filter on the ID field to 00.0.

    vmware-vcenter-filter-physical-pci-device.png

  5. Click TOGGLE PASSTHROUGH.
  6. Reboot the ESXi host.
  7. After the ESXi host has booted, right-click the VM and choose Edit Settings.
  8. From the New Device menu, choose PCI Device and click Add.
  9. On the page that opens, from the New Device drop-down list, select the GPU.
  10. Click Reserve all memory and click OK.
  11. Start the VM.

For more information about vDGA, see the following topics in the VMware Horizon documentation:

After configuring a vSphere VM with vDGA, install the NVIDIA graphics driver in the guest OS on the VM as explained in Installing the NVIDIA vGPU Software Graphics Driver.

The process for installing the NVIDIA vGPU software graphics driver depends on the OS that you are using. However, for any OS, the process for installing the driver is the same in a VM configured with vGPU, in a VM that is running pass-through GPU, or on a physical host in a bare-metal deployment.

After you install the NVIDIA vGPU software graphics driver, you can license any NVIDIA vGPU software licensed products that you are using.

4.1. Installing the NVIDIA vGPU Software Graphics Driver on Windows

Installation in a VM: After you create a Windows VM on the hypervisor and boot the VM, the VM should boot to a standard Windows desktop in VGA mode at 800×600 resolution. You can use the Windows screen resolution control panel to increase the resolution to other standard resolutions, but to fully enable GPU operation, the NVIDIA vGPU software graphics driver must be installed. Windows guest VMs are supported only on Q-series, B-series, and A-series NVIDIA vGPU types. They are not supported on C-series NVIDIA vGPU types.

Installation on bare metal: When the physical host is booted before the NVIDIA vGPU software graphics driver is installed, boot and the primary display are handled by an on-board graphics adapter. To install the NVIDIA vGPU software graphics driver, access the Windows desktop on the host by using a display connected through the on-board graphics adapter.

The procedure for installing the driver is the same in a VM and on bare metal.

  1. Copy the NVIDIA Windows driver package to the guest VM or physical host where you are installing the driver.
  2. Execute the package to unpack and run the driver installer.

    Figure 17. NVIDIA driver installation

    install-nvidia-windows-guest-driver.png

  3. Click through the license agreement.
  4. Select Express Installation and click NEXT. After the driver installation is complete, the installer may prompt you to restart the platform.
  5. If prompted to restart the platform, do one of the following:
    • Select Restart Now to reboot the VM or physical host.
    • Exit the installer and reboot the VM or physical host when you are ready.

    After the VM or physical host restarts, it boots to a Windows desktop.

  6. Verify that the NVIDIA driver is running.
    1. Right-click on the desktop.
    2. From the menu that opens, choose NVIDIA Control Panel.
    3. In the NVIDIA Control Panel, from the Help menu, choose System Information.

      NVIDIA Control Panel reports the vGPU or physical GPU that is being used, its capabilities, and the NVIDIA driver version that is loaded.

      Figure 18. Verifying NVIDIA driver operation using NVIDIA Control Panel

      verify-nvidia-driver-nvcp.png

Installation in a VM: After you install the NVIDIA vGPU software graphics driver, you can license any NVIDIA vGPU software licensed products that you are using. For instructions, refer to Virtual GPU Client Licensing User Guide.

Installation on bare metal: After you install the NVIDIA vGPU software graphics driver, complete the bare-metal deployment as explained in Bare-Metal Deployment.

4.2. Installing the NVIDIA vGPU Software Graphics Driver on Linux

Installation in a VM: After you create a Linux VM on the hypervisor and boot the VM, install the NVIDIA vGPU software graphics driver in the VM to fully enable GPU operation. 64-bit Linux guest VMs are supported only on Q-series, C-series, and B-series NVIDIA vGPU types. They are not supported on A-series NVIDIA vGPU types.

Installation on bare metal: When the physical host is booted before the NVIDIA vGPU software graphics driver is installed, the vesa Xorg driver starts the X server. If a primary display device is connected to the host, use the device to access the desktop. Otherwise, use secure shell (SSH) to log in to the host from a remote host.

The procedure for installing the driver is the same in a VM and on bare metal.

Installation of the NVIDIA vGPU software graphics driver for Linux requires:

  • Compiler toolchain
  • Kernel headers

Before installing the NVIDIA vGPU software graphics driver, ensure that the following prerequisites are met:

  • NVIDIA Direct Rendering Manager Kernel Modesetting (DRM KMS) is disabled. By default, DRM KMS is disabled. However, if it has been enabled, remove nvidia-drm.modeset=1 from the kernel command-line options.
  • If the VM uses UEFI boot, ensure that secure boot is disabled.
  • If the Nouveau driver for NVIDIA graphics cards is present, disable it. For instructions, refer to as explained in Disabling the Nouveau Driver for NVIDIA Graphics Cards.
  • If you are using a Linux OS for which the Wayland display server protocol is enabled by default, disable it as explained in Disabling the Wayland Display Server Protocol for Red Hat Enterprise Linux.

  1. Copy the NVIDIA vGPU software Linux driver package, for example NVIDIA-Linux_x86_64-470.256.02-grid.run, to the guest VM or physical host where you are installing the driver.
  2. Before attempting to run the driver installer, exit the X server and terminate all OpenGL applications.
    • On Red Hat Enterprise Linux and CentOS systems, exit the X server by transitioning to runlevel 3:
      Copy
      Copied!
                  

      [nvidia@localhost ~]$ sudo init 3

    • On Ubuntu platforms, do the following:
      1. Switch to a console login prompt.
        • If you have access to the terminal's function keys, press CTRL-ALT-F1.
        • If you are accessing the guest VM or physical host through VNC or a web browser and do not have access to the terminal's function keys, run the chvt command of the OS as root.
          Copy
          Copied!
                      

          [nvidia@localhost ~]$ sudo chvt 3

      2. Log in and shut down the display manager:
        • For Ubuntu 18 and later releases, stop the gdm service

          Copy
          Copied!
                      

          [nvidia@localhost ~]$ sudo service gdm stop

        • For releases earlier than Ubuntu 18, stop the lightdm service.

          Copy
          Copied!
                      

          [nvidia@localhost ~]$ sudo service lightdm stop

  3. From a console shell, run the driver installer as the root user.
    Copy
    Copied!
                

    sudo sh ./NVIDIA-Linux_x86_64-470.256.02-grid.run

    If DKMS is enabled, set the -dkms option. This option requires the dkms package to be installed.
    Copy
    Copied!
                

    sudo sh ./NVIDIA-Linux_x86_64-470.256.02-grid.run -dkms

    In some instances the installer may fail to detect the installed kernel headers and sources. In this situation, re-run the installer, specifying the kernel source path with the --kernel-source-path option.
    Copy
    Copied!
                

    sudo sh ./NVIDIA-Linux_x86_64-470.256.02-grid.run \ –kernel-source-path=/usr/src/kernels/3.10.0-229.11.1.el7.x86_64

  4. When prompted, accept the option to update the X configuration file (xorg.conf).

    Figure 19. Update xorg.conf settings

    update-xorg-conf-linux-driver-installer.png

  5. Once installation has completed, select OK to exit the installer.
  6. Verify that the NVIDIA driver is operational.
    1. Reboot the system and log in.
    2. Run nvidia-settings.
      Copy
      Copied!
                  

      [nvidia@localhost ~]$ nvidia-settings

      The NVIDIA X Server Settings dialog box opens to show that the NVIDIA driver is operational.

Installation in a VM: After you install the NVIDIA vGPU software graphics driver, you can license any NVIDIA vGPU software licensed products that you are using. For instructions, refer to Virtual GPU Client Licensing User Guide.

Installation on bare metal: After you install the NVIDIA vGPU software graphics driver, complete the bare-metal deployment as explained in Bare-Metal Deployment.

4.3. Disabling the Nouveau Driver for NVIDIA Graphics Cards

If the Nouveau driver for NVIDIA graphics cards is present, disable it before installing the NVIDIA vGPU software graphics driver.

Note:

If you are using SUSE Linux Enterprise Server, you can skip this task because the Nouveau driver is not present in SUSE Linux Enterprise Server.


Run the following command and if the command prints any output, the Nouveau driver is present and must be disabled.

Copy
Copied!
            

$ lsmod | grep nouveau


  1. Create the file /etc/modprobe.d/blacklist-nouveau.conf with the following contents:
    Copy
    Copied!
                

    blacklist nouveau options nouveau modeset=0

  2. Regenerate the kernel initial RAM file system (initramfs). The command to run to regenerate the kernel initramfs depends on the Linux distribution that you are using.
    Linux Distribution Command
    CentOS
    Copy
    Copied!
                

    $ sudo dracut --force

    Debian
    Copy
    Copied!
                

    $ sudo update-initramfs -u

    Red Hat Enterprise Linux
    Copy
    Copied!
                

    $ sudo dracut --force

    Ubuntu
    Copy
    Copied!
                

    $ sudo update-initramfs -u

  3. Reboot the host or guest VM.

4.4. Disabling the Wayland Display Server Protocol for Red Hat Enterprise Linux

Starting with Red Hat Enterprise Linux Desktop 8.0, the Wayland display server protocol is used by default on supported GPU and graphics driver configurations. However, the NVIDIA vGPU software graphics driver for Linux requires the X Window System. Before installing the driver, you must disable the Wayland display server protocol to revert to the X Window System.

Perform this task from the host or guest VM that is running Red Hat Enterprise Linux Desktop.
This task requires administrative access.

  1. In a plain text editor, edit the file /etc/gdm/custom.conf and remove the comment from the option WaylandEnable=false.
  2. Save your changes to /etc/gdm/custom.conf.
  3. Reboot the host or guest VM.

NVIDIA vGPU is a licensed product. When booted on a supported GPU, a vGPU initially operates at full capability but its performance is degraded over time if the VM fails to obtain a license. If the performance of a vGPU has been degraded, the full capability of the vGPU is restored when a license is acquired. For information about how the performance of an unlicensed vGPU is degraded, see Virtual GPU Client Licensing User Guide.

After you license NVIDIA vGPU, the VM that is set up to use NVIDIA vGPU is capable of running the full range of DirectX and OpenGL graphics applications.

If licensing is configured, the virtual machine (VM) obtains a license from the license server when a vGPU is booted on these GPUs. The VM retains the license until it is shut down. It then releases the license back to the license server. Licensing settings persist across reboots and need only be modified if the license server address changes, or the VM is switched to running GPU pass through.

How to license an NVIDIA vGPU depends on whether your licenses are served from NVIDIA License System or the legacy NVIDIA vGPU software license server.

Note:

For complete information about configuring and using NVIDIA vGPU software licensed features, including vGPU, refer to Virtual GPU Client Licensing User Guide.


5.1. Prerequisites for Configuring a Licensed Client of NVIDIA License System

A client with a network connection obtains a license by leasing it from a NVIDIA License System service instance. The service instance serves the license to the client over the network from a pool of floating licenses obtained from the NVIDIA Licensing Portal. The license is returned to the service instance when the licensed client no longer requires the license.

Before configuring a licensed client, ensure that the following prerequisites are met:

  • The NVIDIA vGPU software graphics driver is installed on the client.
  • The client configuration token that you want to deploy on the client has been created from the NVIDIA Licensing Portal or the DLS as explained in NVIDIA License System User Guide.
  • Ports 443 and 80 in your firewall or proxy must be open to allow HTTPS traffic between a service instance and its the licensed clients. These ports must be open for both CLS instances and DLS instances.
    Note:

    For DLS releases before DLS 1.1, ports 8081 and 8082 were also required to be open to allow HTTPS traffic between a DLS instance and its licensed clients. Although these ports are no longer required, they remain supported for backward compatibility.

The graphics driver creates a default location in which to store the client configuration token on the client.

The process for configuring a licensed client is the same for CLS and DLS instances but depends on the OS that is running on the client.

5.1.1. Configuring a Licensed Client on Windows with Default Settings

Perform this task from the client.

  1. Copy the client configuration token to the %SystemDrive%:\Program Files\NVIDIA Corporation\vGPU Licensing\ClientConfigToken folder.
  2. Restart the NvDisplayContainer service.

The NVIDIA service on the client should now automatically obtain a license from the CLS or DLS instance.

5.1.2. Configuring a Licensed Client on Linux with Default Settings

Perform this task from the client.

  1. As root, open the file /etc/nvidia/gridd.conf in a plain-text editor, such as vi.
    Copy
    Copied!
                

    $ sudo vi /etc/nvidia/gridd.conf

    Note:

    You can create the /etc/nvidia/gridd.conf file by copying the supplied template file /etc/nvidia/gridd.conf.template.

  2. Add the FeatureType configuration parameter to the file /etc/nvidia/gridd.conf on a new line as FeatureType="value".

    value depends on the type of the GPU assigned to the licensed client that you are configuring.

    GPU Type Value
    NVIDIA vGPU 1. NVIDIA vGPU software automatically selects the correct type of license based on the vGPU type.
    Physical GPU The feature type of a GPU in pass-through mode or a bare-metal deployment:
    • 0: NVIDIA Virtual Applications
    • 2: NVIDIA RTX Virtual Workstation
    • 4: NVIDIA Virtual Compute Server

    Note:

    You can also perform this step from NVIDIA X Server Settings. Before using NVIDIA X Server Settings to perform this step, ensure that this option has been enabled as explained in Virtual GPU Client Licensing User Guide.

    This example shows how to configure a licensed Linux client for NVIDIA RTX Virtual Workstation.

    Copy
    Copied!
                

    # /etc/nvidia/gridd.conf.template - Configuration file for NVIDIA Grid Daemon … # Description: Set Feature to be enabled # Data type: integer # Possible values: # 0 => for unlicensed state # 1 => for NVIDIA vGPU # 2 => for NVIDIA RTX Virtual Workstation # 4 => for NVIDIA Virtual Compute Server FeatureType=2 ...

  3. Copy the client configuration token to the /etc/nvidia/ClientConfigToken directory.
  4. Ensure that the file access modes of the client configuration token allow the owner to read, write, and execute the token, and the group and others only to read the token.
    1. Determine the current file access modes of the client configuration token.
      Copy
      Copied!
                  

      # ls -l client-configuration-token-directory

    2. If necessary, change the mode of the client configuration token to 744.
      Copy
      Copied!
                  

      # chmod 744 client-configuration-token-directory/client_configuration_token_*.tok

    client-configuration-token-directory
    The directory to which you copied the client configuration token in the previous step.
  5. Save your changes to the /etc/nvidia/gridd.conf file and close the file.
  6. Restart the nvidia-gridd service.

The NVIDIA service on the client should now automatically obtain a license from the CLS or DLS instance.

5.1.3. Verifying the NVIDIA vGPU Software License Status of a Licensed Client

After configuring a client with an NVIDIA vGPU software license, verify the license status by displaying the licensed product name and status.

To verify the license status of a licensed client, run nvidia-smi with the –q or --query option from the licensed client, not the hypervisor host. If the product is licensed, the expiration date is shown in the license status.

Copy
Copied!
            

nvidia-smi -q ==============NVSMI LOG============== Timestamp : Wed Mar 31 01:49:28 2020 Driver Version : 440.88 CUDA Version : 10.0 Attached GPUs : 1 GPU 00000000:00:08.0 Product Name : Tesla T4 Product Brand : Grid Display Mode : Enabled Display Active : Disabled Persistence Mode : N/A Accounting Mode : Disabled Accounting Mode Buffer Size : 4000 Driver Model Current : WDDM Pending : WDDM Serial Number : 0334018000638 GPU UUID : GPU-ba2310b6-95d1-802b-f96f-5865410fe517 Minor Number : N/A VBIOS Version : 90.04.21.00.01 MultiGPU Board : No Board ID : 0x8 GPU Part Number : 699-2G183-0200-100 Inforom Version Image Version : G183.0200.00.02 OEM Object : 1.1 ECC Object : 5.0 Power Management Object : N/A GPU Operation Mode Current : N/A Pending : N/A GPU Virtualization Mode Virtualization mode : Pass-Through vGPU Software Licensed Product Product Name : NVIDIA Virtual Compute Server License Status : Licensed (Expiry: 2021-11-13 18:29:59 GMT) … …

5.2. Licensing NVIDIA vGPU from the Legacy License Server

How to license NVIDIA vGPU depends on the guest OS that is running in the VM.

5.2.1. Licensing an NVIDIA vGPU on Windows

Perform this task from the guest VM to which the vGPU is assigned.

The NVIDIA Control Panel tool that you use to perform this task detects that a vGPU is assigned to the VM and, therefore, provides no options for selecting the license type. After you license the vGPU, NVIDIA vGPU software automatically selects the correct type of license based on the vGPU type.

  1. Open NVIDIA Control Panel:
    • Right-click on the Windows desktop and select NVIDIA Control Panel from the menu.
    • Open Windows Control Panel and double-click the NVIDIA Control Panel icon.
  2. In NVIDIA Control Panel, select the Manage License task in the Licensing section of the navigation pane.
    Note:

    If the Licensing section and Manage License task are not displayed in NVIDIA Control Panel, the system has been configured to hide licensing controls in NVIDIA Control Panel. For information about registry settings, refer to Virtual GPU Client Licensing User Guide.

    The Manage License task pane shows that NVIDIA vGPU is currently unlicensed.

    Figure 20. Managing vGPU licensing in NVIDIA Control Panel

    grid-vgpu-license-windows-nvcp.png

  3. In the Primary License Server field, enter the address of your primary NVIDIA vGPU software License Server. The address can be a fully-qualified domain name such as gridlicense1.example.com, or an IP address such as 10.31.20.45. If you have only one license server configured, enter its address in this field.
  4. Leave the Port Number field under the Primary License Server field unset. The port defaults to 7070, which is the default port number used by NVIDIA vGPU software License Server.
  5. In the Secondary License Server field, enter the address of your secondary NVIDIA vGPU software License Server. If you have only one license server configured, leave this field unset. The address can be a fully-qualified domain name such as gridlicense2.example.com, or an IP address such as 10.31.20.46.
  6. Leave the Port Number field under the Secondary License Server field unset. The port defaults to 7070, which is the default port number used by NVIDIA vGPU software License Server.
  7. Click Apply to assign the settings. The system requests the appropriate license for the current vGPU from the configured license server.

The vGPU within the VM should now operate at full capability without any performance degradation over time for as long as the vGPU is licensed.
If the system fails to obtain a license, see Virtual GPU Client Licensing User Guide for guidance on troubleshooting.

5.2.2. Licensing an NVIDIA vGPU on Linux

Perform this task from the guest VM to which the vGPU is assigned.

The NVIDIA X Server Settings tool that you use to perform this task detects that a vGPU is assigned to the VM and, therefore, provides no options for selecting the license type. After you license the vGPU, NVIDIA vGPU software automatically selects the correct type of license based on the vGPU type.

Ensure that the Manage License option is enabled as explained in Virtual GPU Client Licensing User Guide.

Note:

Do not enable the Manage License option with Red Hat Enterprise Linux 6.8 and 6.9 or CentOS 6.8 and 6.9. To prevent a segmentation fault in DBus code from causing the nvidia-gridd service from exiting, the GUI for licensing must be disabled with these OS versions.


  1. Start NVIDIA X Server Settings by using the method for launching applications provided by your Linux distribution. For example, on Ubuntu Desktop, open the Dash, search for NVIDIA X Server Settings, and click the NVIDIA X Server Settings icon.
  2. In the NVIDIA X Server Settings window that opens, click Manage GRID License. The License Edition section of the NVIDIA X Server Settings window shows that NVIDIA vGPU is currently unlicensed.
  3. In the Primary Server field, enter the address of your primary NVIDIA vGPU software License Server. The address can be a fully-qualified domain name such as gridlicense1.example.com, or an IP address such as 10.31.20.45. If you have only one license server configured, enter its address in this field.
  4. Leave the Port Number field under the Primary Server field unset. The port defaults to 7070, which is the default port number used by NVIDIA vGPU software License Server.
  5. In the Secondary Server field, enter the address of your secondary NVIDIA vGPU software License Server. If you have only one license server configured, leave this field unset. The address can be a fully-qualified domain name such as gridlicense2.example.com, or an IP address such as 10.31.20.46.
  6. Leave the Port Number field under the Secondary Server field unset. The port defaults to 7070, which is the default port number used by NVIDIA vGPU software License Server.
  7. Click Apply to assign the settings. The system requests the appropriate license for the current vGPU from the configured license server.

The vGPU within the VM should now operate at full capability without any performance degradation over time for as long as the vGPU is licensed.
If the system fails to obtain a license, refer to Virtual GPU Client Licensing User Guide for guidance on troubleshooting.

You can modify a VM's NVIDIA vGPU configuration by removing the NVIDIA vGPU configuration from a VM or by modifying GPU allocation policy.

6.1. Removing a VM’s NVIDIA vGPU Configuration

Remove a VM’s NVIDIA vGPU configuration when you no longer require the VM to use a virtual GPU.

6.1.1. Removing a Citrix Virtual Apps and Desktops VM’s vGPU configuration

You can remove a virtual GPU assignment from a VM, such that it no longer uses a virtual GPU, by using either XenCenter or the xe command.

Note:

The VM must be in the powered-off state in order for its vGPU configuration to be modified or removed.


6.1.1.1. Removing a VM’s vGPU configuration by using XenCenter

  1. Set the GPU type to None in the VM’s GPU Properties, as shown in Figure 21.

    Figure 21. Using XenCenter to remove a vGPU configuration from a VM

    remove-vgpu-configuration-citrix-xenserver-xencenter-screen.png

  2. Click OK.

6.1.1.2. Removing a VM’s vGPU configuration by using xe

  1. Use vgpu-list to discover the vGPU object UUID associated with a given VM:
    Copy
    Copied!
                

    [root@xenserver ~]# xe vgpu-list vm-uuid=e71afda4-53f4-3a1b-6c92-a364a7f619c2 uuid ( RO) : c1c7c43d-4c99-af76-5051-119f1c2b4188 vm-uuid ( RO): e71afda4-53f4-3a1b-6c92-a364a7f619c2 gpu-group-uuid ( RO): d53526a9-3656-5c88-890b-5b24144c3d96

  2. Use vgpu-destroy to delete the virtual GPU object associated with the VM:
    Copy
    Copied!
                

    [root@xenserver ~]# xe vgpu-destroy uuid=c1c7c43d-4c99-af76-5051-119f1c2b4188 [root@xenserver ~]#

6.1.2. Removing a vSphere VM’s vGPU Configuration

To remove a vSphere vGPU configuration from a VM:

  1. Select Edit settings after right-clicking on the VM in the vCenter Web UI.
  2. Select the Virtual Hardware tab.
  3. Mouse over the PCI Device entry showing NVIDIA GRID vGPU and click on the (X) icon to mark the device for removal.
  4. Click OK to remove the device and update the VM settings.

6.2. Modifying GPU Allocation Policy

Citrix Hypervisor and VMware vSphere both support the breadth first and depth-first GPU allocation policies for vGPU-enabled VMs.

breadth-first
The breadth-first allocation policy attempts to minimize the number of vGPUs running on each physical GPU. Newly created vGPUs are placed on the physical GPU that can support the new vGPU and that has the fewest vGPUs already resident on it. This policy generally leads to higher performance because it attempts to minimize sharing of physical GPUs, but it may artificially limit the total number of vGPUs that can run.
depth-first
The depth-first allocation policy attempts to maximize the number of vGPUs running on each physical GPU. Newly created vGPUs are placed on the physical GPU that can support the new vGPU and that has the most vGPUs already resident on it. This policy generally leads to higher density of vGPUs, particularly when different types of vGPUs are being run, but may result in lower performance because it attempts to maximize sharing of physical GPUs.

Each hypervisor uses a different GPU allocation policy by default.

  • Citrix Hypervisor uses the depth-first allocation policy.
  • VMware vSphere ESXi uses the breadth-first allocation policy.

If the default GPU allocation policy does not meet your requirements for performance or density of vGPUs, you can change it.

6.2.1. Modifying GPU Allocation Policy on Citrix Hypervisor

You can modify GPU allocation policy on Citrix Hypervisor by using XenCenter or the xe command.

6.2.1.1. Modifying GPU Allocation Policy by Using xe

The allocation policy of a GPU group is stored in the allocation-algorithm parameter of the gpu-group object.

To change the allocation policy of a GPU group, use gpu-group-param-set:

Copy
Copied!
            

[root@xenserver ~]# xe gpu-group-param-get uuid=be825ba2-01d7-8d51-9780-f82cfaa64924 param-name=allocation-algorithmdepth-first [root@xenserver ~]# xe gpu-group-param-set uuid=be825ba2-01d7-8d51-9780-f82cfaa64924 allocation-algorithm=breadth-first [root@xenserver ~]#

6.2.1.2. Modifying GPU Allocation Policy GPU by Using XenCenter

You can modify GPU allocation policy from the GPU tab in XenCenter.

Figure 22. Modifying GPU placement policy in XenCenter

controlling-gpu-allocation-policy-citrix-xenserver-xencenter-screen.png

6.2.2. Modifying GPU Allocation Policy on VMware vSphere

How to switch to a depth-first allocation scheme depends on the version of VMware vSphere that you are using.

  • Supported versions earlier than 6.5: Add the following parameter to /etc/vmware/config:

    Copy
    Copied!
                

    vGPU.consolidation = true

  • Version 6.5: Use the vSphere Web Client.

Before using the vSphere Web Client to change the allocation scheme, ensure that the ESXi host is running and that all VMs on the host are powered off.

  1. Log in to vCenter Server by using the vSphere Web Client.
  2. In the navigation tree, select your ESXi host and click the Configure tab.
  3. From the menu, choose Graphics and then click the Host Graphics tab.
  4. On the Host Graphics tab, click Edit.

    Figure 23. Breadth-first allocation scheme setting for vGPU-enabled VMs

    configure-host-graphics-tab-shared-mode-vmware-vsphere.png

  5. In the Edit Host Graphics Settings dialog box that opens, select these options and click OK.
    1. If not already selected, select Shared Direct.
    2. Select Group VMs on GPU until full.

    Figure 24. Host graphics settings for vGPU

    shared-to-shared-direct-mode-depth-first-vmware-vsphere.png

    After you click OK, the default graphics type changes to Shared Direct and the allocation scheme for vGPU-enabled VMs is breadth-first.

    Figure 25. Depth-first allocation scheme setting for vGPU-enabled VMs

    configure-host-graphics-tab-shared-direct-mode-depth-first-vmware-vsphere.png

  6. Restart the ESXi host or the Xorg service on the host.

See also the following topics in the VMware vSphere documentation:

6.3. Migrating a VM Configured with vGPU

On some hypervisors, NVIDIA vGPU software supports migration of VMs that are configured with vGPU.
Before migrating a VM configured with vGPU, ensure that the following prerequisites are met:

  • The VM is configured with vGPU.
  • The VM is running.
  • The VM obtained a suitable vGPU license when it was booted.
  • The destination host has a physical GPU of the same type as the GPU where the vGPU currently resides.
  • ECC memory configuration (enabled or disabled) on both the source and destination hosts must be identical.
  • The GPU topologies (including NVLink widths) on both the source and destination hosts must be identical.
Note:

vGPU migration is disabled for a VM for which any of the following NVIDIA CUDA Toolkit features is enabled:

  • Unified memory
  • Debuggers
  • Profilers


How to migrate a VM configured with vGPU depends on the hypervisor that you are using.

After migration, the vGPU type of the vGPU remains unchanged.

The time required for migration depends on the amount of frame buffer that the vGPU has. Migration for a vGPU with a large amount of frame buffer is slower than for a vGPU with a small amount of frame buffer.

6.3.1. Migrating a VM Configured with vGPU on Citrix Hypervisor

NVIDIA vGPU software supports XenMotion for VMs that are configured with vGPU. XenMotion enables you to move a running virtual machine from one physical host machine to another host with very little disruption or downtime. For a VM that is configured with vGPU, the vGPU is migrated with the VM to an NVIDIA GPU on the other host. The NVIDIA GPUs on both host machines must be of the same type.

For details about which Citrix Hypervisor versions, NVIDIA GPUs, and guest OS releases support XenMotion with vGPU, see Virtual GPU Software for Citrix Hypervisor Release Notes.

For best performance, the physical hosts should be configured to use the following:

  • Shared storage, such as NFS, iSCSI, or Fiberchannel

    If shared storage is not used, migration can take a very long time because vDISK must also be migrated.

  • 10 GB networking.
  1. In Citrix XenCenter, context-click the VM and from the menu that opens, choose Migrate.
  2. From the list of available hosts, select the destination host to which you want to migrate the VM. The destination host must have a physical GPU of the same type as the GPU where the vGPU currently resides. Furthermore, the physical GPU must be capable of hosting the vGPU. If these requirements are not met, no available hosts are listed.

6.3.2. Migrating a VM Configured with vGPU on VMware vSphere

NVIDIA vGPU software supports VMware vMotion for VMs that are configured with vGPU. VMware vMotion enables you to move a running virtual machine from one physical host machine to another host with very little disruption or downtime. For a VM that is configured with vGPU, the vGPU is migrated with the VM to an NVIDIA GPU on the other host. The NVIDIA GPUs on both host machines must be of the same type.

For details about which VMware vSphere versions, NVIDIA GPUs, and guest OS releases support suspend and resume, see Virtual GPU Software for VMware vSphere Release Notes.

Perform this task in the VMware vSphere web client by using the Migration wizard.
Before migrating a VM configured with vGPU on VMware vSphere, ensure that the following prerequisites are met:

  1. Context-click the VM and from the menu that opens, choose Migrate.
  2. For the type of migration, select Change compute resource only and click Next. If you select Change both compute resource and storage, the time required for the migration increases.
  3. Select the destination host and click Next. The destination host must have a physical GPU of the same type as the GPU where the vGPU currently resides. Furthermore, the physical GPU must be capable of hosting the vGPU. If these requirements are not met, no available hosts are listed.
  4. Select the destination network and click Next.
  5. Select the migration priority level and click Next.
  6. Review your selections and click Finish.

For more information, see the following topics in the VMware documentation:

If NVIDIA vGPU migration is not configured, any attempt to migrate a VM with an NVIDIA vGPU fails and a window containing the following error message is displayed:

Copy
Copied!
            

Compatibility Issue/Host Migration was temporarily disabled due to another migration activity. vGPU hot migration is not enabled.

The window appears as follows:

vgpu-hot-migration-not-enabled.jpg

If you see this error, configure NVIDIA vGPU migration as explained in Configuring VMware vMotion with vGPU for VMware vSphere.

If your version of VMware vSpehere ESXi does not support vMotion for VMs configured with NVIDIA vGPU, any attempt to migrate a VM with an NVIDIA vGPU fails and a window containing the following error message is displayed:

Copy
Copied!
            

Compatibility Issues ... A required migration feature is not supported on the "Source" host 'host-name'. A warning or error occurred when migrating the virtual machine. Virtual machine relocation, or power on after relocation or cloning can fail if vGPU resources are not available on the destination host.

The window appears as follows:

vgpu-migration-compatibility-issues.png

For details about which VMware vSphere versions, NVIDIA GPUs, and guest OS releases support suspend and resume, see Virtual GPU Software for VMware vSphere Release Notes.

6.3.3. Suspending and Resuming a VM Configured with vGPU on VMware vSphere

NVIDIA vGPU software supports suspend and resume for VMs that are configured with vGPU.

For details about which VMware vSphere versions, NVIDIA GPUs, and guest OS releases support suspend and resume, see Virtual GPU Software for VMware vSphere Release Notes.

Perform this task in the VMware vSphere web client.

  • To suspend a VM, context-click the VM that you want to suspend, and from the context menu that pops up, choose Power > Suspend.
  • To resume a VM, context-click the VM that you want to resume, and from the context menu that pops up, choose Power > Power On.

6.4. Modifying a MIG-Backed vGPU's Configuration

If compute instances weren't created within the GPU instances when the GPU was configured for MIG-backed vGPUs, you can add the compute instances for an individual vGPU from within the guest VM. If you want to replace the compute instances that were created when the GPU was configured for MIG-backed vGPUs, you can delete them before adding the compute instances from within the guest VM.

Ensure that the following prerequisites are met:

  • You have root user privileges in the guest VM.
  • The GPU instance is not being used by any other processes, such as CUDA applications, monitoring applications, or the nvidia-smi command.

Perform this task in a guest VM command shell.

  1. Open a command shell as the root user in the guest VM. You can use secure shell (SSH) for this purpose.
  2. List the available GPU instance.
    Copy
    Copied!
                

    $ nvidia-smi mig -lgi +----------------------------------------------------+ | GPU instances: | | GPU Name Profile Instance Placement | | ID ID Start:Size | |====================================================| | 0 MIG 2g.10gb 0 0 0:8 | +----------------------------------------------------+

  3. Optional: If compute instances were created when the GPU was configured for MIG-backed vGPUs that you no longer require, delete them.
    Copy
    Copied!
                

    $ nvidia-smi mig -dci -ci compute-instance-id -gi gpu-instance-id

    compute-instance-id
    The ID of the compute instance that you want to delete.
    gpu-instance-id
    The ID of the GPU instance from which you want to delete the compute instance.
    Note:

    If the GPU instance is being used by another process, this command fails. In this situation, stop all processes that are using the GPU instance and retry the command.

    This example deletes compute instance 0 from GPU instance 0 on GPU 0.

    Copy
    Copied!
                

    $ nvidia-smi mig -dci -ci 0 -gi 0 Successfully destroyed compute instance ID 0 from GPU 0 GPU instance ID 0

  4. List the compute instance profiles that are available for your GPU instance.
    Copy
    Copied!
                

    $ nvidia-smi mig -lcip

    This example shows that one MIG 2g.10gb compute instance or two MIG 1c.2g.10gb compute instances can be created within the GPU instance.
    Copy
    Copied!
                

    $ nvidia-smi mig -lcip +-------------------------------------------------------------------------------+ | Compute instance profiles: | | GPU GPU Name Profile Instances Exclusive Shared | | Instance ID Free/Total SM DEC ENC OFA | | ID CE JPEG | |===============================================================================| | 0 0 MIG 1c.2g.10gb 0 2/2 14 1 0 0 | | 2 0 | +-------------------------------------------------------------------------------+ | 0 0 MIG 2g.10gb 1* 1/1 28 1 0 0 | | 2 0 | +-------------------------------------------------------------------------------+

  5. Create the compute instances that you need within the available GPU instance.

    Create each compute instance individually by running the following command.

    Copy
    Copied!
                

    $ nvidia-smi mig -cci compute-instance-profile-id -gi gpu-instance-id

    compute-instance-profile-id
    The compute instance profile ID that specifies the compute instance.
    gpu-instance-id
    The GPU instance ID that specifies the GPU instance within which you want to create the compute instance.
    Note:

    If the GPU instance is being used by another process, this command fails. In this situation, stop all processes that are using the GPU and retry the command.

    This example creates a MIG 2g.10gb compute instance on GPU instance 0.

    Copy
    Copied!
                

    $ nvidia-smi mig -cci 1 -gi 0 Successfully created compute instance ID 0 on GPU 0 GPU instance ID 0 using profile MIG 2g.10gb (ID 1)

    This example creates two MIG 1c.2g.10gb compute instances on GPU instance 0 by running the same command twice.

    Copy
    Copied!
                

    $ nvidia-smi mig -cci 0 -gi 0 Successfully created compute instance ID 0 on GPU 0 GPU instance ID 0 using profile MIG 1c.2g.10gb (ID 0) $ nvidia-smi mig -cci 0 -gi 0 Successfully created compute instance ID 1 on GPU 0 GPU instance ID 0 using profile MIG 1c.2g.10gb (ID 0)

  6. Verify that the compute instances were created within the GPU instance. Use the nvidia-smi command for this purpose.

    This example confirms that a MIG 2g.10gb compute instance was created on GPU instance 0.

    Copy
    Copied!
                

    nvidia-smi Mon Jul 15 19:01:24 2024 +-----------------------------------------------------------------------------+ | NVIDIA-SMI 470.256.02 Driver Version: 470.256.02 CUDA Version: 11.4 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |===============================+======================+======================| | 0 GRID A100X-2-10C On | 00000000:00:08.0 Off | On | | N/A N/A P0 N/A / N/A | 1058MiB / 10235MiB | N/A Default | | | | Enabled | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | MIG devices: | +------------------+----------------------+-----------+-----------------------+ | GPU GI CI MIG | Memory-Usage | Vol| Shared | | ID ID Dev | BAR1-Usage | SM Unc| CE ENC DEC OFA JPG| | | | ECC| | |==================+======================+===========+=======================| | 0 0 0 0 | 1058MiB / 10235MiB | 28 0 | 2 0 1 0 0 | | | 0MiB / 4096MiB | | | +------------------+----------------------+-----------+-----------------------+ +-----------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=============================================================================| | No running processes found | +-----------------------------------------------------------------------------+

    This example confirms that two MIG 1c.2g.10gb compute instances were created on GPU instance 0.

    Copy
    Copied!
                

    $ nvidia-smi Mon Jul 15 19:01:24 2024 +-----------------------------------------------------------------------------+ | NVIDIA-SMI 470.256.02 Driver Version: 470.256.02 CUDA Version: 11.4 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |===============================+======================+======================| | 0 GRID A100X-2-10C On | 00000000:00:08.0 Off | On | | N/A N/A P0 N/A / N/A | 1058MiB / 10235MiB | N/A Default | | | | Enabled | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | MIG devices: | +------------------+----------------------+-----------+-----------------------+ | GPU GI CI MIG | Memory-Usage | Vol| Shared | | ID ID Dev | BAR1-Usage | SM Unc| CE ENC DEC OFA JPG| | | | ECC| | |==================+======================+===========+=======================| | 0 0 0 0 | 1058MiB / 10235MiB | 14 0 | 2 0 1 0 0 | | | 0MiB / 4096MiB | | | +------------------+ +-----------+-----------------------+ | 0 0 1 1 | | 14 0 | 2 0 1 0 0 | | | | | | +------------------+----------------------+-----------+-----------------------+ +-----------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=============================================================================| | No running processes found | +-----------------------------------------------------------------------------+

6.5. Enabling Unified Memory for a vGPU

Unified memory is disabled by default. If used, you must enable unified memory individually for each vGPU that requires it by setting a vGPU plugin parameter. How to enable unified memory for a vGPU depends on the hypervisor that you are using.

6.5.1. Enabling Unified Memory for a vGPU on Citrix Hypervisor

On Citrix Hypervisor, enable unified memory by setting the enable_uvm vGPU plugin parameter.

Perform this task for each vGPU that requires unified memory by using the xe command.

Set the enable_uvm vGPU plugin parameter for the vGPU to 1 as explained in Setting vGPU Plugin Parameters on Citrix Hypervisor.

This example enables unified memory for the vGPU that has the UUID d15083f8-5c59-7474-d0cb-fbc3f7284f1b.

Copy
Copied!
            

[root@xenserver ~] xe vgpu-param-set uuid=d15083f8-5c59-7474-d0cb-fbc3f7284f1b extra_args='enable_uvm=1'

6.5.2. Enabling Unified Memory for a vGPU on Red Hat Enterprise Linux KVM

On Red Hat Enterprise Linux KVM, enable unified memory by setting the enable_uvm vGPU plugin parameter.

Ensure that the mdev device file that represents the vGPU has been created as explained in Creating an NVIDIA vGPU on a Linux with KVM Hypervisor.
Perform this task for each vGPU that requires unified memory.
Set the enable_uvm vGPU plugin parameter for the mdev device file that represents the vGPU to 1 as explained in Setting vGPU Plugin Parameters on a Linux with KVM Hypervisor.

6.5.3. Enabling Unified Memory for a vGPU on VMware vSphere

On VMware vSphere, enable unified memory by setting the pciPassthruvgpu-id.cfg.enable_uvm configuration parameter in advanced VM attributes.

Ensure that the VM to which the vGPU is assigned is powered off.
Perform this task in the vSphere Client for each vGPU that requires unified memory.

In advanced VM attributes, set the pciPassthruvgpu-id.cfg.enable_uvm vGPU plugin parameter for the vGPU to 1 as explained in Setting vGPU Plugin Parameters on VMware vSphere.

vgpu-id
A positive integer that identifies the vGPU assigned to a VM. For the first vGPU assigned to a VM, vgpu-id is 0. For example, if two vGPUs are assigned to a VM and you are enabling unified memory for both vGPUs, set pciPassthru0.cfg.enable_uvm and pciPassthru1.cfg.enable_uvm to 1.

6.6. Enabling NVIDIA CUDA Toolkit Development Tools for NVIDIA vGPU

By default, NVIDIA CUDA Toolkit development tools are disabled on NVIDIA vGPU. If used, you must enable NVIDIA CUDA Toolkit development tools individually for each VM that requires them by setting vGPU plugin parameters. One parameter must be set for enabling NVIDIA CUDA Toolkit debuggers and a different parameter must be set for enabling NVIDIA CUDA Toolkit profilers.

6.6.1. Enabling NVIDIA CUDA Toolkit Debuggers for NVIDIA vGPU

By default, NVIDIA CUDA Toolkit debuggers are disabled. If used, you must enable them for each vGPU VM that requires them by setting a vGPU plugin parameter. How to set the parameter to enable NVIDIA CUDA Toolkit debuggers for a vGPU VM depends on the hypervisor that you are using.

You can enable NVIDIA CUDA Toolkit debuggers for any number of VMs configured with vGPUs on the same GPU. When NVIDIA CUDA Toolkit debuggers are enabled for a VM, the VM cannot be migrated.

Perform this task for each VM for which you want to enable NVIDIA CUDA Toolkit debuggers.

Enabling NVIDIA CUDA Toolkit Debuggers for NVIDIA vGPU on Citrix Hypervisor

Set the enable_debugging vGPU plugin parameter for the vGPU that is assigned to the VM to 1 as explained in Setting vGPU Plugin Parameters on Citrix Hypervisor.

This example enables NVIDIA CUDA Toolkit debuggers for the vGPU that has the UUID d15083f8-5c59-7474-d0cb-fbc3f7284f1b.

Copy
Copied!
            

[root@xenserver ~] xe vgpu-param-set uuid=d15083f8-5c59-7474-d0cb-fbc3f7284f1b extra_args='enable_debugging=1'

The setting of this parameter is preserved after a guest VM is restarted and after the hypervisor host is restarted.

Enabling NVIDIA CUDA Toolkit Debuggers for NVIDIA vGPU on Red Hat Enterprise Linux KVM

Set the enable_debugging vGPU plugin parameter for the mdev device file that represents the vGPU that is assigned to the VM to 1 as explained in Setting vGPU Plugin Parameters on a Linux with KVM Hypervisor.

The setting of this parameter is preserved after a guest VM is restarted. However, this parameter is reset to its default value after the hypervisor host is restarted.

Enabling NVIDIA CUDA Toolkit Debuggers for NVIDIA vGPU on on VMware vSphere

Ensure that the VM for which you want to enable NVIDIA CUDA Toolkit debuggers is powered off.

In advanced VM attributes, set the pciPassthruvgpu-id.cfg.enable_debugging vGPU plugin parameter for the vGPU that is assigned to the VM to 1 as explained in Setting vGPU Plugin Parameters on VMware vSphere.

vgpu-id
A positive integer that identifies the vGPU assigned to the VM. For the first vGPU assigned to a VM, vgpu-id is 0. For example, if two vGPUs are assigned to a VM and you are enabling debuggers for both vGPUs, set pciPassthru0.cfg.enable_debugging and pciPassthru1.cfg.enable_debugging to 1.

The setting of this parameter is preserved after a guest VM is restarted. However, this parameter is reset to its default value after the hypervisor host is restarted.

6.6.2. Enabling NVIDIA CUDA Toolkit Profilers for NVIDIA vGPU

By default, only GPU workload trace is enabled. If you want to use all NVIDIA CUDA Toolkit profiler features that NVIDIA vGPU supports, you must enable them for each vGPU VM that requires them.

Note:

Enabling profiling for a VM gives the VM access to the GPU’s global performance counters, which may include activity from other VMs executing on the same GPU. Enabling profiling for a VM also allows the VM to lock clocks on the GPU, which impacts all other VMs executing on the same GPU.


6.6.2.1. Supported NVIDIA CUDA Toolkit Profiler Features

You can enable the following NVIDIA CUDA Toolkit profiler features for a vGPU VM:

  • NVIDIA Nsight™ Compute
  • NVIDIA Nsight Systems
  • CUDA Profiling Tools Interface (CUPTI)

6.6.2.2. Clock Management for a vGPU VM for Which NVIDIA CUDA Toolkit Profilers Are Enabled

Clocks are not locked for periodic sampling use cases such as NVIDIA Nsight Systems profiling.

Clocks are locked for multipass profiling such as:

  • NVIDIA Nsight Compute kernel profiling
  • CUPTI range profiling

Clocks are locked automatically when profiling starts and are unlocked automatically when profiling ends.

6.6.2.3. Limitations on the Use of NVIDIA CUDA Toolkit Profilers with NVIDIA vGPU

The following limitations apply when NVIDIA CUDA Toolkit profilers are enabled for NVIDIA vGPU:

  • NVIDIA CUDA Toolkit profilers can be used on only one VM at a time.
  • Multiple CUDA contexts cannot be profiled simultaneously.
  • Profiling data is collected separately for each context.
  • A VM for which NVIDIA CUDA Toolkit profilers are enabled cannot be migrated.

Because NVIDIA CUDA Toolkit profilers can be used on only one VM at a time, you should enable them for only one VM assigned a vGPU on a GPU. However, NVIDIA vGPU software cannot enforce this requirement. If NVIDIA CUDA Toolkit profilers are enabled on more than one VM assigned a vGPU on a GPU, profiling data is collected only for the first VM to start the profiler.

6.6.2.4. Enabling NVIDIA CUDA Toolkit Profilers for a vGPU VM

You enable NVIDIA CUDA Toolkit profilers for a vGPU VM by setting a vGPU plugin parameter. How to set the parameter to enable NVIDIA CUDA Toolkit profilers for a vGPU VM depends on the hypervisor that you are using.

Perform this task for the VM for which you want to enable NVIDIA CUDA Toolkit profilers.

Enabling NVIDIA CUDA Toolkit Profilers for NVIDIA vGPU on Citrix Hypervisor

Set the enable_profiling vGPU plugin parameter for the vGPU that is assigned to the VM to 1 as explained in Setting vGPU Plugin Parameters on Citrix Hypervisor.

This example enables NVIDIA CUDA Toolkit profilers for the vGPU that has the UUID d15083f8-5c59-7474-d0cb-fbc3f7284f1b.

Copy
Copied!
            

[root@xenserver ~] xe vgpu-param-set uuid=d15083f8-5c59-7474-d0cb-fbc3f7284f1b extra_args='enable_profiling=1'

The setting of this parameter is preserved after a guest VM is restarted and after the hypervisor host is restarted.

Enabling NVIDIA CUDA Toolkit Profilers for NVIDIA vGPU on Red Hat Enterprise Linux KVM

Set the enable_profiling vGPU plugin parameter for the mdev device file that represents the vGPU that is assigned to the VM to 1 as explained in Setting vGPU Plugin Parameters on a Linux with KVM Hypervisor.

The setting of this parameter is preserved after a guest VM is restarted. However, this parameter is reset to its default value after the hypervisor host is restarted.

Enabling NVIDIA CUDA Toolkit Profilers for NVIDIA vGPU on on VMware vSphere

Ensure that the VM for which you want to enable NVIDIA CUDA Toolkit profilers is powered off.

In advanced VM attributes, set the pciPassthruvgpu-id.cfg.enable_profiling vGPU plugin parameter for the vGPU that is assigned to the VM to 1 as explained in Setting vGPU Plugin Parameters on VMware vSphere.

vgpu-id
A positive integer that identifies the vGPU assigned to the VM. For the first vGPU assigned to a VM, vgpu-id is 0. For example, if two vGPUs are assigned to a VM and you are enabling profilers for the second vGPU, set pciPassthru1.cfg.enable_profiling to 1.

The setting of this parameter is preserved after a guest VM is restarted. However, this parameter is reset to its default value after the hypervisor host is restarted.

NVIDIA vGPU software enables you to monitor the performance of physical GPUs and virtual GPUs from the hypervisor and from within individual guest VMs. You can use several tools for monitoring GPU performance:

  • From any supported hypervisor, and from a guest VM that is running a 64-bit edition of Windows or Linux, you can use NVIDIA System Management Interface, nvidia-smi.
  • From Citrix Hypervisor, you can use Citrix XenCenter.
  • From a Windows guest VM, you can use these tools:
    • Windows Performance Monitor
    • Windows Management Instrumentation (WMI)

7.1. NVIDIA System Management Interface nvidia-smi

NVIDIA System Management Interface, nvidia-smi, is a command-line tool that reports management information for NVIDIA GPUs.

The nvidia-smi tool is included in the following packages:

  • NVIDIA Virtual GPU Manager package for each supported hypervisor
  • NVIDIA driver package for each supported guest OS

The scope of the reported management information depends on where you run nvidia-smi from:

  • From a hypervisor command shell, such as the Citrix Hypervisor dom0 shell or VMware ESXi host shell, nvidia-smi reports management information for NVIDIA physical GPUs and virtual GPUs present in the system.

    Note:

    When run from a hypervisor command shell, nvidia-smi will not list any GPU that is currently allocated for GPU pass-through.

  • From a guest VM, nvidia-smi retrieves usage statistics for vGPUs or pass-through GPUs that are assigned to the VM.

    From a Windows guest VM, you can run nvidia-smi from a command prompt by changing to the C:\Program Files\NVIDIA Corporation\NVSMI folder and running the nvidia-smi.exe command.

7.2. Monitoring GPU Performance from a Hypervisor

You can monitor GPU performance from any supported hypervisor by using the NVIDIA System Management Interface nvidia-smi command-line utility. On Citrix Hypervisor platforms, you can also use Citrix XenCenter to monitor GPU performance.

Note:

You cannot monitor from the hypervisor the performance of GPUs that are being used for GPU pass-through. You can monitor the performance of pass-through GPUs only from within the guest VM that is using them.


7.2.1. Using nvidia-smi to Monitor GPU Performance from a Hypervisor

You can get management information for the NVIDIA physical GPUs and virtual GPUs present in the system by running nvidia-smi from a hypervisor command shell such as the Citrix Hypervisor dom0 shell or the VMware ESXi host shell.

Without a subcommand, nvidia-smi provides management information for physical GPUs. To examine virtual GPUs in more detail, use nvidia-smi with the vgpu subcommand.

From the command line, you can get help information about the nvidia-smi tool and the vgpu subcommand.

Help Information Command
A list of subcommands supported by the nvidia-smi tool. Note that not all subcommands apply to GPUs that support NVIDIA vGPU software. nvidia-smi -h
A list of all options supported by the vgpu subcommand. nvidia-smi vgpu –h


7.2.1.1. Getting a Summary of all Physical GPUs in the System

To get a summary of all physical GPUs in the system, along with PCI bus IDs, power state, temperature, current memory usage, and so on, run nvidia-smi without additional arguments.

Each vGPU instance is reported in the Compute processes section, together with its physical GPU index and the amount of frame-buffer memory assigned to it.

In the example that follows, three vGPUs are running in the system: One vGPU is running on each of the physical GPUs 0, 1, and 2.

Copy
Copied!
            

[root@vgpu ~]# nvidia-smi Fri Jul 12 09:26:18 2024 +-----------------------------------------------------------------------------+ | NVIDIA-SMI 470.256.02 Driver Version: 470.256.02 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 Tesla M60 On | 0000:83:00.0 Off | Off | | N/A 31C P8 23W / 150W | 1889MiB / 8191MiB | 7% Default | +-------------------------------+----------------------+----------------------+ | 1 Tesla M60 On | 0000:84:00.0 Off | Off | | N/A 26C P8 23W / 150W | 926MiB / 8191MiB | 9% Default | +-------------------------------+----------------------+----------------------+ | 2 Tesla M10 On | 0000:8A:00.0 Off | N/A | | N/A 23C P8 10W / 53W | 1882MiB / 8191MiB | 12% Default | +-------------------------------+----------------------+----------------------+ | 3 Tesla M10 On | 0000:8B:00.0 Off | N/A | | N/A 26C P8 10W / 53W | 10MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ | 4 Tesla M10 On | 0000:8C:00.0 Off | N/A | | N/A 34C P8 10W / 53W | 10MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ | 5 Tesla M10 On | 0000:8D:00.0 Off | N/A | | N/A 32C P8 10W / 53W | 10MiB / 8191MiB | 0% Default | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | 0 11924 C+G /usr/lib64/xen/bin/vgpu 1856MiB | | 1 11903 C+G /usr/lib64/xen/bin/vgpu 896MiB | | 2 11908 C+G /usr/lib64/xen/bin/vgpu 1856MiB | +-----------------------------------------------------------------------------+ [root@vgpu ~]#

7.2.1.2. Getting a Summary of all vGPUs in the System

To get a summary of the vGPUs currently that are currently running on each physical GPU in the system, run nvidia-smi vgpu without additional arguments.

Copy
Copied!
            

[root@vgpu ~]# nvidia-smi vgpu Fri Jul 12 09:27:06 2024 +-----------------------------------------------------------------------------+ | NVIDIA-SMI 470.256.02 Driver Version: 470.256.02 | |-------------------------------+--------------------------------+------------+ | GPU Name | Bus-Id | GPU-Util | | vGPU ID Name | VM ID VM Name | vGPU-Util | |===============================+================================+============| | 0 Tesla M60 | 0000:83:00.0 | 7% | | 11924 GRID M60-2Q | 3 Win7-64 GRID test 2 | 6% | +-------------------------------+--------------------------------+------------+ | 1 Tesla M60 | 0000:84:00.0 | 9% | | 11903 GRID M60-1B | 1 Win8.1-64 GRID test 3 | 8% | +-------------------------------+--------------------------------+------------+ | 2 Tesla M10 | 0000:8A:00.0 | 12% | | 11908 GRID M10-2Q | 2 Win7-64 GRID test 1 | 10% | +-------------------------------+--------------------------------+------------+ | 3 Tesla M10 | 0000:8B:00.0 | 0% | +-------------------------------+--------------------------------+------------+ | 4 Tesla M10 | 0000:8C:00.0 | 0% | +-------------------------------+--------------------------------+------------+ | 5 Tesla M10 | 0000:8D:00.0 | 0% | +-------------------------------+--------------------------------+------------+ [root@vgpu ~]#

7.2.1.3. Getting vGPU Details

To get detailed information about all the vGPUs on the platform, run nvidia-smi vgpu with the –q or --query option.

To limit the information retrieved to a subset of the GPUs on the platform, use the –i or --id option to select one or more GPUs.

Copy
Copied!
            

[root@vgpu ~]# nvidia-smi vgpu -q -i 1 GPU 00000000:86:00.0 Active vGPUs : 1 vGPU ID : 3251634178 VM ID : 1 VM Name : Win7 vGPU Name : GRID M60-8Q vGPU Type : 22 vGPU UUID : b8c6d0e1-d167-11e8-b8c9-55705e5a54a6 Guest Driver Version : 411.81 License Status : Unlicensed Accounting Mode : Disabled Accounting Buffer Size: 4000 Frame Rate Limit : 3 FPS FB Memory Usage : Total : 8192 MiB Used : 675 MiB Free : 7517 MiB Utilization : Gpu : 3 % Memory : 0 % Encoder : 0 % Decoder : 0 % Encoder Stats : Active Sessions : 0 Average FPS : 0 Average Latency : 0 FBC Stats : Active Sessions : 1 Average FPS : 227 Average Latency : 4403 [root@vgpu ~]#

7.2.1.4. Monitoring vGPU engine usage

To monitor vGPU engine usage across multiple vGPUs, run nvidia-smi vgpu with the –u or --utilization option.

For each vGPU, the usage statistics in the following table are reported once every second. The table also shows the name of the column in the command output under which each statistic is reported.

Statistic Column
3D/Compute sm
Memory controller bandwidth mem
Video encoder enc
Video decoder dec

Each reported percentage is the percentage of the physical GPU’s capacity that a vGPU is using. For example, a vGPU that uses 20% of the GPU’s graphics engine’s capacity will report 20%.

To modify the reporting frequency, use the –l or --loop option.

To limit monitoring to a subset of the GPUs on the platform, use the –i or --id option to select one or more GPUs.

Copy
Copied!
            

[root@vgpu ~]# nvidia-smi vgpu -u # gpu vgpu sm mem enc dec # Idx Id % % % % 0 11924 6 3 0 0 1 11903 8 3 0 0 2 11908 10 4 0 0 3 - - - - - 4 - - - - - 5 - - - - - 0 11924 6 3 0 0 1 11903 9 3 0 0 2 11908 10 4 0 0 3 - - - - - 4 - - - - - 5 - - - - - 0 11924 6 3 0 0 1 11903 8 3 0 0 2 11908 10 4 0 0 3 - - - - - 4 - - - - - 5 - - - - - ^C[root@vgpu ~]#

7.2.1.5. Monitoring vGPU engine usage by applications

To monitor vGPU engine usage by applications across multiple vGPUs, run nvidia-smi vgpu with the –p option.

For each application on each vGPU, the usage statistics in the following table are reported once every second. Each application is identified by its process ID and process name. The table also shows the name of the column in the command output under which each statistic is reported.

Statistic Column
3D/Compute sm
Memory controller bandwidth mem
Video encoder enc
Video decoder dec

Each reported percentage is the percentage of the physical GPU’s capacity used by an application running on a vGPU that resides on the physical GPU. For example, an application that uses 20% of the GPU’s graphics engine’s capacity will report 20%.

To modify the reporting frequency, use the –l or --loop option.

To limit monitoring to a subset of the GPUs on the platform, use the –i or --id option to select one or more GPUs.

Copy
Copied!
            

[root@vgpu ~]# nvidia-smi vgpu -p # GPU vGPU process process sm mem enc dec # Idx Id Id name % % % % 0 38127 1528 dwm.exe 0 0 0 0 1 37408 4232 DolphinVS.exe 32 25 0 0 1 257869 4432 FurMark.exe 16 12 0 0 1 257969 4552 FurMark.exe 48 37 0 0 0 38127 1528 dwm.exe 0 0 0 0 1 37408 4232 DolphinVS.exe 16 12 0 0 1 257911 656 DolphinVS.exe 32 24 0 0 1 257969 4552 FurMark.exe 48 37 0 0 0 38127 1528 dwm.exe 0 0 0 0 1 257869 4432 FurMark.exe 38 30 0 0 1 257911 656 DolphinVS.exe 19 14 0 0 1 257969 4552 FurMark.exe 38 30 0 0 0 38127 1528 dwm.exe 0 0 0 0 1 257848 3220 Balls64.exe 16 12 0 0 1 257869 4432 FurMark.exe 16 12 0 0 1 257911 656 DolphinVS.exe 16 12 0 0 1 257969 4552 FurMark.exe 48 37 0 0 0 38127 1528 dwm.exe 0 0 0 0 1 257911 656 DolphinVS.exe 32 25 0 0 1 257969 4552 FurMark.exe 64 50 0 0 0 38127 1528 dwm.exe 0 0 0 0 1 37408 4232 DolphinVS.exe 16 12 0 0 1 257911 656 DolphinVS.exe 16 12 0 0 1 257969 4552 FurMark.exe 64 49 0 0 0 38127 1528 dwm.exe 0 0 0 0 1 37408 4232 DolphinVS.exe 16 12 0 0 1 257869 4432 FurMark.exe 16 12 0 0 1 257969 4552 FurMark.exe 64 49 0 0 [root@vgpu ~]#

7.2.1.6. Monitoring Encoder Sessions

Note:

Encoder sessions can be monitored only for vGPUs assigned to Windows VMs. No encoder session statistics are reported for vGPUs assigned to Linux VMs.


To monitor the encoder sessions for processes running on multiple vGPUs, run nvidia-smi vgpu with the –es or --encodersessions option.

For each encoder session, the following statistics are reported once every second:

  • GPU ID
  • vGPU ID
  • Encoder session ID
  • PID of the process in the VM that created the encoder session
  • Codec type, for example, H.264 or H.265
  • Encode horizontal resolution
  • Encode vertical resolution
  • One-second trailing average encoded FPS
  • One-second trailing average encode latency in microseconds

To modify the reporting frequency, use the –l or --loop option.

To limit monitoring to a subset of the GPUs on the platform, use the –i or --id option to select one or more GPUs.

Copy
Copied!
            

[root@vgpu ~]# nvidia-smi vgpu -es # GPU vGPU Session Process Codec H V Average Average # Idx Id Id Id Type Res Res FPS Latency(us) 1 21211 2 2308 H.264 1920 1080 424 1977 1 21206 3 2424 H.264 1920 1080 0 0 1 22011 1 3676 H.264 1920 1080 374 1589 1 21211 2 2308 H.264 1920 1080 360 807 1 21206 3 2424 H.264 1920 1080 325 1474 1 22011 1 3676 H.264 1920 1080 313 1005 1 21211 2 2308 H.264 1920 1080 329 1732 1 21206 3 2424 H.264 1920 1080 352 1415 1 22011 1 3676 H.264 1920 1080 434 1894 1 21211 2 2308 H.264 1920 1080 362 1818 1 21206 3 2424 H.264 1920 1080 296 1072 1 22011 1 3676 H.264 1920 1080 416 1994 1 21211 2 2308 H.264 1920 1080 444 1912 1 21206 3 2424 H.264 1920 1080 330 1261 1 22011 1 3676 H.264 1920 1080 436 1644 1 21211 2 2308 H.264 1920 1080 344 1500 1 21206 3 2424 H.264 1920 1080 393 1727 1 22011 1 3676 H.264 1920 1080 364 1945 1 21211 2 2308 H.264 1920 1080 555 1653 1 21206 3 2424 H.264 1920 1080 295 925 1 22011 1 3676 H.264 1920 1080 372 1869 1 21211 2 2308 H.264 1920 1080 326 2206 1 21206 3 2424 H.264 1920 1080 318 1366 1 22011 1 3676 H.264 1920 1080 464 2015 1 21211 2 2308 H.264 1920 1080 305 1167 1 21206 3 2424 H.264 1920 1080 445 1892 1 22011 1 3676 H.264 1920 1080 361 906 1 21211 2 2308 H.264 1920 1080 353 1436 1 21206 3 2424 H.264 1920 1080 354 1798 1 22011 1 3676 H.264 1920 1080 373 1310 ^C[root@vgpu ~]#

7.2.1.7. Monitoring Frame Buffer Capture (FBC) Sessions

To monitor the FBC sessions for processes running on multiple vGPUs, run nvidia-smi vgpu with the -fs or --fbcsessions option.

For each FBC session, the following statistics are reported once every second:

  • GPU ID
  • vGPU ID
  • FBC session ID
  • PID of the process in the VM that created the FBC session
  • Display ordinal associated with the FBC session.
  • FBC session type
  • FBC session flags
  • Capture mode
  • Maximum horizontal resolution supported by the session
  • Maximum vertical resolution supported by the session
  • Horizontal resolution requested by the caller in the capture call
  • Vertical resolution requested by the caller in the capture call
  • Moving average of new frames captured per second by the session
  • Moving average new frame capture latency in microseconds for the session

To modify the reporting frequency, use the –l or --loop option.

To limit monitoring to a subset of the GPUs on the platform, use the –i or --id option to select one or more GPUs.

Copy
Copied!
            

[root@vgpu ~]# nvidia-smi vgpu -fs # GPU vGPU Session Process Display Session Diff. Map Class. Map Capture Max H Max V H V Average Average # Idx Id Id Id Ordinal Type State State Mode Res Res Res Res FPS Latency(us) 0 - - - - - - - - - - - - - - 1 3251634178 - - - - - - - - - - - - - 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 - - - - - - - - - - - - - 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 - - - - - - - - - - - - - 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 - - - - - - - - - - - - - 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 - - - - - - - - - - - - - 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Unknown 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Unknown 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - # GPU vGPU Session Process Display Session Diff. Map Class. Map Capture Max H Max V H V Average Average # Idx Id Id Id Ordinal Type State State Mode Res Res Res Res FPS Latency(us) 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Unknown 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Unknown 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Unknown 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Unknown 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Unknown 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Unknown 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Unknown 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 1600 900 25 39964 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 1600 900 25 39964 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - # GPU vGPU Session Process Display Session Diff. Map Class. Map Capture Max H Max V H V Average Average # Idx Id Id Id Ordinal Type State State Mode Res Res Res Res FPS Latency(us) 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 1600 900 135 7400 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 1600 900 227 4403 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 1600 900 227 4403 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - # GPU vGPU Session Process Display Session Diff. Map Class. Map Capture Max H Max V H V Average Average # Idx Id Id Id Ordinal Type State State Mode Res Res Res Res FPS Latency(us) 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - 0 - - - - - - - - - - - - - - 1 3251634178 1 3984 0 ToSys Disabled Disabled Blocking 4096 2160 0 0 0 0 2 - - - - - - - - - - - - - - ^C[root@vgpu ~]#

7.2.1.8. Listing Supported vGPU Types

To list the virtual GPU types that the GPUs in the system support, run nvidia-smi vgpu with the –s or --supported option.

To limit the retrieved information to a subset of the GPUs on the platform, use the –i or --id option to select one or more GPUs.

Copy
Copied!
            

[root@vgpu ~]# nvidia-smi vgpu -s -i 0 GPU 0000:83:00.0 GRID M60-0B GRID M60-0Q GRID M60-1A GRID M60-1B GRID M60-1Q GRID M60-2A GRID M60-2Q GRID M60-4A GRID M60-4Q GRID M60-8A GRID M60-8Q [root@vgpu ~]#

To view detailed information about the supported vGPU types, add the –v or --verbose option:

Copy
Copied!
            

[root@vgpu ~]# nvidia-smi vgpu -s -i 0 -v | less GPU 00000000:83:00.0 vGPU Type ID : 0xb Name : GRID M60-0B Class : NVS Max Instances : 16 Device ID : 0x13f210de Sub System ID : 0x13f21176 FB Memory : 512 MiB Display Heads : 2 Maximum X Resolution : 2560 Maximum Y Resolution : 1600 Frame Rate Limit : 45 FPS GRID License : GRID-Virtual-PC,2.0;GRID-Virtual-WS,2.0;GRID-Virtual-WS-Ext,2.0;Quadro-Virtual-DWS,5.0 vGPU Type ID : 0xc Name : GRID M60-0Q Class : Quadro Max Instances : 16 Device ID : 0x13f210de Sub System ID : 0x13f2114c FB Memory : 512 MiB Display Heads : 2 Maximum X Resolution : 2560 Maximum Y Resolution : 1600 Frame Rate Limit : 60 FPS GRID License : GRID-Virtual-WS,2.0;GRID-Virtual-WS-Ext,2.0;Quadro-Virtual-DWS,5.0 vGPU Type ID : 0xd Name : GRID M60-1A Class : NVS Max Instances : 8 … [root@vgpu ~]#

7.2.1.9. Listing the vGPU Types that Can Currently Be Created

To list the virtual GPU types that can currently be created on GPUs in the system, run nvidia-smi vgpu with the –c or --creatable option.

This property is a dynamic property that varies for each GPU depending on whether MIG mode is enabled for the GPU.

  • If MIG mode is not enabled for the GPU, or if the GPU does not support MIG, this property reflects the number and type of vGPUs that are already running on the GPU.
    • If no vGPUs are running on the GPU, all vGPU types that the GPU supports are listed.
    • If one or more vGPUs are running on the GPU, but the GPU is not fully loaded, only the type of the vGPUs that are already running is listed.
    • If the GPU is fully loaded, no vGPU types are listed.
  • If MIG mode is enabled for the GPU, the result reflects the number and type of GPU instances on which no vGPUs are already running.
    • If no GPU instances have been created, no vGPU types are listed.
    • If GPU instances have been created, only the vGPU types that correspond to GPU instances on which no vGPU is running are listed.
    • If a vGPU is running on every GPU instance, no vGPU types are listed.

To limit the retrieved information to a subset of the GPUs on the platform, use the –i or --id option to select one or more GPUs.

Copy
Copied!
            

[root@vgpu ~]# nvidia-smi vgpu -c -i 0 GPU 0000:83:00.0 GRID M60-2Q [root@vgpu ~]#

To view detailed information about the vGPU types that can currently be created, add the –v or --verbose option.

7.2.2. Using Citrix XenCenter to monitor GPU performance

If you are using Citrix Hypervisor as your hypervisor, you can monitor GPU performance in XenCenter.

  1. Click on a server’s Performance tab.
  2. Right-click on the graph window, then select Actions and New Graph.
  3. Provide a name for the graph.
  4. In the list of available counter resources, select one or more GPU counters.

Counters are listed for each physical GPU not currently being used for GPU pass-through.

Figure 26. Using Citrix XenCenter to monitor GPU performance

vgu-performance-stats-in-citrix-xencenter.png

7.3. Monitoring GPU Performance from a Guest VM

You can use monitoring tools within an individual guest VM to monitor the performance of vGPUs or pass-through GPUs that are assigned to the VM. The scope of these tools is limited to the guest VM within which you use them. You cannot use monitoring tools within an individual guest VM to monitor any other GPUs in the platform.

For a vGPU, only these metrics are reported in a guest VM:

  • 3D/Compute
  • Memory controller
  • Video encoder
  • Video decoder
  • Frame buffer usage

Other metrics normally present in a GPU are not applicable to a vGPU and are reported as zero or N/A, depending on the tool that you are using.

7.3.1. Using nvidia-smi to Monitor GPU Performance from a Guest VM

In guest VMs, you can use the nvidia-smi command to retrieve statistics for the total usage by all applications running in the VM and usage by individual applications of the following resources:

  • GPU
  • Video encoder
  • Video decoder
  • Frame buffer

To use nvidia-smi to retrieve statistics for the total resource usage by all applications running in the VM, run the following command:

Copy
Copied!
            

nvidia-smi dmon

The following example shows the result of running nvidia-smi dmon from within a Windows guest VM.

Figure 27. Using nvidia-smi from a Windows guest VM to get total resource usage by all applications

nvidia-smi-guest-vm-windows-command-prompt.png

To use nvidia-smi to retrieve statistics for resource usage by individual applications running in the VM, run the following command:

Copy
Copied!
            

nvidia-smi pmon

Figure 28. Using nvidia-smi from a Windows guest VM to get resource usage by individual applications

nvidia-smi-guest-vm-by-application-windows-command-prompt.png

7.3.2. Using Windows Performance Counters to monitor GPU performance

In Windows VMs, GPU metrics are available as Windows Performance Counters through the NVIDIA GPU object.

Any application that is enabled to read performance counters can access these metrics. You can access these metrics directly through the Windows Performance Monitor application that is included with the Windows OS.

The following example shows GPU metrics in the Performance Monitor application.

Figure 29. Using Windows Performance Monitor to monitor GPU performance

windows-performance-monitor-screen.png

On vGPUs, the following GPU performance counters read as 0 because they are not applicable to vGPUs:

  • % Bus Usage
  • % Cooler rate
  • Core Clock MHz
  • Fan Speed
  • Memory Clock MHz
  • PCI-E current speed to GPU Mbps
  • PCI-E current width to GPU
  • PCI-E downstream width to GPU
  • Power Consumption mW
  • Temperature C

7.3.3. Using NVWMI to monitor GPU performance

In Windows VMs, Windows Management Instrumentation (WMI) exposes GPU metrics in the ROOT\CIMV2\NV namespace through NVWMI. NVWMI is included with the NVIDIA driver package. After the driver is installed, NVWMI help information in Windows Help format is available as follows:

Copy
Copied!
            

C:\Program Files\NVIDIA Corporation\NVIDIA WMI Provider>nvwmi.chm

Any WMI-enabled application can access these metrics. The following example shows GPU metrics in the third-party application WMI Explorer, which is available for download from the from the CodePlex WMI Explorer page.

Figure 30. Using WMI Explorer to monitor GPU performance

wmi-explorer-screen.png

On vGPUs, some instance properties of the following classes do not apply to vGPUs:

  • Gpu
  • PcieLink

Gpu instance properties that do not apply to vGPUs

Gpu Instance Property Value reported on vGPU
gpuCoreClockCurrent -1
memoryClockCurrent -1
pciDownstreamWidth 0
pcieGpu.curGen 0
pcieGpu.curSpeed 0
pcieGpu.curWidth 0
pcieGpu.maxGen 1
pcieGpu.maxSpeed 2500
pcieGpu.maxWidth 0
power -1
powerSampleCount -1
powerSamplingPeriod -1
verVBIOS.orderedValue 0
verVBIOS.strValue -
verVBIOS.value 0


PcieLink instance properties that do not apply to vGPUs

No instances of PcieLink are reported for vGPU.

NVIDIA GPUs based on the NVIDIA Maxwell™ graphic architecture implement a best effort vGPU scheduler that aims to balance performance across vGPUs. The best effort scheduler allows a vGPU to use GPU processing cycles that are not being used by other vGPUs. Under some circumstances, a VM running a graphics-intensive application may adversely affect the performance of graphics-light applications running in other VMs.

GPUs based on NVIDIA GPU architectures after the Maxwell architecture additionally support equal share and fixed share vGPU schedulers. These schedulers impose a limit on GPU processing cycles used by a vGPU, which prevents graphics-intensive applications running in one VM from affecting the performance of graphics-light applications running in other VMs. On GPUs that support multiple vGPU schedulers, you can select the vGPU scheduler to use. You can also set the length of the time slice for the equal share and fixed share vGPU schedulers.

Note:

If you use the equal share or fixed share vGPU scheduler, the frame-rate limiter (FRL) is disabled.

The best effort scheduler is the default scheduler for all supported GPU architectures.

If you are unsure of the NVIDIA GPU architecture of your GPU, consult the release notes for your hypervisor at NVIDIA Virtual GPU Software Documentation.

8.1. Scheduling Policies for Time-Sliced vGPUs

In addition to the default best effort scheduler, GPUs based on NVIDIA GPU architectures after the Maxwell architecture support equal share and fixed share vGPU schedulers.

Equal share scheduler
The physical GPU is shared equally amongst the running vGPUs that reside on it. As vGPUs are added to or removed from a GPU, the share of the GPU's processing cycles allocated to each vGPU changes accordingly. As a result, the performance of a vGPU may increase as other vGPUs on the same GPU are stopped, or decrease as other vGPUs are started on the same GPU.
Fixed share scheduler
Each vGPU is given a fixed share of the physical GPU's processing cycles, the amount of which depends on the vGPU type, which in turn determines the maximum number of vGPUs per physical GPU. For example, the maximum number of T4-4C vGPUs per physical GPU is 4. When the scheduling policy is fixed share, each T4-4C vGPU is given one quarter, or 25%, the physical GPU's processing cycles. As vGPUs are added to or removed from a GPU, the share of the GPU's processing cycles allocated to each vGPU remains constant. As a result, the performance of a vGPU remains unchanged as other vGPUs are stopped or started on the same GPU.

8.2. Scheduler Time Slice for Time-Sliced vGPUs

When multiple VMs access the vGPUs on a single GPU, the GPU performs the work for each VM serially. The vGPU scheduler time slice represents the amount of time that the work of a VM is allowed to run on the GPU before it is preempted and the work of the next VM is performed.

For the equal share and fixed share vGPU schedulers, you can set the length of the time slice. The length of the time slice affects latency and throughput. The optimal length of the time slice depends the workload that the GPU is handling.

  • For workloads that require low latency, a shorter time slice is optimal. Typically, these workloads are applications that must generate output at a fixed interval, such as graphics applications that generate output at a frame rate of 60 FPS. These workloads are sensitive to latency and should be allowed to run at least once per interval. A shorter time slice reduces latency and improves responsiveness by causing the scheduler to switch more frequently between VMs.

  • For workloads that require maximum throughput, a longer time slice is optimal. Typically, these workloads are applications that must complete their work as quickly as possible and do not require responsiveness, such as CUDA applications. A longer time slice increases throughput by preventing frequent switching between VMs.

8.3.  RmPVMRL Registry Key

The RmPVMRL registry key controls the scheduling behavior for NVIDIA vGPUs by setting the scheduling policy and the length of the time slice.

Note:

You can change the vGPU scheduling behavior only on GPUs that support multiple vGPU schedulers, that is, GPUs based on NVIDIA GPU architectures after the Maxwell architecture.


Type

Dword

Contents

Value Meaning
0x00 (default) Best effort scheduler
0x01 Equal share scheduler with the default time slice length
0x00TT0001 Equal share scheduler with a user-defined time slice length TT
0x11 Fixed share scheduler with the default time slice length
0x00TT0011 Fixed share scheduler with a user-defined time slice length TT

The default time slice length depends on the maximum number of vGPUs per physical GPU allowed for the vGPU type.

Maximum Number of vGPUs Default Time Slice Length
Less than or equal to 8 2 ms
Greater than 8 1 ms

TT
Two hexadecimal digits in the range 01 to 1E that set the length of the time slice in milliseconds (ms) for the equal share and fixed share schedulers. The minimum length is 1 ms and the maximum length is 30 ms.

If TT is 00, the length is set to the default length for the vGPU type.

If TT is greater than 1E, the length is set to 30 ms.

Examples

This example sets the vGPU scheduler to equal share scheduler with the default time slice length.

Copy
Copied!
            

RmPVMRL=0x01

This example sets the vGPU scheduler to equal share scheduler with a time slice that is 3 ms long.

Copy
Copied!
            

RmPVMRL=0x00030001

This example sets the vGPU scheduler to fixed share scheduler with the default time slice length.

Copy
Copied!
            

RmPVMRL=0x11

This example sets the vGPU scheduler to fixed share scheduler with a time slice that is 24 (0x18) ms long.

Copy
Copied!
            

RmPVMRL=0x00180011

8.4. Getting the Current Time-Sliced vGPU Scheduling Behavior for All GPUs

Get the current scheduling behavior before changing the scheduling behavior of one or more GPUs to determine if you need to change it or after changing it to confirm the change.

Perform this task in your hypervisor command shell.

  1. Open a command shell on your hypervisor host machine. On all supported hypervisors, you can use secure shell (SSH) for this purpose. Individual hypervisors may provide additional means for logging in. For details, refer to the documentation for your hypervisor.
  2. Use the dmesg command to display messages from the kernel that contain the strings NVRM and scheduler.
    Copy
    Copied!
                

    $ dmesg | grep NVRM | grep scheduler

    The scheduling behavior is indicated in these messages by the following strings:

    • BEST_EFFORT
    • EQUAL_SHARE
    • FIXED_SHARE

    If the scheduling behavior is equal share or fixed share, the scheduler time slice in ms is also displayed.

    This example gets the scheduling behavior of the GPUs in a system in which the behavior of one GPU is set to best effort, one GPU is set to equal share, and one GPU is set to fixed share.

    Copy
    Copied!
                

    $ dmesg | grep NVRM | grep scheduler 2020-10-05T02:58:08.928Z cpu79:2100753)NVRM: GPU at 0000:3d:00.0 has software scheduler DISABLED with policy BEST_EFFORT. 2020-10-05T02:58:09.818Z cpu79:2100753)NVRM: GPU at 0000:5e:00.0 has software scheduler ENABLED with policy EQUAL_SHARE. NVRM: Software scheduler timeslice set to 1 ms. 2020-10-05T02:58:12.115Z cpu79:2100753)NVRM: GPU at 0000:88:00.0 has software scheduler ENABLED with policy FIXED_SHARE. NVRM: Software scheduler timeslice set to 1 ms.

8.5. Changing the Time-Sliced vGPU Scheduling Behavior for All GPUs

Note:

You can change the vGPU scheduling behavior only on GPUs that support multiple vGPU schedulers, that is, GPUs based on NVIDIA GPU architectures after the Maxwell architecture.

Perform this task in your hypervisor command shell.

  1. Open a command shell on your hypervisor host machine. On all supported hypervisors, you can use secure shell (SSH) for this purpose. Individual hypervisors may provide additional means for logging in. For details, refer to the documentation for your hypervisor.
  2. Set the RmPVMRL registry key to the value that sets the GPU scheduling policy and the length of the time slice that you want.
    • On Citrix Hypervisor, Red Hat Enterprise Linux KVM, or Red Hat Virtualization (RHV), add the following entry to the /etc/modprobe.d/nvidia.conf file.

      Copy
      Copied!
                  

      options nvidia NVreg_RegistryDwords="RmPVMRL=value"

      If the /etc/modprobe.d/nvidia.conf file does not already exist, create it.

    • On VMware vSphere, use the esxcli set command.

      Copy
      Copied!
                  

      # esxcli system module parameters set -m nvidia -p "NVreg_RegistryDwords=RmPVMRL=value"

    value

    The value that sets the GPU scheduling policy and the length of the time slice that you want, for example:

    0x01
    Sets the vGPU scheduling policy to equal share scheduler with the default time slice length.
    0x00030001
    Sets the GPU scheduling policy to equal share scheduler with a time slice that is 3 ms long.
    0x11
    Sets the vGPU scheduling policy to fixed share scheduler with the default time slice length.
    0x00180011
    Sets the GPU scheduling policy to fixed share scheduler with a time slice that is 24 (0x18) ms long.

    For all supported values, see RmPVMRL Registry Key.

  3. Reboot your hypervisor host machine.

Confirm that the scheduling behavior was changed as required as explained in Getting the Current Time-Sliced vGPU Scheduling Behavior for All GPUs.

8.6. Changing the Time-Sliced vGPU Scheduling Behavior for Select GPUs

Note:

You can change the vGPU scheduling behavior only on GPUs that support multiple vGPU schedulers, that is, GPUs based on NVIDIA GPU architectures after the Maxwell architecture.

Perform this task in your hypervisor command shell.

  1. Open a command shell on your hypervisor host machine. On all supported hypervisors, you can use secure shell (SSH) for this purpose. Individual hypervisors may provide additional means for logging in. For details, refer to the documentation for your hypervisor.
  2. Use the lspci command to obtain the PCI domain and bus/device/function (BDF) of each GPU for which you want to change the scheduling behavior.
    • On Citrix Hypervisor, Red Hat Enterprise Linux KVM, or Red Hat Virtualization (RHV), add the -D option to display the PCI domain and the -d 10de: option to display information only for NVIDIA GPUs.
      Copy
      Copied!
                  

      # lspci -D -d 10de:

    • On VMware vSphere, pipe the output of lspci to the grep command to display information only for NVIDIA GPUs.
      Copy
      Copied!
                  

      # lspci | grep NVIDIA

    The NVIDIA GPU listed in this example has the PCI domain 0000 and BDF 86:00.0.

    Copy
    Copied!
                

    0000:86:00.0 3D controller: NVIDIA Corporation GP104GL [Tesla P4] (rev a1)

  3. Use the module parameter NVreg_RegistryDwordsPerDevice to set the pci and RmPVMRL registry keys for each GPU.
    • On Citrix Hypervisor, Red Hat Enterprise Linux KVM, or RHV, add the following entry to the /etc/modprobe.d/nvidia.conf file.

      Copy
      Copied!
                  

      options nvidia NVreg_RegistryDwordsPerDevice="pci=pci-domain:pci-bdf;RmPVMRL=value [;pci=pci-domain:pci-bdf;RmPVMRL=value...]"

      If the /etc/modprobe.d/nvidia.conf file does not already exist, create it.

    • On VMware vSphere, use the esxcli set command.

      Copy
      Copied!
                  

      # esxcli system module parameters set -m nvidia \ -p "NVreg_RegistryDwordsPerDevice=pci=pci-domain:pci-bdf;RmPVMRL=value\ [;pci=pci-domain:pci-bdf;RmPVMRL=value...]"

    For each GPU, provide the following information:

    pci-domain
    The PCI domain of the GPU.
    pci-bdf
    The PCI device BDF of the GPU.
    value

    The value that sets the GPU scheduling policy and the length of the time slice that you want, for example:

    0x01
    Sets the GPU scheduling policy to equal share scheduler with the default time slice length.
    0x00030001
    Sets the GPU scheduling policy to equal share scheduler with a time slice that is 3 ms long.
    0x11
    Sets the GPU scheduling policy to fixed share scheduler with the default time slice length.
    0x00180011
    Sets the GPU scheduling policy to fixed share scheduler with a time slice that is 24 (0x18) ms long.

    For all supported values, see RmPVMRL Registry Key.

    This example adds an entry to the /etc/modprobe.d/nvidia.conf file to change the scheduling behavior of a single GPU. The entry sets the GPU scheduling policy of the GPU at PCI domain 0000 and BDF 86:00.0 to fixed share scheduler with the default time slice length.

    Copy
    Copied!
                

    options nvidia NVreg_RegistryDwordsPerDevice= "pci=0000:86:00.0;RmPVMRL=0x11"

    This example adds an entry to the /etc/modprobe.d/nvidia.conf file to change the scheduling behavior of a single GPU. The entry sets the scheduling policy of the GPU at PCI domain 0000 and BDF 86:00.0 to fixed share scheduler with a time slice that is 24 (0x18) ms long.

    Copy
    Copied!
                

    options nvidia NVreg_RegistryDwordsPerDevice= "pci=0000:86:00.0;RmPVMRL=0x00180011"

    This example changes the scheduling behavior of a single GPU on a hypervisor host that is running VMware vSphere. The command sets the scheduling policy of the GPU at PCI domain 0000 and BDF 15:00.0 to fixed share scheduler with the default time slice length.

    Copy
    Copied!
                

    # esxcli system module parameters set -m nvidia -p \ "NVreg_RegistryDwordsPerDevice=pci=0000:15:00.0;RmPVMRL=0x11[;pci=0000:15:00.0;RmPVMRL=0x11]"

    This example changes the scheduling behavior of a single GPU on a hypervisor host that is running VMware vSphere. The command sets the scheduling policy of the GPU at PCI domain 0000 and BDF 15:00.0 to fixed share scheduler with a time slice that is 24 (0x18) ms long.

    Copy
    Copied!
                

    # esxcli system module parameters set -m nvidia -p \ "NVreg_RegistryDwordsPerDevice=pci=0000:15:00.0;RmPVMRL=0x11[;pci=0000:15:00.0;RmPVMRL=0x00180011]"

  4. Reboot your hypervisor host machine.

Confirm that the scheduling behavior was changed as required as explained in Getting the Current Time-Sliced vGPU Scheduling Behavior for All GPUs.

8.7. Restoring Default Time-Sliced vGPU Scheduler Settings

Perform this task in your hypervisor command shell.

  1. Open a command shell on your hypervisor host machine. On all supported hypervisors, you can use secure shell (SSH) for this purpose. Individual hypervisors may provide additional means for logging in. For details, refer to the documentation for your hypervisor.
  2. Unset the RmPVMRL registry key.
    • On Citrix Hypervisor, Red Hat Enterprise Linux KVM, or Red Hat Virtualization (RHV), comment out the entries in the /etc/modprobe.d/nvidia.conf file that set RmPVMRL by prefixing each entry with the # character.

    • On VMware vSphere, set the module parameter to an empty string.

      Copy
      Copied!
                  

      # esxcli system module parameters set -m nvidia -p "module-parameter="

      module-parameter
      The module parameter to set, which depends on whether the scheduling behavior was changed for all GPUs or select GPUs:
      • For all GPUs, set the NVreg_RegistryDwords module parameter.
      • For select GPUs, set the NVreg_RegistryDwordsPerDevice module parameter.

      For example, to restore default vGPU scheduler settings after they were changed for all GPUs, enter this command:

      Copy
      Copied!
                  

      # esxcli system module parameters set -m nvidia -p "NVreg_RegistryDwords="

  3. Reboot your hypervisor host machine.

This chapter describes basic troubleshooting steps for NVIDIA vGPU on Citrix Hypervisor, Red Hat Enterprise Linux KVM, Red Hat Virtualization (RHV), and VMware vSphere, and how to collect debug information when filing a bug report.

9.1. Known issues

Before troubleshooting or filing a bug report, review the release notes that accompany each driver release, for information about known issues with the current release, and potential workarounds.

9.2. Troubleshooting steps

If a vGPU-enabled VM fails to start, or doesn’t display any output when it does start, follow these steps to narrow down the probable cause.

9.2.1. Verifying the NVIDIA Kernel Driver Is Loaded

  1. Use the command that your hypervisor provides to verify that the kernel driver is loaded:
    • On Citrix Hypervisor, Red Hat Enterprise Linux KVM, and RHV, use lsmod:
      Copy
      Copied!
                  

      [root@xenserver ~]# lsmod|grep nvidia nvidia 9604895 84 i2c_core 20294 2 nvidia,i2c_i801 [root@xenserver ~]#

    • On VMware vSphere, use vmkload_mod:
      Copy
      Copied!
                  

      [root@esxi:~] vmkload_mod -l | grep nvidia nvidia 5 8420

  2. If the nvidia driver is not listed in the output, check dmesg for any load-time errors reported by the driver (see Examining NVIDIA kernel driver output).
  3. On Citrix Hypervisor, Red Hat Enterprise Linux KVM, and RHV, also use the rpm -q command to verify that the NVIDIA GPU Manager package is correctly installed.
    Copy
    Copied!
                

    rpm -q vgpu-manager-rpm-package-name

    vgpu-manager-rpm-package-name
    The RPM package name of the NVIDIA GPU Manager package, for example NVIDIA-vGPU-NVIDIA-vGPU-CitrixHypervisor-8.2-470.256.02 for Citrix Hypervisor.

    This example verifies that the NVIDIA GPU Manager package for Citrix Hypervisor is correctly installed.

    Copy
    Copied!
                

    [root@xenserver ~]# rpm –q NVIDIA-vGPU-NVIDIA-vGPU-CitrixHypervisor-8.2-470.256.02 [root@xenserver ~]# If an existing NVIDIA GRID package is already installed and you don’t select the upgrade (-U) option when installing a newer GRID package, the rpm command will return many conflict errors. Preparing packages for installation... file /usr/bin/nvidia-smi from install of NVIDIA-vGPU-NVIDIA-vGPU-CitrixHypervisor-8.2-470.256.02.x86_64 conflicts with file from package NVIDIA-vGPU-xenserver-8.2-470.223.02.x86_64 file /usr/lib/libnvidia-ml.so from install of NVIDIA-vGPU-NVIDIA-vGPU-CitrixHypervisor-8.2-470.256.02.x86_64 conflicts with file from package NVIDIA-vGPU-xenserver-8.2-470.223.02.x86_64 ...

9.2.2. Verifying that nvidia-smi works

If the NVIDIA kernel driver is correctly loaded on the physical GPU, run nvidia-smi and verify that all physical GPUs not currently being used for GPU pass-through are listed in the output. For details on expected output, see NVIDIA System Management Interface nvidia-smi.

If nvidia-smi fails to report the expected output, check dmesg for NVIDIA kernel driver messages.

9.2.3. Examining NVIDIA kernel driver output

Information and debug messages from the NVIDIA kernel driver are logged in kernel logs, prefixed with NVRM or nvidia.

Run dmesg on Citrix Hypervisor, Red Hat Enterprise Linux KVM, RHV, and VMware vSphere and check for the NVRM and nvidia prefixes:

Copy
Copied!
            

[root@xenserver ~]# dmesg | grep -E "NVRM|nvidia" [ 22.054928] nvidia: module license 'NVIDIA' taints kernel. [ 22.390414] NVRM: loading [ 22.829226] nvidia 0000:04:00.0: enabling device (0000 -> 0003) [ 22.829236] nvidia 0000:04:00.0: PCI INT A -> GSI 32 (level, low) -> IRQ 32 [ 22.829240] NVRM: This PCI I/O region assigned to your NVIDIA device is invalid: [ 22.829241] NVRM: BAR0 is 0M @ 0x0 (PCI:0000:00:04.0) [ 22.829243] NVRM: The system BIOS may have misconfigured your GPU.

9.2.4. Examining NVIDIA Virtual GPU Manager Messages

Information and debug messages from the NVIDIA Virtual GPU Manager are logged to the hypervisor’s log files, prefixed with vmiop.

9.2.4.1. Examining Citrix Hypervisor vGPU Manager Messages

For Citrix Hypervisor, NVIDIA Virtual GPU Manager messages are written to /var/log/messages.

Look in the /var/log/messages file for the vmiop prefix:

Copy
Copied!
            

[root@xenserver ~]# grep vmiop /var/log/messages Jul 15 10:34:03 localhost vgpu-ll[25698]: notice: vmiop_log: gpu-pci-id : 0000:05:00.0 Jul 15 10:34:03 localhost vgpu-ll[25698]: notice: vmiop_log: vgpu_type : quadro Jul 15 10:34:03 localhost vgpu-ll[25698]: notice: vmiop_log: Framebuffer: 0x74000000 Jul 15 10:34:03 localhost vgpu-ll[25698]: notice: vmiop_log: Virtual Device Id: 0xl3F2:0xll4E Jul 15 10:34:03 localhost vgpu-ll[25698]: notice: vmiop_log: ######## vGPU Manager Information: ######## Jul 15 10:34:03 localhost vgpu-ll[25698]: notice: vmiop_log: Driver Version: 470.256.02 Jul 15 10:34:03 localhost vgpu-ll[25698]: notice: vmiop_log: Init frame copy engine: syncing... Jul 15 10:35:31 localhost vgpu-ll[25698]: notice: vmiop_log: ######## Guest NVIDIA Driver Information: ######## Jul 15 10:35:31 localhost vgpu-ll[25698]: notice: vmiop_log: Driver Version: 475.14 Jul 15 10:35:36 localhost vgpu-ll[25698]: notice: vmiop_log: Current max guest pfn = 0xllbc84! Jul 15 10:35:40 localhost vgpu-ll[25698]: notice: vmiop_log: Current max guest pfn = 0xlleff0! [root@xenserver ~]#

9.2.4.2. Examining Red Hat Enterprise Linux KVM vGPU Manager Messages

For Red Hat Enterprise Linux KVM and RHV, NVIDIA Virtual GPU Manager messages are written to /var/log/messages.

Look in these files for the vmiop_log: prefix:

Copy
Copied!
            

# grep vmiop_log: /var/log/messages [2024-07-12 04:46:12] vmiop_log: [2024-07-12 04:46:12] notice: vmiop-env: guest_max_gpfn:0x11f7ff [2024-07-12 04:46:12] vmiop_log: [2024-07-12 04:46:12] notice: pluginconfig: /usr/share/nvidia/vgx/grid_m60-1q.conf,gpu-pci-id=0000:06:00.0 [2024-07-12 04:46:12] vmiop_log: [2024-07-12 04:46:12] notice: Loading Plugin0: libnvidia-vgpu [2024-07-12 04:46:12] vmiop_log: [2024-07-12 04:46:12] notice: Successfully update the env symbols! [2024-07-12 04:46:12] vmiop_log: [2024-07-12 04:46:12] notice: vmiop_log: gpu-pci-id : 0000:06:00.0 [2024-07-12 04:46:12] vmiop_log: [2024-07-12 04:46:12] notice: vmiop_log: vgpu_type : quadro [2024-07-12 04:46:12] vmiop_log: [2024-07-12 04:46:12] notice: vmiop_log: Framebuffer: 0x38000000 [2024-07-12 04:46:12] vmiop_log: [2024-07-12 04:46:12] notice: vmiop_log: Virtual Device Id: 0x13F2:0x114D [2024-07-12 04:46:12] vmiop_log: [2024-07-12 04:46:12] notice: vmiop_log: ######## vGPU Manager Information: ######## [2024-07-12 04:46:12] vmiop_log: [2024-07-12 04:46:12] notice: vmiop_log: Driver Version: 470.256.02 [2024-07-12 04:46:12] vmiop_log: [2024-07-12 04:46:12] notice: vmiop_log: Init frame copy engine: syncing... [2024-07-12 05:09:14] vmiop_log: [2024-07-12 05:09:14] notice: vmiop_log: ######## Guest NVIDIA Driver Information: ######## [2024-07-12 05:09:14] vmiop_log: [2024-07-12 05:09:14] notice: vmiop_log: Driver Version: 475.14 [2024-07-12 05:09:14] vmiop_log: [2024-07-12 05:09:14] notice: vmiop_log: Current max guest pfn = 0x11a71f! [2024-07-12 05:12:09] vmiop_log: [2024-07-12 05:12:09] notice: vmiop_log: vGPU license state: (0x00000001) #

9.2.4.3. Examining VMware vSphere vGPU Manager Messages

For VMware vSphere, NVIDIA Virtual GPU Manager messages are written to the vmware.log file in the guest VM’s storage directory.

Look in the vmware.log file for the vmiop prefix:

Copy
Copied!
            

[root@esxi:~] grep vmiop /vmfs/volumes/datastore1/win7-vgpu-test1/vmware.log 2024-07-12T14:02:21.275Z| vmx| I120: DICT pciPassthru0.virtualDev = "vmiop" 2024-07-12T14:02:21.344Z| vmx| I120: GetPluginPath testing /usr/lib64/vmware/plugin/libvmx-vmiop.so 2024-07-12T14:02:21.344Z| vmx| I120: PluginLdr_LoadShared: Loaded shared plugin libvmx-vmiop.so from /usr/lib64/vmware/plugin/libvmx-vmiop.so 2024-07-12T14:02:21.344Z| vmx| I120: VMIOP: Loaded plugin libvmx-vmiop.so:VMIOP_InitModule 2024-07-12T14:02:21.359Z| vmx| I120: VMIOP: Initializing plugin vmiop-display 2024-07-12T14:02:21.365Z| vmx| I120: vmiop_log: gpu-pci-id : 0000:04:00.0 2024-07-12T14:02:21.365Z| vmx| I120: vmiop_log: vgpu_type : quadro 2024-07-12T14:02:21.365Z| vmx| I120: vmiop_log: Framebuffer: 0x74000000 2024-07-12T14:02:21.365Z| vmx| I120: vmiop_log: Virtual Device Id: 0x11B0:0x101B 2024-07-12T14:02:21.365Z| vmx| I120: vmiop_log: ######## vGPU Manager Information: ######## 2024-07-12T14:02:21.365Z| vmx| I120: vmiop_log: Driver Version: 470.256.02 2024-07-12T14:02:21.365Z| vmx| I120: vmiop_log: VGX Version: 13.12 2024-07-12T14:02:21.445Z| vmx| I120: vmiop_log: Init frame copy engine: syncing... 2024-07-12T14:02:37.031Z| vthread-12| I120: vmiop_log: ######## Guest NVIDIA Driver Information: ######## 2024-07-12T14:02:37.031Z| vthread-12| I120: vmiop_log: Driver Version: 475.14 2024-07-12T14:02:37.031Z| vthread-12| I120: vmiop_log: VGX Version: 13.12 2024-07-12T14:02:37.093Z| vthread-12| I120: vmiop_log: Clearing BAR1 mapping 2024-07-15T23:39:55.726Z| vmx| I120: VMIOP: Shutting down plugin vmiop-display [root@esxi:~]

9.3. Capturing configuration data for filing a bug report

When filing a bug report with NVIDIA, capture relevant configuration data from the platform exhibiting the bug in one of the following ways:

  • On any supported hypervisor, run nvidia-bug-report.sh.
  • On Citrix Citrix Hypervisor, create a Citrix Hypervisor server status report.

9.3.1. Capturing configuration data by running nvidia-bug-report.sh

The nvidia-bug-report.sh script captures debug information into a gzip-compressed log file on the server.

Run nvidia-bug-report.sh from theCitrix Hypervisor dom0 shell, the Red Hat Enterprise Linux KVM host shell, the Red Hat Virtualization (RHV) host shell, or the VMware ESXi host shell.

This example runs nvidia-bug-report.sh on Citrix Hypervisor, but the procedure is the same on Red Hat Enterprise Linux KVM, RHV, or VMware vSphere ESXi.

Copy
Copied!
            

[root@xenserver ~]# nvidia-bug-report.sh nvidia-bug-report.sh will now collect information about your system and create the file 'nvidia-bug-report.log.gz' in the current directory. It may take several seconds to run. In some cases, it may hang trying to capture data generated dynamically by the Linux kernel and/or the NVIDIA kernel module. While the bug report log file will be incomplete if this happens, it may still contain enough data to diagnose your problem. For Xen open source/XCP users, if you are reporting a domain issue, please run: nvidia-bug-report.sh --domain-name <"domain_name"> Please include the 'nvidia-bug-report.log.gz' log file when reporting your bug via the NVIDIA Linux forum (see devtalk.nvidia.com) or by sending email to 'linux-bugs@nvidia.com'. Running nvidia-bug-report.sh... If the bug report script hangs after this point consider running with --safe-mode command line argument. complete [root@xenserver ~]#

9.3.2. Capturing Configuration Data by Creating a Citrix Hypervisor Status Report

  1. In XenCenter, from the Tools menu, choose Server Status Report.
  2. Select the Citrix Hypervisor instance from which you want to collect a status report.
  3. Select the data to include in the report.
  4. To include NVIDIA vGPU debug information, select NVIDIA-logs in the Report Content Item list.
  5. Generate the report.

    Figure 31. Including NVIDIA logs in a Citrix Hypervisor status report

    nvidia-logs-in-citrix-xenserver-status-report.png

A.1. Virtual GPU Types for Supported GPUs

NVIDIA vGPU is available as a licensed product on supported NVIDIA GPUs. For a list of recommended server platforms and supported GPUs, consult the release notes for supported hypervisors at NVIDIA Virtual GPU Software Documentation.

A.1.1. NVIDIA A100 PCIe 40GB Virtual GPU Types

Physical GPUs per board: 1 This GPU supports MIG-backed virtual GPUs and time-sliced virtual GPUs.

MIG-Backed C-Series Virtual GPU Types for NVIDIA A100 PCIe 40GB

Required license edition: vCS or vWS

For details of GPU instance profiles, see NVIDIA Multi-Instance GPU User Guide.

Virtual GPU Type Intended Use Case Frame Buffer (MB) Maximum vGPUs per GPU Slices per vGPU Compute Instances per vGPU Corresponding GPU Instance Profile
A100-7-40C Training Workloads 40960 1 7 7 MIG 7g.40gb
A100-4-20C Training Workloads 20480 1 4 4 MIG 4g.20gb
A100-3-20C Training Workloads 20480 2 3 3 MIG 3g.20gb
A100-2-10C Training Workloads 10240 3 2 2 MIG 2g.10gb
A100-1-5C Inference Workloads 5120 7 1 1 MIG 1g.5gb
A100-1-5CME Inference Workloads 5120 1 1 1 MIG 1g.5gb+me


Time-Sliced C-Series Virtual GPU Types for NVIDIA A100 PCIe 40GB

Required license edition: vCS or vWS

These vGPU types support a single display with a fixed maximum resolution.

Virtual GPU Type Intended Use Case Frame Buffer (MB) Maximum vGPUs per GPU Maximum vGPUs per Board Maximum Display Resolution Virtual Displays per vGPU
A100-40C Training Workloads 40960 1 1 4096×21602 1
A100-20C Training Workloads 20480 2 2 4096×21602 1
A100-10C Training Workloads 10240 4 4 4096×21602 1
A100-8C Training Workloads 8192 5 5 4096×21602 1
A100-5C Inference Workloads 5120 8 8 4096×21602 1
A100-4C Inference Workloads 4096 10 10 4096×21602 1

A.1.2. NVIDIA A100 HGX 40GB Virtual GPU Types

Physical GPUs per board: 1 This GPU supports MIG-backed virtual GPUs and time-sliced virtual GPUs.

MIG-Backed C-Series Virtual GPU Types for NVIDIA A100 HGX 40GB

Required license edition: vCS or vWS

For details of GPU instance profiles, see NVIDIA Multi-Instance GPU User Guide.

Virtual GPU Type Intended Use Case Frame Buffer (MB) Maximum vGPUs per GPU Slices per vGPU Compute Instances per vGPU Corresponding GPU Instance Profile
A100X-7-40C Training Workloads 40960 1 7 7 MIG 7g.40gb
A100X-4-20C Training Workloads 20480 1 4 4 MIG 4g.20gb
A100X-3-20C Training Workloads 20480 2 3 3 MIG 3g.20gb
A100X-2-10C Training Workloads 10240 3 2 2 MIG 2g.10gb
A100X-1-5C Inference Workloads 5120 7 1 1 MIG 1g.5gb
A100X-1-5CME Inference Workloads 5120 1 1 1 MIG 1g.5gb+me


Time-Sliced C-Series Virtual GPU Types for NVIDIA A100 HGX 40GB

Required license edition: vCS or vWS

These vGPU types support a single display with a fixed maximum resolution.

Virtual GPU Type Intended Use Case Frame Buffer (MB) Maximum vGPUs per GPU Maximum vGPUs per Board Maximum Display Resolution Virtual Displays per vGPU
A100X-40C Training Workloads 40960 1 1 4096×21602 1
A100X-20C Training Workloads 20480 2 2 4096×21602 1
A100X-10C Training Workloads 10240 4 4 4096×21602 1
A100X-8C Training Workloads 8192 5 5 4096×21602 1
A100X-5C Inference Workloads 5120 8 8 4096×21602 1
A100X-4C Inference Workloads 4096 10 10 4096×21602 1

A.1.3. NVIDIA A100 PCIe 80GB Virtual GPU Types

Physical GPUs per board: 1 This GPU supports MIG-backed virtual GPUs and time-sliced virtual GPUs.

MIG-Backed C-Series Virtual GPU Types for NVIDIA A100 PCIe 80GB

Required license edition: vCS or vWS

For details of GPU instance profiles, see NVIDIA Multi-Instance GPU User Guide.

Virtual GPU Type Intended Use Case Frame Buffer (MB) Maximum vGPUs per GPU Slices per vGPU Compute Instances per vGPU Corresponding GPU Instance Profile
A100D-7-80C Training Workloads 81920 1 7 7 MIG 7g.80gb
A100D-4-40C Training Workloads 40960 1 4 4 MIG 4g.40gb
A100D-3-40C Training Workloads 40960 2 3 3 MIG 3g.40gb
A100D-2-20C Training Workloads 20480 3 2 2 MIG 2g.20gb
A100D-1-10C Training Workloads 10240 7 1 1 MIG 1g.10gb
A100D-1-10CME Training Workloads 10240 1 1 1 MIG 1g.10gb+me


Time-Sliced C-Series Virtual GPU Types for NVIDIA A100 PCIe 80GB

Required license edition: vCS or vWS

These vGPU types support a single display with a fixed maximum resolution.

Virtual GPU Type Intended Use Case Frame Buffer (MB) Maximum vGPUs per GPU Maximum vGPUs per Board Maximum Display Resolution Virtual Displays per vGPU
A100D-80C Training Workloads 81920 1 1 4096×21602 1
A100D-40C Training Workloads 40960 2 2 4096×21602 1
A100D-20C Training Workloads 20480 4 4 4096×21602 1
A100D-16C Inference Workloads 16384 5 5 4096×21602 1
A100D-10C Training Workloads 10240 8 8 4096×21602 1
A100D-8C Training Workloads 8192 10 10 4096×21602 1
A100D-4C Inference Workloads 4096 20 20 4096×21602 1

A.1.4. NVIDIA A100 HGX 80GB Virtual GPU Types

Physical GPUs per board: 1 This GPU supports MIG-backed virtual GPUs and time-sliced virtual GPUs.

MIG-Backed C-Series Virtual GPU Types for NVIDIA A100 HGX 80GB

Required license edition: vCS or vWS

For details of GPU instance profiles, see NVIDIA Multi-Instance GPU User Guide.

Virtual GPU Type Intended Use Case Frame Buffer (MB) Maximum vGPUs per GPU Slices per vGPU Compute Instances per vGPU Corresponding GPU Instance Profile
A100DX-7-80C Training Workloads 81920 1 7 7 MIG 7g.80gb
A100DX-4-40C Training Workloads 40960 1 4 4 MIG 4g.40gb
A100DX-3-40C Training Workloads 40960 2 3 3 MIG 3g.40gb
A100DX-2-20C Training Workloads 20480 3 2 2 MIG 2g.20gb
A100DX-1-10C Training Workloads 10240 7 1 1 MIG 1g.10gb
A100DX-1-10CME Training Workloads 10240 1 1 1 MIG 1g.10gb+me


Time-Sliced C-Series Virtual GPU Types for NVIDIA A100 HGX 80GB

Required license edition: vCS or vWS

These vGPU types support a single display with a fixed maximum resolution.

Virtual GPU Type Intended Use Case Frame Buffer (MB) Maximum vGPUs per GPU Maximum vGPUs per Board Maximum Display Resolution Virtual Displays per vGPU
A100DX-80C Training Workloads 81920 1 1 4096×21602 1
A100DX-40C Training Workloads 40960 2 2 4096×21602 1
A100DX-20C Training Workloads 20480 4 4 4096×21602 1
A100DX-16C Inference Workloads 16384 5 5 4096×21602 1
A100DX-10C Training Workloads 10240 8 8 4096×21602 1
A100DX-8C Training Workloads 8192 10 10 4096×21602 1
A100DX-4C Inference Workloads 4096 20 20 4096×21602 1

A.1.5. NVIDIA A40 Virtual GPU Types

Physical GPUs per board: 1

Q-Series Virtual GPU Types for NVIDIA A40

Required license edition: vWS

These vGPU types support a maximum combined resolution based on the number of available pixels, which is determined by their frame buffer size. You can choose between using a small number of high resolution displays or a larger number of lower resolution displays with these vGPU types. The maximum number of displays per vGPU is based on a configuration in which all displays have the same resolution. For examples of configurations with a mixture of display resolutions, see Mixed Display Configurations for B-Series and Q-Series vGPUs.

Virtual GPU Type Intended Use Case Frame Buffer (MB) Maximum vGPUs per GPU Maximum vGPUs per Board Available Pixels Display Resolution Virtual Displays per vGPU