NVIDIA AI Enterprise Solution Guide

This guide aims to provide guidance on how to set up a high-performance multi-node cluster as virtual machines. Within this guide, you will become familiar with GPUDirect RDMA and ATS while using Docker as the platform for running high-performance multi-node Deep Learning Training. ATS is a VMware PCIe support enhancement in vSphere 7 Update 2. GPUDirect RDMA benefits from ATS and is certified and supported by NVIDIA AI Enterprise.

NVIDIA AI Enterprise

• Enterprise-Grade AI Software Platform
  • Enterprise-Grade AI Software Platform From NVIDIA
  • NVIDIA AI Software Supported
  • NVIDIA AI Workflows
  • Production-Ready Pretrained Models
  • NVIDIA NIM
  • Supported Hardware and Software

Multi-Node Deep Learning Training

• Overview
• Compute Workflows
  • Multi-GPU Deep Learning Workflows
    • Single VM – Full or Fractionalized Partitioned vGPU
    • Single VM Using Multiple NVIDIA NVLink vGPU Devices
    • Multiple Nodes (VMs)
  • Remote Direct Memory Access (RDMA) With Address Translation Services (ATS)
  • NVIDIA Container Toolkit
• Requirements
  • Summary of Minimum Requirements
  • Server Configuration
  • VM Requirements
  • VM Configuration
  • Additional VM Configuration
  • Optional: Setting up the A100 vGPU VM with MIG
• Getting Started
  • Configure NVIDIA ConnectX-6 Dx NIC and Spectrum switch for RoCE
  • Enable ATS on VMware ESXi and VMs
  • Enable ATS on the NVIDIA ConnectX-6 Dx NIC
  • Configure NUMA Affinity for the VMs
    • Creating Docker File For Multi-Node Training
  • Setup Keyless Entry Between VMs On The Multi-Node Cluster
    • Generating SSH Keys
      • Host/WorkerSSH Keys
      • Container SSH Keys
      • Creating Config Files for Keyless Entry
      • Add Public SSH Keys To “Authorized_keys”
      • Create mpicont.sh script
        • Add Container SSH Key to the Master’s authorized_keys File
        • Copy ~/.ssh to Worker Nodes and Confirm Keyless entry
        • Change Permissions in the ssh_container on all Nodes
        • Install nv_peer_memory
  • Run Sample ResNet-50 Multi-Node Training
    • Interpreting the Results
• Summary

Support

• Support and Services

Notices

• Notices
  • Notice
  • Trademarks
  • Copyright