RDG for Apache Spark 3.0 on Kubernetes Accelerated with RAPIDS over RoCE Network

NVIDIA Docs Hub NVIDIA Networking Networking Solutions RDG for Apache Spark 3.0 on Kubernetes Accelerated with RAPIDS over RoCE Network

Introduction
Abbreviation and Acronym List
References
Key Components and Technologies
Solution Overview
Deployment
- Physical Network
  - Connectivity
  - Network Configuration
- Network Switch Configuration for RoCE Transport
  - NVIDIA Cumulus Linux Network OS
Deployment Guide
Prepare and Run TPC-H benchmarks
Appendix A
Appendix B
About the Authors

Created on Aug 16, 2020

Introduction

The following Reference Deployment Guide (RDG) demonstrates the process of running Apache Spark 3.0 workload with RAPIDS Accelerator for Apache Spark and 25Gb/s Ethernet RoCE.

The deployment will be provisioned on top of Network accelerated and GPU enabled Kubernetes cluster over NVIDIA Mellanox end-to-end 25 Gb/s Ethernet solution.

In this document we will go through the following processes:

How to deploy K8s cluster with Kubespray over bare metal nodes running Ubuntu 18.04.
How to prepare the network for RoCE traffic using NVIDIA recommended settings on both host and switch sides.
How to deploy and run RAPIDS accelerated Apache Spark 3.0 cluster over NVIDIA accelerated infrastructure.

Abbreviation and Acronym List

Term	Definition	Term	Definition
AOC	Active Optical Cable	NFD	Node Feature Discovery
CNI	Container Network Interface	NGC	NVIDIA GPU Cloud
DAC	Direct Attach Copper cable	PF	Physical Function
DHCP	Dynamic Host Configuration Protocol	RDG	Reference Deployment Guide
DNS	Domain Name System	RDMA	Remote Direct Memory Access
DP	Device Plugin	RoCE	RDMA over Converged Ethernet
GDR	GPUDirect	SR-IOV	Single Root Input Output Virtualization
GPU	G raphics Processing Unit	GPU	G raphics Processing Unit
HWE	Hardware Enablement	UCX	Unified Communication X
K8s	Kubernetes	VF	Virtual Function
MPI	Message Passing Interface	VLAN	Virtual Local Area Network

Key Components and Technologies

NVIDIA ^® T4 GPU
The NVIDIA ^® T4 GPU is based on the NVIDIA Turing ^™ architecture and packaged in an energy-efficient 70-watt small PCIe form factor. T4 is optimized for mainstream computing environments, and features multi-precision Turing Tensor Cores and RT Cores. Combined with accelerated containerized software stacks from NGC, T4 delivers revolutionary performance at scale to accelerate cloud workloads, such as high-performance computing, deep learning training and inference, machine learning, data analytics, and graphics..
NVIDIA Cumulus Linux
Cumulus Linux is the only open network OS that enables building affordable network fabrics and operating them similarly to the world’s largest data center operators, unlocking web-scale networking for businesses of all sizes.
NVIDIA Mellanox ConnectX-5 Ethernet Network Interface Cards
NVIDIA Mellanox ConnectX-5 NICs enable the highest performance and efficiency for data centers hyper-scale, public and private clouds, storage, Machine Learning, Deep Learning, Artificial Intelligence, Big Data and Telco platforms and applications

NVIDIA Mellanox Spectrum® Open Ethernet Switches
The NVIDIA Mellanox Spectrum® switch family provides the most efficient network solution for the ever-increasing performance demands of Data Center applications. The Spectrum product family includes a broad portfolio of Top-of-Rack (TOR) and aggregation switches that range from 16 to 128 physical ports, with Ethernet data rates of 1GbE, 10GbE, 25GbE, 40GbE, 50GbE, 100GbE and 200GbE per port. Spectrum Ethernet switches are ideal to build cost-effective and scalable data center network fabrics that can scale from a few nodes to tens-of-thousands of nodes.

NVIDIA Mellanox LinkX® Ethernet Cables and Transceivers
NVIDIA Mellanox LinkX cables and transceivers make 100Gb/s deployments as easy and as universal as 10Gb/s links. Because Mellanox offers one of industry’s broadest portfolio of 10, 25, 40, 50,100 and 200Gb/s Direct Attach Copper cables (DACs), Copper Splitter cables, Active Optical Cables (AOCs) and Transceivers, every data center reach from 0.5m to 10km is supported. To maximize system performance. Mellanox tests every product in an end-to-end environment ensuring a Bit Error Rate of less than 1e-15. A BER of 1e-15 is 1000x better than many competitors.

Kubernetes
Kubernetes (K8s) is an open-source container orchestration platform for deployment automation, scaling, and management of containerized applications.
Kubespray ( From Kubernetes.io, )
Kubespray is a composition of Ansible playbooks, inventory, provisioning tools, and domain knowledge for generic OS/Kubernetes clusters configuration management tasks and provides:
- A highly available cluster
- Composable attributes
- Support for most popular Linux distributions
NVIDIA GPU Operator
The NVIDIA GPU Operator uses the operator framework within Kubernetes to automate the management of all NVIDIA software components needed to provision GPU. These components include the NVIDIA drivers (to enable CUDA), Kubernetes device plugin for GPUs, the NVIDIA Container Runtime, automatic node labelling, DCGM based monitoring and others.
Multus
Multus is a meta CNI plugin that provides multiple network interface support to pods. For each interface Multus delegates CNI calls to secondary CNI plugins such as Calico, SR-IOV, etc.
SR-IOV Network Device Plugin
Kubernetes device plugin for discovering and advertising SR-IOV virtual functions (VFs) available on a Kubernetes host. The plugin was enhanced by the NVIDIA Mellanox R&D team to use RDMA applications in Kubernetes.
RDMA
RDMA supports zero-copy networking by enabling the network adapter to transfer data from the wire directly to application memory or from application memory directly to the wire, eliminating the need to copy data between application memory and the data buffers in the operating system. Such transfers require no work to be done by CPUs, caches, or context switches, and transfers continue in parallel with other system operations. This reduces latency in message transfer
SR-IOV CNI Plugin
The SR-IOV CNI plugin works with SR-IOV network device plugin for VF allocation in Kubernetes. It enables the configuration and usage of SR-IOV VF networks in containers and orchestration systems like Kubernetes.
Apache Spark™
Apache Spark™ is an open-source, fast and general engine for large-scale data processing. Spark provides an interface for programming entire clusters with implicit data parallelism and fault-tolerance.
RAPIDS
The RAPIDS suite of open source software libraries and APIs provide the ability to execute end-to-end data science and analytics pipelines entirely on GPUs. Licensed under Apache 2.0, RAPIDS is incubated by NVIDIA® based on extensive hardware and data science experience. RAPIDS utilizes NVIDIA CUDA® primitives for low-level compute optimization, and exposes GPU parallelism and high-bandwidth memory speed through user-friendly Python interfaces.
RAPIDS also focuses on common data preparation tasks for analytics and data science. This includes a familiar dataframe API that integrates with a variety of machine learning algorithms for end-to-end pipeline accelerations without paying typical serialization costs. RAPIDS also includes support for multi-node, multi-GPU deployments, enabling vastly accelerated processing and training on much larger dataset sizes.
RAPIDS Accelerator for Apache Spark
The RAPIDS Accelerator for Apache Spark combines the power of the RAPIDS cuDF library and the scale of the Spark distributed computing framework.
The RAPIDS Accelerator library also has a built-in accelerated shuffle based on UCX that can be configured to leverage GPU-to-GPU communication and RDMA capabilities.
Documentation on the current release can be found at here.

Unified Communication X

Unified Communication X (UCX) provides an optimized communication layer for Message Passing (MPI), PGAS/OpenSHMEM libraries and RPC/data-centric applications.

UCX utilizes high-speed networks for extra-node communication, and shared memory mechanisms for efficient intra-node communication.

GPUDirect RDMA
GPUDirect (GDR) RDMA provides a direct P2P (Peer-to-Peer) data path between the GPU Memory directly to and from NVIDIA Mellanox HCA devices, which reduces GPU-to-GPU communication latency and completely offloads the CPU, removing it from all GPU-to-GPU communications across the network.

Solution Overview

Prerequisites

Hardware
All servers K8s Worker nodes have the same hardware specification (see BoM for details).
Host OS
Ubuntu Server 18.04 operating system installed on all servers with OpenSSH server packages.
Switch OS
NVIDIA Cumulus Linux 4.1.1
Management Network
DHCP and DNS services are part of the IT infrastructure. The components Installation and configuration are not covered in this guide.
Expirience with Apache Spark
Make sure to familiarize with the Apache Cluster multi-node cluster architecture. See Overview - Spark 3.0 Documentation for more info.

There are multiple ways to deploy Spark and Spark Rapids Plugin. Standalone Mode, Local Mode (for development and testing only, not production), on a YARN cluster (see this link for more details), and on a Kubernetes cluster which is the method used in this deployment guide.

The logical design includes the following layers:

Two separate networking layers:
1. Management network
2. High-speed RoCE Network
One Compute layer:
1. K8s Master node and Deployment
2. 3 x K8s Worker Nodes with two NVIDIA T4 GPUs and one Mellanox ConnectX-5 adapter.

Bill of Materials (BoM)

This document covers single Kubernetes controller deployment scenario. For highly available cluster deployment refer to https://github.com/kubernetes-sigs/kubespray/blob/master/docs/ha-mode.md

The following hardware setup is utilized in the distributed Spark/K8s configuration described in this guide:

Warning

The above table does not contain Kubernetes Management network connectivity components.

Deployment

Physical Network

Connectivity

The first port of each NVIDIA Mellanox network card on each Worker Node is wired to NVIDIA Mellanox switch using 25Gb/s DAC cables:

Warning

This table does not contain Kubernetes Management network connectivity components.

Network Configuration

Below are the server names with their relevant network configurations:

Server/Switch type	Server/Switch name	IP and NICS
Server/Switch type	Server/Switch name	High-speed network 25 GigE (VLAN -111)	Management network 1 GigE
Master Node	node1		eno0: DHCP
Worker Node	node2	ens1f0: none	eno0: DHCP
Worker Node	node3	ens1f0: none	eno0: DHCP
Worker Node	node4	ens1f0: none	eno0: DHCP
Depl./Driver Node	depl-node		eno0: DHCP
High-speed switch	swx-mld-s01	none	mgmt0: From DHCP

Note

ens1f0 interfaces do not require any additional configuration.

Network Switch Configuration for RoCE Transport

NVIDIA Cumulus Linux Network OS

RoCE transport is utilized to accelerate Spark networking through the UCX library. To get the highest possible results we will configure our network to be lossless.

Run the following commands to configure a lossless networks and for NVIDIA Cumulus version 4.1.1 and above :

Switch console

Copy
Copied!

            
            net add interface swp1-32 storage-optimized pfc
net commit

Add VLAN 111 to ports 1-3 on NVIDIA Cumulus Linux Network OS by running the following commands:

Switch console

Copy
Copied!

            
            net add interface swp1-6 bridge trunk 111
net add interface swp1-6 bridge trunk vlans 111
net commit
net show bridge vlan

Deployment Guide

Nodes Configuration

General Prerequisites:

Ubuntu 18.04 system.
Access to a terminal or command line.
Sudo user or root permissions.

Host OS Prerequisites

Make sure Ubuntu Server 18.04 operating system is installed on all servers with OpenSSH server packages and create a non-root user account with sudo privileges without password.

Update the Ubuntu software packages and install the latest HWE kernel by running the below commands:

Server console

Copy
Copied!

            
            # apt-get update
# apt-get -y install linux-image-generic-hwe-18.04
# reboot
# sudo apt-get upgrade -y

Non-root User Account Prerequisites

In this solution we added the following line to the EOF /etc/sudoers :

Server Console

Copy
Copied!

            
            #includedir /etc/sudoers.d
 
 
#K8s cluster deployment user with sudo privileges without password
user ALL=(ALL) NOPASSWD:ALL

Installation Process

Install general dependencies on the deployment server, run the commands below or paste each line into the terminal:
Server Console

Copy

Copied!
```
            
            > sudo apt-get install git wget scala maven make gcc openssh-server openssh-client -y
        
```

Install the Java 8 software packages:

Server Console

Copy
Copied!

            
            > sudo apt-get install python-software-properties
> sudo add-apt-repository ppa:webupd8team/java
> sudo apt-get update
> sudo apt-get install oracle-java8-installer

Install the general dependencies on the Worker Servers by running the commands below or paste each line into the terminal:
Server Console

Copy

Copied!
```
            
            > sudo apt-get install git wget dkms make gcc
        
```
To install LLDP service on the Worker Servers, run the commands below or paste each line into the terminal:
Server Console

Copy

Copied!
```
            
            # sudo service lldpd start
# sudo systemctl enable lldpd
        
```
To create a Network File System (NFS) Share, Install NFS server on a new or existing server in the environment. For this solution Node 2 (First Worker Node) is used as the NFS server. Follow the procedure in detailed in this guide to install the BFS server.

SR-IOV Configuration (on Worker Nodes only)

Verify that you are using SR-IOV supported server platform and review the BIOS settings in the hardware documentation to enable support for SR-IOV networking.

Enable Virtualization (SR-IOV) in the BIOS.

Enable SR-IOV in the NIC firmware by execute the following commands:

Copy
Copied!

            
            sudo apt-get install mstflint
### request information about ALL Mellanox NIC's on the server ###
sudo mstconfig q
### enable SR-IOV with 8 VF's
mstconfig -d /sys/bus/pci/devices/0000:13:00.0/config set SRIOV_EN=1 NUM_OF_VFS=8
reboot

Warning

All Worker Nodes must have the same configuration and the same PCIe card placement.

K8s Cluster Deployment and Configuration

The Kubernetes cluster in this solution will be installed using Kubespray with a non-root user account from a Deployment Node(K8s Master Node).

Configuring SSH Private Key and SSH Passwordless Login

Login to the Deployment Node as a deployment user (in this case - user) and create SSH private key for configuring the password-less authentication on your computer by running the following command:

Copy
Copied!

            
            $ ssh-keygen

Copy your SSH private key , such as ~/.ssh/id_rsa , to all nodes in your deployment by running the following command:

Copy
Copied!

            
            $ ssh-copy-id -i <filename> user@nodename

Configuring Kubespray

Copy
Copied!

            
            $ cd ~
$ sudo apt -y install python3-pip jq
$ wget https://github.com/kubernetes-sigs/kubespray/archive/v2.13.3.tar.gz
$ tar -zxf v2.13.3.tar.gz
$ cd kubespray-2.13.3
$ sudo pip3 install -r requirements.txt

The default folder for subsequent commands is ~/kubespray-2.13.3.

Create a new cluster configuration:

Copy
Copied!

            
            $ cp -rfp inventory/sample inventory/mycluster
$ declare -a IPS=(192.168.1.6 192.168.1.71 192.168.1.72 192.168.1.73)
$ CONFIG_FILE=inventory/mycluster/hosts.yaml python3 contrib/inventory_builder/inventory.py ${IPS[@]}

Review and change the host configuration file - inventory/mycluster/hosts.yaml. Below is an example output from this solution:

inventory/mycluster/hosts.yaml

Copy
Copied!

            
            all:
  hosts:
    node1:
      ansible_host: 192.168.1.6
      ip: 192.168.1.6
      access_ip: 192.168.1.6
    node2:
      ansible_host: 192.168.1.71
      ip: 192.168.1.71
      access_ip: 192.168.1.71
    node3:
      ansible_host: 192.168.1.72
      ip: 192.168.1.72
      access_ip: 192.168.1.72
    node4:
      ansible_host: 192.168.1.73
      ip: 192.168.1.73
      access_ip: 192.168.1.73
  children:
    kube-master:
      hosts:
        node1:
    kube-node:
      hosts:
        node2:
        node3:
        node4:
    etcd:
      hosts:
        node1:
    k8s-cluster:
      children:
        kube-master:
        kube-node:
    calico-rr:
      hosts: {}

Customizing variables for K8s cluster installation

Review and change cluster installation parameters under inventory/mycluster/group_vars:

Copy
Copied!

            
            $ cat inventory/mycluster/group_vars/all/all.yml
$ cat inventory/mycluster/group_vars/k8s-cluster/k8s-cluster.yml

In “inventory/mycluster/group_vars/all.yml” uncomment the following line so that the metrics can receive data about cluster resources use.

Copy
Copied!

            
            vim inventory/mycluster/group_vars/all/all.yml
 
## The read-only port for the Kubelet to serve on with no authentication/authorization. Uncomment to enable.
kube_read_only_port: 10255

In "inventory/mycluster/group_vars/k8s-cluster/k8s-cluster.yml" enable Multus installation by setting the variable kube_network_plugin_multus to true, and specify the Docker version by adding the variable " docker_version: 19.03 " t o avoid any related issues .

Copy
Copied!

            
            vim inventory/mycluster/group_vars/k8s-cluster/k8s-cluster.yml
 
## Setting multi_networking to true will install Multus: https://github.com/intel/multus-cni
kube_network_plugin_multus: true
 
## Container runtime
## docker for docker, crio for cri-o and containerd for containerd.
container_manager: docker
docker_version: 19.03

Note

The Kubespray version in this deployment suffers from an inconsistency when installing Docker components (see issue details here: https://github.com/kubernetes-sigs/kubespray/issues/6160).

Note

The default Kubernetes CNI can be changed by setting the desired kube_network_plugin value (default: calico) parameter in inventory/mycluster/group_vars/k8s-cluster/k8s-cluster.yml.

It will install Multus and Calico and configure Multus to use Calico as the primary network plugin.

Deploy K8s cluster with Kubespray Ansible Playbook

Deployment Node Console

Copy
Copied!

            
            $ ansible-playbook -i inventory/mycluster/hosts.yaml --become --become-user=root cluster.yml

Note

The execution time for this step may take a while to finalize.

Example of a successful completion of the playbooks looks like:

Deployment Node Console

Copy
Copied!

            
            PLAY RECAP ***************************************************************************************************************************************
localhost                  : ok=1    changed=0    unreachable=0    failed=0   
node1                      : ok=617  changed=101  unreachable=0    failed=0   
node2                      : ok=453  changed=58   unreachable=0    failed=0   
node3                      : ok=410  changed=53   unreachable=0    failed=0   
node4                      : ok=410  changed=53   unreachable=0    failed=0   
 
 
Monday 16 April 2020  17:48:14 +0300 (0:00:00.265)       0:13:49.321 ********** 
=============================================================================== 
kubernetes/master : kubeadm | Initialize first master ------------------------------------------------------------------------------------ 55.94s
kubernetes/kubeadm : Join to cluster ----------------------------------------------------------------------------------------------------- 37.65s
kubernetes/master : Master | wait for kube-scheduler ------------------------------------------------------------------------------------- 21.97s
download : download_container | Download image if required ------------------------------------------------------------------------------- 21.34s
kubernetes-apps/ansible : Kubernetes Apps | Start Resources ------------------------------------------------------------------------------ 14.85s
kubernetes/preinstall : Update package management cache (APT) ---------------------------------------------------------------------------- 12.49s
download : download_file | Download item ------------------------------------------------------------------------------------------------- 11.45s
etcd : Install | Copy etcdctl binary from docker container ------------------------------------------------------------------------------- 10.57s
download : download_file | Download item -------------------------------------------------------------------------------------------------- 9.37s
kubernetes/preinstall : Install packages requirements ------------------------------------------------------------------------------------- 9.18s
etcd : wait for etcd up ------------------------------------------------------------------------------------------------------------------- 8.78s
etcd : Configure | Check if etcd cluster is healthy --------------------------------------------------------------------------------------- 8.62s
download : download_file | Download item -------------------------------------------------------------------------------------------------- 8.24s
kubernetes-apps/network_plugin/multus : Multus | Start resources -------------------------------------------------------------------------- 7.32s
download : download_container | Download image if required -------------------------------------------------------------------------------- 6.61s
policy_controller/calico : Start of Calico kube controllers ------------------------------------------------------------------------------- 4.92s
download : download_file | Download item -------------------------------------------------------------------------------------------------- 4.76s
kubernetes-apps/cluster_roles : Apply workaround to allow all nodes with cert O=system:nodes to register ---------------------------------- 4.56s
download : download_container | Download image if required -------------------------------------------------------------------------------- 4.48s
download : download | Download files / images --------------------------------------------------------------------------------------------- 4.28s

K8s Deployment Verification

The Kubernetes cluster deployment verification must be done from the K8s Master Node.

Copy Kubernetes cluster configuration files from ROOT folder to user home folder or run using root user on the K8s Master Node.

Execute the following command for copy Kubernetes cluster configuration files to a non-root user on the Master Node:

K8s Master Node Console

Copy
Copied!

            
            user@node1:$ sudo su -
root@node1:~# cp -r .kube/ /home/user/
root@node1:~# chown -R `id -u user`:`id -g user` /home/user/.kube/
root@node1:~# exit

Verify that the Kubernetes cluster is installed properly. Execute the following commands:

K8s Master Node Console

Copy
Copied!

            
            user@node1:$ kubectl get nodes -o wide
NAME    STATUS   ROLES    AGE   VERSION   INTERNAL-IP    EXTERNAL-IP   OS-IMAGE             KERNEL-VERSION     CONTAINER-RUNTIME
node1   Ready    master   14d   v1.17.9   192.168.1.6    <none>        Ubuntu 18.04.4 LTS   5.4.0-42-generic   docker://18.9.7
node2   Ready    <none>   14d   v1.17.9   192.168.1.71   <none>        Ubuntu 18.04.4 LTS   5.4.0-42-generic   docker://18.9.7
node3   Ready    <none>   14d   v1.17.9   192.168.1.72   <none>        Ubuntu 18.04.4 LTS   5.4.0-42-generic   docker://18.9.7
node4   Ready    <none>   14d   v1.17.9   192.168.1.73   <none>        Ubuntu 18.04.4 LTS   5.4.0-42-generic   docker://18.9.7
 
user@node1:~$ kubectl get pod -n kube-system -o wide
NAME                                          READY   STATUS    RESTARTS   AGE   IP             NODE    NOMINATED NODE   READINESS GATES
calico-kube-controllers-8567d5fdd6-zfch2      1/1     Running   0          12d   192.168.1.71   node2   <none>           <none>
calico-node-fzv78                             1/1     Running   1          14d   192.168.1.6    node1   <none>           <none>
calico-node-gsplr                             1/1     Running   2          14d   192.168.1.71   node2   <none>           <none>
calico-node-n6mgq                             1/1     Running   4          14d   192.168.1.73   node4   <none>           <none>
calico-node-pdcft                             1/1     Running   5          14d   192.168.1.72   node3   <none>           <none>
coredns-76798d84dd-2tt2f                      1/1     Running   0          12d   10.233.92.60   node3   <none>           <none>
coredns-76798d84dd-7sndq                      1/1     Running   0          14d   10.233.90.1    node1   <none>           <none>
dns-autoscaler-85f898cd5c-5ldrz               1/1     Running   0          14d   10.233.90.2    node1   <none>           <none>
kube-apiserver-node1                          1/1     Running   0          14d   192.168.1.6    node1   <none>           <none>
kube-controller-manager-node1                 1/1     Running   0          14d   192.168.1.6    node1   <none>           <none>
kube-multus-ds-amd64-7s445                    1/1     Running   1          14d   192.168.1.72   node3   <none>           <none>
kube-multus-ds-amd64-8g7br                    1/1     Running   1          14d   192.168.1.71   node2   <none>           <none>
kube-multus-ds-amd64-dncpc                    1/1     Running   2          14d   192.168.1.73   node4   <none>           <none>
kube-multus-ds-amd64-h2n76                    1/1     Running   0          14d   192.168.1.6    node1   <none>           <none>
kube-proxy-f7zgz                              1/1     Running   1          14d   192.168.1.71   node2   <none>           <none>
kube-proxy-ml4s4                              1/1     Running   3          14d   192.168.1.73   node4   <none>           <none>
kube-proxy-mlk7c                              1/1     Running   4          14d   192.168.1.72   node3   <none>           <none>
kube-proxy-pqc8m                              1/1     Running   0          14d   192.168.1.6    node1   <none>           <none>
kube-scheduler-node1                          1/1     Running   0          14d   192.168.1.6    node1   <none>           <none>
kubernetes-dashboard-77475cf576-rsl5h         1/1     Running   0          12d   10.233.92.62   node3   <none>           <none>
kubernetes-metrics-scraper-747b4fd5cd-tjwl2   1/1     Running   0          12d   10.233.96.51   node2   <none>           <none>
nginx-proxy-node2                             1/1     Running   1          14d   192.168.1.71   node2   <none>           <none>
nginx-proxy-node3                             1/1     Running   4          14d   192.168.1.72   node3   <none>           <none>
nginx-proxy-node4                             1/1     Running   3          14d   192.168.1.73   node4   <none>           <none>
nodelocaldns-bpqvr                            1/1     Running   3          14d   192.168.1.73   node4   <none>           <none>
nodelocaldns-f85lh                            1/1     Running   3          14d   192.168.1.72   node3   <none>           <none>
nodelocaldns-jsknr                            1/1     Running   0          14d   192.168.1.6    node1   <none>           <none>
nodelocaldns-t9dts                            1/1     Running   1          14d   192.168.1.71   node2   <none>           <none>

NVIDIA GPU Operator Installation for K8s cluster

The preferred method to deploy the device plugin is as a daemonset using helm. Install Helm from the official installer script:

K8s Master Node Console

Copy
Copied!

            
            $ curl -fsSL -o get_helm.sh https://raw.githubusercontent.com/helm/helm/master/scripts/get-helm-3
$ chmod 700 get_helm.sh
$ ./get_helm.sh

Add the NVIDIA repository:

K8s Master Node Console

Copy
Copied!

            
            $ helm repo add nvidia https://nvidia.github.io/gpu-operator 
$ helm repo update

NVIDIA GPU Operator uses hostNetwork by default. The defaults must be modified as it is not suitable for this solution.
Deploy the device plugin:

K8s Master Node Console

Copy
Copied!

            
            # helm install --wait --generate-name nvidia/gpu-operator

K8s Master Node Console

Copy
Copied!

            
            # helm ls
NAME                    NAMESPACE       REVISION        UPDATED                                 STATUS          CHART                   APP VERSION
gpu-operator-1596563035 default         1               2020-10-30 20:43:59.550042604 +0300 IDT deployed        gpu-operator-1.3.0      1.3.0

K8s Master Node Console

Copy
Copied!

            
            # kubectl get pods -A NAMESPACE                NAME default                  gpu-operator-1596563035-no default                  gpu-operator-1596563035-no default                  gpu-operator-1596563035-no default                  gpu-operator-1596563035-no default                  gpu-operator-1596563035-no default                  gpu-operator-74c97448d9-pmwjh gpu-operator-resources   nvidia-container-toolkit-d gpu-operator-resources   nvidia-container-toolkit-d gpu-operator-resources   nvidia-container-toolkit-d gpu-operator-resources   nvidia-dcgm-exporter-dlzjv gpu-operator-resources   nvidia-dcgm-exporter-hbd4t gpu-operator-resources   nvidia-dcgm-exporter-hhjcb gpu-operator-resources   nvidia-device-plugin-daemonset-jnprj gpu-operator-resources   nvidia-device-plugin-daemonset-rcj9n gpu-operator-resources   nvidia-device-plugin-daemonset-slkpj gpu-operator-resources   nvidia-device-plugin-validation gpu-operator-resources   nvidia-driver-daemonset-8b92r gpu-operator-resources   nvidia-driver-daemonset-lkwdk gpu-operator-resources   nvidia-driver-daemonset-mcdf2 gpu-operator-resources   nvidia-driver-validation kube-system              calico-kube-controllers-85 kube-system              calico-node-fzv78 kube-system              calico-node-gsplr kube-system              calico-node-n6mgq kube-system              calico-node-pdcft kube-system              coredns-76798d84dd-2tt2f kube-system              coredns-76798d84dd-7sndq kube-system              dns-autoscaler-85f898cd5c-5ldrz kube-system              kube-apiserver-node1 kube-system              kube-controller-manager-node1 kube-system              kube-multus-ds-amd64-7s445 kube-system              kube-multus-ds-amd64-8g7br kube-system              kube-multus-ds-amd64-dncpc kube-system              kube-multus-ds-amd64-h2n76 kube-system              kube-proxy-f7zgz kube-system              kube-proxy-ml4s4 kube-system              kube-proxy-mlk7c kube-system              kube-proxy-pqc8m kube-system              kube-scheduler-node1 kube-system              kubernetes-dashboard-77475cf576-rsl5h kube-system              kubernetes-metrics-scraper kube-system              nginx-proxy-node2 kube-system              nginx-proxy-node3 kube-system              nginx-proxy-node4 kube-system              nodelocaldns-bpqvr kube-system              nodelocaldns-f85lh kube-system              nodelocaldns-jsknr kube-system              nodelocaldns-t9dts

READY   STATUS      RESTARTS   AGE de-feature-discovery-master-6468bznxc   1/1     Running     0          15d de-feature-discovery-worker-jdz5f       1/1     Running     33         15d de-feature-discovery-worker-lwqks       1/1     Running     13         15d de-feature-discovery-worker-sb7sv       1/1     Running     16         15d de-feature-discovery-worker-xv5xb       1/1     Running     8          15d 1/1     Running     1          15d aemonset-hlgp9                          1/1     Running     3          15d aemonset-wgxnx                          1/1     Running     2          15d aemonset-zxqxj                          1/1     Running     1          15d 1/1     Running     8          15d 1/1     Running     6          15d 1/1     Running     13         15d 1/1     Running     13         15d 1/1     Running     8          15d 1/1     Running     0          13d 0/1     Completed   0          13d 1/1     Running     1          15d 1/1     Running     8          15d 1/1     Running     18         15d 0/1     Completed   0          13d 67d5fdd6-zfch2                          1/1     Running     0          13d 1/1     Running     1          15d 1/1     Running     2          15d 1/1     Running     4          15d 1/1     Running     5          15d 1/1     Running     0          13d 1/1     Running     0          15d 1/1     Running     0          15d 1/1     Running     0          15d 1/1     Running     0          15d 1/1     Running     1          15d 1/1     Running     1          15d 1/1     Running     2          15d 1/1     Running     0          15d 1/1     Running     1          15d 1/1     Running     3          15d 1/1     Running     4          15d 1/1     Running     0          15d 1/1     Running     0          15d 1/1     Running     0          13d -747b4fd5cd-tjwl2                       1/1     Running     0          13d 1/1     Running     1          15d 1/1     Running     4          15d 1/1     Running     3          15d 1/1     Running     3          15d 1/1     Running     3          15d 1/1     Running     0          15d 1/1     Running     1          15d

To run a Sample GPU Application: https://github.com/NVIDIA/gpu-operator#running-a-sample-gpu-application For GPU monitoring: https://github.com/NVIDIA/gpu-operator#gpu-monitoring

SR-IOV Components Installation for K8s Cluster

The RoCE enhancement for the K8s cluster enables containerizing SR-IOV virtual functions in Pod deployments for additional RoCE enabled NIC's.
During the installation process, a role will use the Kubespray inventory/mycluster/hosts.yaml file for Kubernetes components deployment and provisioning.

The RoCE components are configured by a separate role in the installation. The RoCE role will install the following components:

Virtual Function activation
Node Feature Discovery v0.6.0
The latest Multus CNI for attaching multiple network interfaces to the Pod
Specific configuration of the universal SR-IOV device plugin
Universal SR-IOV CNI
Specific network provisioning with "NetworkAttachmentDefinition"
DHCP CNI for providing IP addresses for SR-IOV based NIC's in the Pod deployment from existing infrastructure

Important

RoCE role deployment is validated using Ubuntu 18.04 OS and Kubespray v2.13.3

Prerequisites

Copy the RoCE role installation package(provided in Appendix) to Kubespray folder:

K8s Master Node Console

Copy
Copied!

            
            $ cd ~/kubespray-2.13.3
$ tar -xf roce.tar

Customize Role Variables

Set the variables for the RoCE role in the yml file - roles/roce_sriov/vars/main.yml.

Set the following parameters:

sriov_resources:
- pf_name: "ens3f0"
vendor: "15b3"
dev_id: "1018"
vlan_id: 111
resourcePrefix: "mellanox.com"
res_name: "sriov_111"
network_name: "sriov111"
mtu: 1500
cidr: "192.168.111.0/24"
num_vf to 8
hugePages to false
install_mofed to true

vars/main.yml

Copy
Copied!

            
            ---
# vars file for roce_sriov
# Physical adapter names must be connected to RoCE backend fabric
# Virtual function device ID. Default - "MT28908 Family [ConnectX-6 Virtual Function]"
# Detailed information about all Mellanox Device ID can be found - https://devicehunt.com/view/type/pci/vendor/15B3.
# Supported values 
#    101c - MT28908 Family [ConnectX-6 Virtual Function]
#    101a - MT28800 Family [ConnectX-5 Ex Virtual Function]
#    1018 - MT27800 Family [ConnectX-5 Virtual Function]   
#    1016 - MT27710 Family [ConnectX-4 Lx Virtual Function]
#    1014 - MT27700 Family [ConnectX-4 Virtual Function]
sriov_resources:
  - pf_name: "ens3f0"
    vendor: "15b3"
    dev_id: "1018"
    vlan_id: 111
    resourcePrefix: "mellanox.com"
    res_name: "sriov_111"
    network_name: "sriov111"
    mtu: 1500
    cidr: "192.168.111.0/24"

# CNI spec version
cniVersion: "0.4.0"

# Number virtual function for activation
num_vf: 8

# HugePages
# Activated only on Worker Nodes with MLNX nic's
# Supported only HugePages mode - hugepages_2048KB
# num_huge parameter set hugepages size for each NUMA. For Node with two NUMA's - 16384

hugePages: false
num_huge: 8192

# DHCP server settings

# If set to TRUE in your network must be installed DHCP server for provide IP address assignment for corresponded VLAN's 
# and will be used DHCP plugin for IPAM

# If set to FALSE will be installed https://github.com/openshift/whereabouts-cni as IPAM plugin
# cidr parametr from sriov_resources used for IP address assignment
# Supported ONLY with K8s v1.16 and above

dhcp_server: false


# Using for Host OS the Ubuntu LTS enablement (also called HWE or Hardware Enablement) kernel
HWE_kernel: true

# New MOFED installation
# If install_mofed is FALSE, will be used kernel inbox driver

install_mofed: false
mlnx_ofed_package: "mlnx-ofed-kernel-only"
mlnx_ofed_version: "latest"
upstream_libs: true


# Install Kubeflow MPI-Operator - https://github.com/kubeflow/mpi-operator

kubeflow_mpi_operator: false
mpi_operator_dep: "https://raw.githubusercontent.com/kubeflow/mpi-operator/master/deploy/v1/mpi-operator.yaml"

# K8s SR-IOV daemonset's
# before K8s 1.16
#multus_ds: "https://raw.githubusercontent.com/intel/multus-cni/master/images/multus-daemonset-pre-1.16.yml"
#sriov_dp_ds: "https://raw.githubusercontent.com/intel/sriov-network-device-plugin/master/deployments/k8s-v1.10-v1.15/sriovdp-daemonset.yaml"
#sriov_cni_ds: "https://raw.githubusercontent.com/intel/sriov-cni/master/images/k8s-v1.10-v1.15/sriov-cni-daemonset.yaml"
#dhcp_cni_ds: "https://raw.githubusercontent.com/Mellanox/dhcp-cni/master/dhcp-cni-ds.yaml"

# from K8s 1.16
multus_ds: "https://raw.githubusercontent.com/intel/multus-cni/master/images/multus-daemonset.yml"
sriov_dp_ds: "https://raw.githubusercontent.com/intel/sriov-network-device-plugin/master/deployments/k8s-v1.16/sriovdp-daemonset.yaml"
sriov_cni_ds: "https://raw.githubusercontent.com/intel/sriov-cni/master/images/k8s-v1.16/sriov-cni-daemonset.yaml"
dhcp_cni_ds: "https://raw.githubusercontent.com/Mellanox/dhcp-cni/master/k8s-1.16/dhcp-cni-ds.yaml"
nfd_release: "https://github.com/kubernetes-sigs/node-feature-discovery/archive/v0.6.0.tar.gz"
wh_cni_ds: "https://raw.githubusercontent.com/openshift/whereabouts-cni/master/doc/daemonset-install.yaml"
wh_iptools_ds: "https://raw.githubusercontent.com/openshift/whereabouts-cni/master/doc/whereabouts.cni.cncf.io_ippools.yaml"

Role Execution

Run the playbook from the Kubespray deployment folder using the following command:

Deployment Node Console

Copy
Copied!

            
            $ ansible-playbook -i inventory/mycluster/hosts.yaml  --become --become-user=root roce.yaml

Note

The execution time for this step may take a while to finalize.

The following is an example of a successful playbooks execution:

Deployment Node Console

Copy
Copied!

            
            PLAY RECAP ***************************************************************************************************************************************
node1                      : ok=47   changed=24   unreachable=0    failed=0    skipped=18   rescued=0    ignored=0   
node2                      : ok=32   changed=13   unreachable=0    failed=0    skipped=7    rescued=0    ignored=0   
node3                      : ok=32   changed=13   unreachable=0    failed=0    skipped=7    rescued=0    ignored=0   
node4                      : ok=32   changed=13   unreachable=0    failed=0    skipped=7    rescued=0    ignored=0   
 
Sunday 26 July 2020  15:24:42 +0300 (0:00:00.668)       0:01:07.966 *********** 
=============================================================================== 
roce_sriov : Update additional packages --------------------------------------------------------------------------------------------------- 8.54s
roce_sriov : Set Hugepages ---------------------------------------------------------------------------------------------------------------- 4.92s
roce_sriov : Set Hugepages ---------------------------------------------------------------------------------------------------------------- 4.89s
roce_sriov : Install Universal SR-IOV device plugin --------------------------------------------------------------------------------------- 4.47s
roce_sriov : Install WH CNI --------------------------------------------------------------------------------------------------------------- 3.00s
roce_sriov : Update cache ----------------------------------------------------------------------------------------------------------------- 2.60s
roce_sriov : Extract NFD daemonset's ------------------------------------------------------------------------------------------------------ 2.43s
roce_sriov : Create netattdef ------------------------------------------------------------------------------------------------------------- 2.28s
Gathering Facts --------------------------------------------------------------------------------------------------------------------------- 2.24s
roce_sriov : Install WH IPPOOL CNI -------------------------------------------------------------------------------------------------------- 2.20s
roce_sriov : Install Openshift pip module ------------------------------------------------------------------------------------------------- 2.09s
roce_sriov : Install SR-IOV CNI ----------------------------------------------------------------------------------------------------------- 1.99s
roce_sriov : Remove old Multus DS for amd64 ----------------------------------------------------------------------------------------------- 1.87s
roce_sriov : Remove DHCP CNI -------------------------------------------------------------------------------------------------------------- 1.77s
roce_sriov : Create netattdef ------------------------------------------------------------------------------------------------------------- 1.33s
roce_sriov : Multus update ---------------------------------------------------------------------------------------------------------------- 1.28s
roce_sriov : Install HWE kernel. It takes a while. ---------------------------------------------------------------------------------------- 1.27s
roce_sriov : create configmap ------------------------------------------------------------------------------------------------------------- 1.25s
roce_sriov : Install aptitude ------------------------------------------------------------------------------------------------------------- 1.24s
roce_sriov : Remove Multus DS for ppc64le ------------------------------------------------------------------------------------------------- 0.67s

Role Installation Summary:

After the installation, the role with default variable parameters in your K8s cluster will have the following:

Node Feature Discovery for Kubernetes installed
The required amount of VF for each specified Mellanox NIC name activated and configured
configmap for "SRIOV NETWORK DEVICE PLUGIN" will be configured and ready for creating resources
DaemonSet's with "SRIOV NETWORK DEVICE PLUGIN" and "SRIOV CNI" installed
Multus meta CNI updated to the latest version
DaemonSet with "whereabouts-cni" installed to provide IP address management for SRIOV based NIC's
One network-attachment-definitions SRIOV111

Role Deployment Verification

The role deployment verification must be done from the K8s Master node. Execute the following commands to initiate the verification process:

K8s Master Node Console

Copy
Copied!

            
            root@node1:~# kubectl get pod -n kube-system -o wide | egrep "multus|cni|device|whereabouts"
kube-multus-ds-amd64-7s445                    1/1     Running   1          15d   192.168.1.72   node3   <none>           <none>
kube-multus-ds-amd64-8g7br                    1/1     Running   1          15d   192.168.1.71   node2   <none>           <none>
kube-multus-ds-amd64-dncpc                    1/1     Running   2          15d   192.168.1.73   node4   <none>           <none>
kube-multus-ds-amd64-h2n76                    1/1     Running   0          15d   192.168.1.6    node1   <none>           <none>
kube-sriov-cni-ds-amd64-g7rgp                 1/1     Running   1          15d   192.168.1.72   node3   <none>           <none>
kube-sriov-cni-ds-amd64-vzl9s                 1/1     Running   1          15d   192.168.1.71   node2   <none>           <none>
kube-sriov-cni-ds-amd64-zdvmh                 1/1     Running   2          15d   192.168.1.73   node4   <none>           <none>
kube-sriov-device-plugin-amd64-6qpwr          1/1     Running   1          15d   192.168.1.71   node2   <none>           <none>
kube-sriov-device-plugin-amd64-8lvdt          1/1     Running   1          15d   192.168.1.72   node3   <none>           <none>
kube-sriov-device-plugin-amd64-cjhjx          1/1     Running   2          15d   192.168.1.73   node4   <none>           <none>
whereabouts-9f4r5                             1/1     Running   0          15d   192.168.1.6    node1   <none>           <none>
whereabouts-dbxzz                             1/1     Running   1          15d   192.168.1.72   node3   <none>           <none>
whereabouts-fcsxr                             1/1     Running   2          15d   192.168.1.73   node4   <none>           <none>
whereabouts-qd8xm                             1/1     Running   1          15d   192.168.1.71   node2   <none>           <none>

Worker Node resources

Copy
Copied!

            
            root@node1:~# kubectl describe nodes node2
...
Capacity:
  cpu:                     32
  ephemeral-storage:       229700940Ki
  hugepages-1Gi:           0
  hugepages-2Mi:           0
  mellanox.com/sriov_111:  8
  memory:                  197746780Ki
  nvidia.com/gpu:          2
  pods:                    110
Allocatable:
  cpu:                     31900m
  ephemeral-storage:       211692385954
  hugepages-1Gi:           0
  hugepages-2Mi:           0
  mellanox.com/sriov_111:  8
  memory:                  197394380Ki
  nvidia.com/gpu:          2
  pods:                    110
...

network-attachment-definitions

Copy
Copied!

            
            root@node1:~# kubectl get network-attachment-definitions.k8s.cni.cncf.io
NAME       AGE
sriov111   10m
 
user@node1:~# kubectl get network-attachment-definitions.k8s.cni.cncf.io sriov111 -o yaml
apiVersion: k8s.cni.cncf.io/v1
kind: NetworkAttachmentDefinition
metadata:
  annotations:
    k8s.v1.cni.cncf.io/resourceName: mellanox.com/sriov_111
  creationTimestamp: "2020-08-04T18:17:32Z"
  generation: 1
  name: sriov111
  namespace: default
  resourceVersion: "9404"
  selfLink: /apis/k8s.cni.cncf.io/v1/namespaces/default/network-attachment-definitions/sriov111
  uid: 07a5e65b-f42f-41fd-b8e9-59c196e77056
spec:
  config: |-
    {
        "cniVersion": "0.4.0",
        "name": "sriov111",
        "plugins": [
            {
                "ipam": {
                    "datastore": "kubernetes",
                    "kubernetes": {
                        "kubeconfig": "/etc/cni/net.d/whereabouts.d/whereabouts.kubeconfig"
                    },
                    "log_file": "/tmp/whereabouts.log",
                    "log_level": "debug",
                    "range": "192.168.111.0/24",
                    "type": "whereabouts"
                },
                "spoofChk": "off",
                "type": "sriov",
                "vlan": 111
            },
            {
                "mtu": 1500,
                "type": "tuning"
            }
        ]
    }

Lossless Fabric with L3 (DSCP) Configuration

Warning

Before starting the below process, make sure you are familiar with this configuration example for NVIDIA Mellanox devices installed with MLNX_OFED running RoCE over a lossless network in DSCP-based QoS mode.

For this configuration, make sure to know your network interface name (for example ens3f0) and its parent NVIDIA Mellanox device (for example mlx5_0). To get this information, run the ibdev2netdev command:

Shell

Copy
Copied!

            
            # ibdev2netdev -v | grep ens3f0
mlx5_0 port 1 ==> ens3f0 (Up)
mlx5_1 port 1 ==> ens3f1 (Down)

Configuration:

Shell

Copy
Copied!

            
            # mlnx_qos -i ens3f0 --trust dscp
# echo 106 > /sys/class/infiniband/mlx5_0/tc/1/traffic_class
# cma_roce_tos -d mlx5_0 -t 106
# sysctl -w net.ipv4.tcp_ecn=1
# mlnx_qos -i ens3f0 --pfc 0,0,0,1,0,0,0,0

Installing Mellanox GPUDirect RDMA

The below listed software is required to install and run the GPUDirect RDMA:

NVIDIA compatible driver. (Contact NVIDIA support for more information)
MLNX_OFED (latest).

Copy the NVIDIA driver from /run/nvidia/driver/usr/src/nvidia-440.64.00/kernel/nvidia.ko file to the /lib/modules/5.4.0-42-generic/updates/dkms/ directory.
Server Console

Copy

Copied!
```
            
            # cp /run/nvidia/driver/usr/src/nvidia-440.64.00/kernel/nvidia.ko /lib/modules/5.4.0-42-generic/updates/dkms/
        
```

Download the latest nv_peer_memory:

Server Console

Copy
Copied!

            
            # cd
# git clone https://github.com/Mellanox/nv_peer_memory

Build source packages (src.rpm for RPM based OS and tarball for DEB based OS) and run the build_module.sh script:

Node cli

Copy
Copied!

            
            # ./build_module.sh
 
 
Building source rpm for nvidia_peer_memory...
Building debian tarball for nvidia-peer-memory...
 
 
Built: /tmp/nvidia_peer_memory-1.0-9.src.rpm
Built: /tmp/nvidia-peer-memory_1.0.orig.tar.gz

Install the packages:

Node cli

Copy
Copied!

            
            # cd /tmp 
# tar xzf /tmp/nvidia-peer-memory_1.0.orig.tar.gz
# cd nvidia-peer-memory-1.0
# dpkg-buildpackage -us -uc
# dpkg -i <path to generated deb files>

( e.g. dpkg -i nv-peer-memory_1.0-9_all.deb nv-peer-memory-dkms_1.0-9_all.deb )

After a successful installation:

nv_peer_mem.ko kernel module will be installed.
service file /etc/init.d/nv_peer_mem to control the kernel module by start/stop/status will be added.
/etc/infiniband/nv_peer_mem.conf configuration file to control whether kernel module will be loaded on boot (default value is YES).

Warning

We recommend both the NIC and the GPU to be physically placed on the same PCI switch to achieve better performance.

Installing OFED Performance Tests

Perftools is a collection of tests written over uverbs intended for use as a performance micro-benchmark. The tests may be used for HW or SW tuning as well as for functional testing. The collection contains a set of bandwidth and latency benchmarks such as:

Send - ib_send_bw and ib_send_lat
RDMA Read - ib_read_bw and ib_read_lat|
RDMA Write - ib_write_bw and ib_write_lat
RDMA Atomic - ib_atomic_bw and ib_atomic_lat
Native Ethernet (when working with MOFED2) - raw_ethernet_bw, raw_ethernet_lat.

To install perftools, run the following commands:

Server Console

Copy
Copied!

            
            cd
git clone https://github.com/linux-rdma/perftest
cd perftest/
make clean
./autogen.sh
./configure --prefix=/usr --libdir=/usr/lib64 --sysconfdir=/etc CUDA_H_PATH=/usr/local/cuda/include/cuda.h
make

Run a RDMA Write - ib_write_bw bandwidth stress benchmark over RoCE.

Server	*ib_write_bw -a -d mlx5_0 &*
Client	*ib_write_bw -a -F $server_IP -d mlx5_0 --report_gbits*

Distributed Spark Deployment

Installing Apache Spark Driver Node

Download Apache Spark (Deployment Node), from Downloads | Apache Spark by selecting a Spark release and package type spark-3.0.0-bin-hadoop3.2.tgz file and copy the file to the /opt folder.

Note

Optional: You can download Spark by running following commands:
cd /opt ; wget http://mirror.cogentco.com/pub/apache/spark/spark-3.0.0/spark-3.0.0-bin-hadoop3.2.tgz
To set up Apache Spark, extract the compressed file using the tar command:
Server Console

Copy

Copied!
```
            
            tar -xvf spark-3.0.0-bin-hadoop3.2.tgz
        
```
Make a symbolic link for spark-3.0.0-bin-hadoop3.2 folder and s et the permissions for your user to spark-3.0.0-bin-hadoop3.2 and spark folders :
Server Console

Copy

Copied!
```
            
            sudo ln -s spark-3.0.0-bin-hadoop3.2/ spark
sudo chown -R user spark
sudo chown -R user spark-3.0.0-bin-hadoop3.2/
        
```
To Configure Spark environment variables (bashrc), edit the .bashrc shell configuration file:
Server Console

Copy

Copied!
```
            
            sudo vim .bashrc
        
```

Define the Spark environment variables by adding the following content to the end of the file:

Server Console

Copy
Copied!

            
            #Spark Related Options
echo "export SPARK_HOME=/opt/spark"
echo "export PATH=$PATH:$SPARK_HOME/bin:$SPARK_HOME/sbin"
echo "export PYSPARK_PYTHON=/usr/bin/python3"

Once you save and close the text file, you can return to your original terminal and type this command to reload your .bashrc file:
Server Console

Copy

Copied!
```
            
            $ source ~/.bashrc
        
```

Check that the path have been properly modified:

Server Console

Copy
Copied!

            
            echo $SPARK_HOME
echo $PATH
echo $PYSPARK_PYTHON

Run a sample Spark job in local mode to ensure Spark is working:

Server Console

Copy
Copied!

            
            $SPARK_HOME/bin/run-example SparkPi 10

Install the Spark Rapids plugin jars, create /opt/sparkRapidsPlugin on the Deployment Node and s et the permissions for your user to the directory :
Server Console

Copy

Copied!
```
            
            sudo mkdir -p /opt/sparkRapidsPlugin
sudo chown -R kuser sparkRapidsPlugin
        
```
Download the 2 jars from the https://github.com/NVIDIA/spark-rapids/blob/branch-0.2/docs/version/stable-release.md .
Place them into the /opt/sparkRapidsPlugin directory. Make sure to pull the cudf jar version for the cuda version you are running 10.2.

Prepare K8s Cluster to run Apache Spark Executor

Prerequisites

Run following steps on the Deployment Node:

To run an Apache Spark and use NFS as a storage to execute benchmarks over a K8s cluster we need to enable the IPC_LOCK capability for memory registration and the nfs-volume configuration.

Create /opt/executor-template directory:

Deployment Node Console

Copy
Copied!

            
            # mkdir /opt/executor-template

Define an executor-template by c reating /opt/executor-template/executor-template.yaml file which contains the following configurations.

Warning

It is important to change the NFS server IP to an IP relevant to your environment.

Deployment Node Console

Copy
Copied!

            
            #
# Licensed to the Apache Software Foundation (ASF) under one or more
# contributor license agreements.  See the NOTICE file distributed with
# this work for additional information regarding copyright ownership.
# The ASF licenses this file to You under the Apache License, Version 2.0
# (the "License"); you may not use this file except in compliance with
# the License.  You may obtain a copy of the License at
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
apiVersion: v1
Kind: Pod
metadata:
  labels:
    template-label-key: executor-template-label-value
spec:
  containers:
  - name: test-executor-container
    image: will-be-overwritten
    securityContext:
      capabilities:
        add: [ "IPC_LOCK" ]
    volumeMounts:
        - name: nfs-volume
          mountPath: /data
  volumes:
    - name: nfs-volume
      nfs:
        # URL for the NFS server
        server: 192.168.1.71
        path: /data

Install and run Docker service on the Deployment Node.

Info

You can skip steps 2-6 and use my docker image from DockerHub.
Deployment Node Console

Copy

Copied!
```
            
            # apt install docker.io
# service docker start
        
```
Create /opt/spark/docker directory on the Deployment Node:
Deployment Node Console

Copy

Copied!
```
            
            # mkdir /opt/spark/docker
        
```

On the Deployment Node, create the /opt/spark/docker/getSRIOVResources.sh file to get the resource information about NVIDIA Mellanox SR-IOV device:

Deployment Node Console

Copy
Copied!

            
            #!/usr/bin/env bash

#
# Licensed to the Apache Software Foundation (ASF) under one or more
# contributor license agreements. See the NOTICE file distributed with
# this work for additional information regarding copyright ownership.
# The ASF licenses this file to You under the Apache License, Version 2.0
# (the "License"); you may not use this file except in compliance with
# the License. You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#

# This script is a basic example script to get resource information about Mellanox SR-IOV devices.
# It assumes the InfiniBand drivers and K8s SR-IOV components are properly installed.
# It is not guaranteed to work on all setups so please test and customize as needed
# for your environment. It can be passed into SPARK via the config
# spark.{driver/executor}.resource.RES-NAME.discoveryScript to allow the driver or executor to discover
# the SR-IOV device it was allocated. It assumes you are running within an isolated container where the
# SR-IOV device are allocated exclusively to that driver or executor.
# It outputs a JSON formatted string that is expected by the
# spark.{driver/executor}.resource.sriov_111.discoveryScript config.
#
# Execution: ./getSRIOVResources.sh
# Example output: {"name": "sriov_111", "addresses":["0000:05:02.4"]}


addr_sriov="$(env | grep -i sriov_111 | awk -F "=" '{print $2}')"
echo {\"name\": \"sriov_111\", \"addresses\":[\"${addr_sriov}\"]}

Create a Dockerfile with the following:

Deployment Node Console

Copy
Copied!

            
            #
# Licensed to the Apache Software Foundation (ASF) under one or more
# contributor license agreements. See the NOTICE file distributed with
# this work for additional information regarding copyright ownership.
# The ASF licenses this file to You under the Apache License, Version 2.0
# (the "License"); you may not use this file except in compliance with
# the License. You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
 
ARG CUDA_VER=10.2
ARG spark_uid=185
 
FROM nvidia/cuda:${CUDA_VER}-devel-ubuntu18.04
 
# Set MOFED version, OS version and platform
ENV MOFED_VER 5.0-2.1.8.0
ENV OS_VER ubuntu18.04
ENV PLATFORM x86_64
ENV DEBIAN_FRONTEND=noninteractive
 
# Set Unified Communication X {UCX) version, OS, MOFED and CUDA versions
ENV UCX_VER v1.8.1
ENV OS_VER ubuntu18.04
ENV MOFED mofed5.0
ENV CUDA 10.2
 
# Set Spark, Hadoop, cuDF and RAPIDS versions
ENV SPARK_VER 3.0.0
ENV HADOOP_VER 3.2
ENV CUDF_VER 0.14
ENV RAPIDS_VER 0.1.0
ENV RAPIDS 4
ENV SPARK_RAPIDS spark_2.12
ENV CUDA_RAPIDS cuda10-2
 
# Install dependencies
RUN apt-get update && \
apt install -y git wget apt-utils scala libnuma1 udev libudev1 libcap2 dpatch libnl-3-200 gfortran automake lsof ethtool chrpath libmnl0 pkg-config m4 libnl-route-3-200 autoconf debhelper swig bison libltdl-dev kmod tcl libnl-route-3-dev pciutils tk autotools-dev flex libnl-3-dev graphviz libgfortran4 iproute2 iputils-ping
RUN ln -fs /usr/share/zoneinfo/America/New_York /etc/localtime
RUN dpkg-reconfigure --frontend noninteractive tzdata
 
# Install java dependencies
RUN apt-get install -y --no-install-recommends openjdk-8-jdk openjdk-8-jre
ENV JAVA_HOME /usr/lib/jvm/java-1.8.0-openjdk-amd64
ENV PATH $PATH:/usr/lib/jvm/java-1.8.0-openjdk-amd64/jre/bin:/usr/lib/jvm/java-1.8.0-openjdk-amd64/bin
 
RUN mkdir -p /opt/spark/python
# TODO: Investigate running both pip and pip3 via virtualenvs
RUN apt install -y python python-pip && \
apt install -y python3 python3-pip && \
# Install numpy and pandas for xgboost python api tests
pip install --upgrade pip && \
pip3 install --upgrade pip && \
pip install numpy pandas && \
pip3 install numpy pandas && \
# We remove ensurepip since it adds no functionality since pip is
# installed on the image and it just takes up 1.6MB on the image
rm -r /usr/lib/python*/ensurepip && \
pip install --upgrade pip setuptools && \
# You may install with python3 packages by using pip3.6
# Removed the .cache to save space
rm -r /root/.cache && rm -rf /var/cache/apt/*
 
# MOFED install
ENV OFED_FQN MLNX_OFED_LINUX-${MOFED_VER}-${OS_VER}-${PLATFORM}
RUN wget --quiet http://content.mellanox.com/ofed/MLNX_OFED-${MOFED_VER}/${OFED_FQN}.tgz && \
tar -xf ${OFED_FQN}.tgz && \
${OFED_FQN}/mlnxofedinstall --user-space-only --without-fw-update -q
RUN cd .. && \
rm -rf ${MOFED_DIR} && \
rm -rf *.tgz
 
# UCX install
RUN wget https://github.com/openucx/ucx/releases/download/${UCX_VER}/ucx-${UCX_VER}-${OS_VER}-${MOFED}-cuda${CUDA}.deb && \
dpkg -i ucx-${UCX_VER}-${OS_VER}-${MOFED}-cuda${CUDA}.deb
 
# Before building the docker image, first build and make a Spark distribution following
# the instructions in http://spark.apache.org/docs/latest/building-spark.html.
# If this docker file is being used in the context of building your images from a Spark
# distribution, the docker build command should be invoked from the top level directory
# of the Spark distribution. E.g.:
# docker build -t spark:latest -f kubernetes/dockerfiles/spark/Dockerfile .
 
RUN set -ex && \
ln -s /lib /lib64 && \
mkdir -p /opt/spark && \
mkdir -p /opt/spark/jars && \
mkdir -p /opt/tpch && \
mkdir -p /opt/spark/examples && \
mkdir -p /opt/spark/work-dir && \
mkdir -p /opt/sparkRapidsPlugin && \
touch /opt/spark/RELEASE && \
rm /bin/sh && \
ln -sv /bin/bash /bin/sh && \
echo "auth required pam_wheel.so use_uid" >> /etc/pam.d/su && \
chgrp root /etc/passwd && chmod ug+rw /etc/passwd
 
COPY spark-${SPARK_VER}-bin-hadoop${HADOOP_VER}/jars /opt/spark/jars
COPY spark-${SPARK_VER}-bin-hadoop${HADOOP_VER}/bin /opt/spark/bin
COPY spark-${SPARK_VER}-bin-hadoop${HADOOP_VER}/sbin /opt/spark/sbin
COPY spark-${SPARK_VER}-bin-hadoop${HADOOP_VER}/kubernetes/dockerfiles/spark/entrypoint.sh /opt/
COPY spark-${SPARK_VER}-bin-hadoop${HADOOP_VER}/examples /opt/spark/examples
COPY spark-${SPARK_VER}-bin-hadoop${HADOOP_VER}/kubernetes/tests /opt/spark/tests
COPY spark-${SPARK_VER}-bin-hadoop${HADOOP_VER}/data /opt/spark/data
 
# Download RAPIDS Spark, cuDF Packages and a get GPU resources script
RUN cd /opt/sparkRapidsPlugin && \
wget https://repo1.maven.org/maven2/com/nvidia/rapids-${RAPIDS}-${SPARK_RAPIDS}/${RAPIDS_VER}/rapids-${RAPIDS}-${SPARK_RAPIDS}-${RAPIDS_VER}.jar && \
wget https://repo1.maven.org/maven2/ai/rapids/cudf/${CUDF_VER}/cudf-${CUDF_VER}-${CUDA_RAPIDS}.jar && \
wget https://raw.githubusercontent.com/apache/spark/master/examples/src/main/scripts/getGpusResources.sh
 
COPY getSRIOVResources.sh /opt/sparkRapidsPlugin/getSRIOVResources.sh
 
COPY spark-${SPARK_VER}-bin-hadoop${HADOOP_VER}/python/pyspark /opt/spark/python/pyspark
COPY spark-${SPARK_VER}-bin-hadoop${HADOOP_VER}/python/lib /opt/spark/python/lib
 
ENV SPARK_HOME /opt/spark
WORKDIR /opt/spark/work-dir
RUN chmod g+w /opt/spark/work-dir
 
ENV TINI_VERSION v0.18.0
ADD https://github.com/krallin/tini/releases/download/${TINI_VERSION}/tini /usr/bin/tini
RUN chmod +rx /usr/bin/tini
RUN chmod -R 777 /opt/sparkRapidsPlugin
 
ENTRYPOINT [ "/opt/entrypoint.sh" ]
 
# Specify the User that the actual main process will run as
USER ${spark_uid}

Build a docker image from the Dockerfile and push to the Docker Hub registry.

Deployment Node Console

Copy
Copied!

            
            # docker login <your DockerHub credentials>
# docker build -t <DockerHub repository>/<Docker image name> . 
# docker push <DockerHub repository>/<Docker image name>

Pull the Docker image from the Docker Hub registry on all Workers Nodes .
Worker Node Console

Copy

Copied!
```
            
            # docker pull <DockerHub repository>/<Docker image name>
        
```
You can use my precompiled docker image: bkovalev/spark3rapids

Prepare and Run TPC-H benchmarks

Prepare TPC-H dataset

To prepare the TPCH application, download the generated dataset and run the following commands on the Worker Node or on a server with access to NFS share:

Deployment Node Console

Copy
Copied!

            
            # mkdir /data/tpc
# mkdir /data/db
# sudo git clone https://github.com/databricks/tpch-dbgen.git
# cd tpch-dbgen
# sudo make
# sudo dbgen -s 70
# sudo hdfs dfs -put *.tbl
# cp *.tbl /data/db/
# ll /data/db/

Examples to run TPC-H application

To run TPC-H application with RAPIDS you need the following jar file:

Put the jar file into the /opt/sparkRapidsPlugin directory on the Deployment Node together with Spark Rapids Plugin jar files.

Here you can find examples to run TPC-H application:

Example 1:

For TCP without GPU/RAPIDS/UCX running we are using the tpch-tcp-only.sh file with the following configs when running Spark on K8s.

Deployment Node Console

Copy
Copied!

            
            ./spark-submit \
       --master k8s://https://192.168.1.6:6443 \
       --driver-memory 1G \
       --conf spark.executor.instances=6 \
       --conf spark.executor.memory=30G \
       --conf spark.executor.cores=8 \
       --conf spark.task.cpus=4 \
       --conf spark.sql.files.maxPartitionBytes=512m \
       --conf spark.sql.shuffle.partitions=300 \
       --conf spark.locality.wait=0s \
       --conf spark.kubernetes.executor.podTemplateFile=/opt/executor-template/executor-template.yaml \
       --conf spark.executor.resource.sriov_111.discoveryScript=/opt/sparkRapidsPlugin/getSRIOVResources.sh \
       --conf spark.executor.resource.sriov_111.amount=1 \
       --conf spark.executor.resource.sriov_111.vendor=mellanox.com \
       --conf spark.kubernetes.executor.annotation.k8s.v1.cni.cncf.io/networks=sriov111 \
       --conf spark.kubernetes.container.image=bkovalev/spark3rapids \
       --class com.nvidia.spark.rapids.tests.tpch.CSV /opt/sparkRapidsPlugin/rapids-4-spark-integration-tests_2.12-0.1-SNAPSHOT.jar file:///data/db file:///data/result 22

Example 2:

For TCP with GPU/RAPIDS and without UCX running we are using the tpch-rapids-tcp-noucx.sh file with the following configs when running Spark on K8s.

Deployment Node Console

Copy
Copied!

            
            ./spark-submit \
       --master k8s://https://192.168.1.6:6443 \
       --driver-memory 1G \
       --conf spark.executor.instances=6 \
       --conf spark.executor.memory=30G \
       --conf spark.executor.cores=8 \
       --conf spark.task.cpus=4 \
       --conf spark.rapids.memory.pinnedPool.size=2G \
       --conf spark.task.resource.gpu.amount=0.25 \
       --conf spark.rapids.sql.concurrentGpuTasks=1 \
       --conf spark.sql.files.maxPartitionBytes=512m \
       --conf spark.sql.shuffle.partitions=300 \
       --conf spark.locality.wait=0s \
       --conf spark.driver.extraClassPath=/opt/sparkRapidsPlugin/* \
       --conf spark.executor.extraClassPath=/opt/sparkRapidsPlugin/*:/usr/lib/:/data/jar/* \
       --conf spark.plugins=com.nvidia.spark.SQLPlugin \
       --conf spark.executor.resource.gpu.discoveryScript=/opt/sparkRapidsPlugin/getGpusResources.sh \
       --conf spark.executor.resource.gpu.amount=1 \
       --conf spark.executor.resource.gpu.vendor=nvidia.com \
       --conf spark.kubernetes.executor.podTemplateFile=/opt/executor-template/executor-template.yaml \
       --conf spark.executor.resource.sriov_111.discoveryScript=/opt/sparkRapidsPlugin/getSRIOVResources.sh \
       --conf spark.executor.resource.sriov_111.amount=1 \
       --conf spark.executor.resource.sriov_111.vendor=mellanox.com \
       --conf spark.kubernetes.executor.annotation.k8s.v1.cni.cncf.io/networks=sriov111 \
       --conf spark.executorEnv.UCX_TLS=cuda_copy,cuda_ipc,tcp \
       --conf spark.executorEnv.UCX_NET_DEVICES=net1 \
       --conf spark.executorEnv.LD_LIBRARY_PATH=/usr/local/cuda/lib64 \
       --conf spark.executorEnv.CUDA_HOME=/usr/local/cuda \     
       --conf spark.kubernetes.container.image=bkovalev/spark3rapids \
       --class com.nvidia.spark.rapids.tests.tpch.CSV /opt/sparkRapidsPlugin/rapids-4-spark-integration-tests_2.12-0.1-SNAPSHOT.jar file:///data/db file:///data/result 22

Example 3:

For TCP with GPU/RAPIDS/UCX running we are using the tpch-rapids-tcp-ucx.sh file with the following configs when running Spark on K8s.

Deployment Node Console

Copy
Copied!

            
            ./spark-submit \
       --master k8s://https://192.168.1.6:6443 \
       --driver-memory 1G \
       --conf spark.executor.instances=6 \
       --conf spark.executor.memory=30G \
       --conf spark.executor.cores=8 \
       --conf spark.task.cpus=4 \
       --conf spark.rapids.memory.pinnedPool.size=2G \
       --conf spark.task.resource.gpu.amount=0.25 \
       --conf spark.rapids.sql.concurrentGpuTasks=1 \
       --conf spark.sql.files.maxPartitionBytes=512m \
       --conf spark.sql.shuffle.partitions=300 \
       --conf spark.locality.wait=0s \
       --conf spark.driver.extraClassPath=/opt/sparkRapidsPlugin/* \
       --conf spark.executor.extraClassPath=/opt/sparkRapidsPlugin/*:/usr/lib/:/data/jar/* \
       --conf spark.plugins=com.nvidia.spark.SQLPlugin \
       --conf spark.shuffle.manager=com.nvidia.spark.RapidsShuffleManager \
       --conf spark.rapids.shuffle.transport.enabled=true \
       --conf spark.executor.resource.gpu.discoveryScript=/opt/sparkRapidsPlugin/getGpusResources.sh \
       --conf spark.executor.resource.gpu.amount=1 \
       --conf spark.executor.resource.gpu.vendor=nvidia.com \
       --conf spark.kubernetes.executor.podTemplateFile=/opt/executor-template/executor-template.yaml \
       --conf spark.executor.resource.sriov_111.discoveryScript=/opt/sparkRapidsPlugin/getSRIOVResources.sh \
       --conf spark.executor.resource.sriov_111.amount=1 \
       --conf spark.executor.resource.sriov_111.vendor=mellanox.com \
       --conf spark.kubernetes.executor.annotation.k8s.v1.cni.cncf.io/networks=sriov111 \
       --conf spark.executorEnv.UCX_TLS=cuda_copy,cuda_ipc,tcp \
       --conf spark.executorEnv.UCX_NET_DEVICES=net1 \
       --conf spark.executorEnv.LD_LIBRARY_PATH=/usr/local/cuda/lib64 \
       --conf spark.executorEnv.CUDA_HOME=/usr/local/cuda \
       --conf spark.kubernetes.container.image=bkovalev/spark3rapids \
       --class com.nvidia.spark.rapids.tests.tpch.CSV /opt/sparkRapidsPlugin/rapids-4-spark-integration-tests_2.12-0.1-SNAPSHOT.jar file:///data/db file:///data/result 22

Example 4:

For RDMA/RC with GPU/RAPIDS/UCX without GPUDirect running we are using the t pch-rapids -rc-ucx-nogdr.sh file with the following configs when running Spark on K8s.

Deployment Node Console

Copy
Copied!

            
            ./spark-submit \
       --master k8s://https://192.168.1.6:6443 \
       --driver-memory 1G \
       --conf spark.executor.instances=6 \
       --conf spark.executor.memory=30G \
       --conf spark.executor.cores=8 \
       --conf spark.task.cpus=4 \
       --conf spark.rapids.memory.pinnedPool.size=2G \
       --conf spark.task.resource.gpu.amount=0.25 \
       --conf spark.rapids.sql.concurrentGpuTasks=1 \
       --conf spark.sql.files.maxPartitionBytes=512m \
       --conf spark.sql.shuffle.partitions=300 \
       --conf spark.locality.wait=0s \
       --conf spark.driver.extraClassPath=/opt/sparkRapidsPlugin/* \
       --conf spark.executor.extraClassPath=/opt/sparkRapidsPlugin/*:/usr/lib/:/data/jar/* \
       --conf spark.plugins=com.nvidia.spark.SQLPlugin \
       --conf spark.shuffle.manager=com.nvidia.spark.RapidsShuffleManager \
       --conf spark.rapids.shuffle.transport.enabled=true \
       --conf spark.executor.resource.gpu.discoveryScript=/opt/sparkRapidsPlugin/getGpusResources.sh \
       --conf spark.executor.resource.gpu.amount=1 \
       --conf spark.executor.resource.gpu.vendor=nvidia.com \
       --conf spark.kubernetes.executor.podTemplateFile=/opt/executor-template/executor-template.yaml \
       --conf spark.executor.resource.sriov_111.discoveryScript=/opt/sparkRapidsPlugin/getSRIOVResources.sh \
       --conf spark.executor.resource.sriov_111.amount=1 \
       --conf spark.executor.resource.sriov_111.vendor=mellanox.com \
       --conf spark.kubernetes.executor.annotation.k8s.v1.cni.cncf.io/networks=sriov111 \
       --conf spark.executorEnv.UCX_TLS=cuda_copy,cuda_ipc,rc \
       --conf spark.executorEnv.UCX_NET_DEVICES=net1 \
       --conf spark.executorEnv.LD_LIBRARY_PATH=/usr/local/cuda/lib64 \
       --conf spark.executorEnv.CUDA_HOME=/usr/local/cuda \
       --conf spark.executorEnv.UCX_RNDV_SCHEME=get_zcopy \
       --conf spark.executorEnv.UCX_IB_GPU_DIRECT_RDMA=no \
       --conf spark.kubernetes.container.image=bkovalev/spark3rapids \
       --class com.nvidia.spark.rapids.tests.tpch.CSV /opt/sparkRapidsPlugin/rapids-4-spark-integration-tests_2.12-0.1-SNAPSHOT.jar file:///data/db
file:///data/result 22

Example 5:

For RDMA/RC with GPU/RAPIDS/UCX/ GPUDirect running we are using the t pch-rapids -rc-ucx-gdr.sh file with the following configs when running Spark on K8s.

Deployment Node Console

Copy
Copied!

            
            ./spark-submit \
       --master k8s://https://192.168.1.6:6443 \
       --driver-memory 1G \
       --conf spark.executor.instances=6 \
       --conf spark.executor.memory=30G \
       --conf spark.executor.cores=8 \
       --conf spark.task.cpus=4 \
       --conf spark.rapids.memory.pinnedPool.size=2G \
       --conf spark.task.resource.gpu.amount=0.25 \
       --conf spark.rapids.sql.concurrentGpuTasks=1 \
       --conf spark.sql.files.maxPartitionBytes=512m \
       --conf spark.sql.shuffle.partitions=300 \
       --conf spark.locality.wait=0s \
       --conf spark.driver.extraClassPath=/opt/sparkRapidsPlugin/* \
       --conf spark.executor.extraClassPath=/opt/sparkRapidsPlugin/*:/usr/lib/:/data/jar/* \
       --conf spark.plugins=com.nvidia.spark.SQLPlugin \
       --conf spark.shuffle.manager=com.nvidia.spark.RapidsShuffleManager \
       --conf spark.rapids.shuffle.transport.enabled=true \
       --conf spark.executor.resource.gpu.discoveryScript=/opt/sparkRapidsPlugin/getGpusResources.sh \
       --conf spark.executor.resource.gpu.amount=1 \
       --conf spark.executor.resource.gpu.vendor=nvidia.com \
       --conf spark.kubernetes.executor.podTemplateFile=/opt/executor-template/executor-template.yaml \
       --conf spark.executor.resource.sriov_111.discoveryScript=/opt/sparkRapidsPlugin/getSRIOVResources.sh \
       --conf spark.executor.resource.sriov_111.amount=1 \
       --conf spark.executor.resource.sriov_111.vendor=mellanox.com \
       --conf spark.kubernetes.executor.annotation.k8s.v1.cni.cncf.io/networks=sriov111 \
       --conf spark.executorEnv.UCX_TLS=cuda_copy,cuda_ipc,rc \
       --conf spark.executorEnv.UCX_NET_DEVICES=net1 \
       --conf spark.executorEnv.LD_LIBRARY_PATH=/usr/local/cuda/lib64 \
       --conf spark.executorEnv.CUDA_HOME=/usr/local/cuda \
       --conf spark.executorEnv.UCX_RNDV_SCHEME=get_zcopy \
       --conf spark.kubernetes.container.image=bkovalev/spark3rapids \
       --class com.nvidia.spark.rapids.tests.tpch.CSV /opt/sparkRapidsPlugin/rapids-4-spark-integration-tests_2.12-0.1-SNAPSHOT.jar file:///data/db
file:///data/result 22

Done!

Appendix A

Archive of the RoCE role for deployment with Kubespray - roce.tar.

Appendix B

General Recommendations

Fewer large input files are better than many small ones. You may not have control over this but it is worth knowing.
Larger input sizes spark.sql.files.maxPartitionBytes=512m are generally better as long as things fit into the GPU.
The GPU does better with larger data chunks as long as they fit into memory. When using the default spark.sql.shuffle.partitions=200 it may be beneficial to make this smaller. Base this on the amount of data the task is reading. Start with 512MB / task.
Out of GPU Memory.

GPU out of memory can show up in multiple ways. You can see an error that it is out of memory or it can also manifest as it just crashes. Generally this means your partition size is too big, go back to the Configuration section for the partition size and/or the number of partitions. Possibly reduce the number of concurrent gpu tasks to 1. The Spark UI may give you a hint at the size of the data. Look at either the input data or the shuffle data size for the stage that failed.

RAPIDS Accelerator for Apache Spark Tuning Guide

Tuning a Spark job configuration settings from the defaults can improve the job performance, and this remains true for jobs leveraging the RAPIDS Accelerator plugin for Apache Spark.

This document provides guidelines on how to tune a Spark job’s configuration settings for improved performance when using the RAPIDS Accelerator plugin.

Monitoring

Since the plugin runs without any API changes, the easiest way to see what is running on the GPU is to look at the "SQL" tab in the Spark Web UI. The SQL tab only shows up after you have actually executed a query. Go to the SQL tab in the UI, click on the query you are interested in and it shows a DAG picture with details. You can also scroll down and twist the "Details" section to see the text representation.

If you want to look at the Spark plan via the code you can use the explain() function call. For example: query.explain() will print the physical plan from Spark and you can see what nodes were replaced with GPU calls.

Debugging

For now, the best way to debug is how you would normally do it on Spark. Look at the UI and log files to see what failed. If you got a seg fault from the GPU find the hs_err_pid.log file. To make sure your hs_err_pid.log file goes into the YARN application log, you may add in the config: --conf spark.executor.extraJavaOptions="-XX:ErrorFile=<LOG_DIR>/hs_err_pid_%p.log"

If you want to see why an operation didn’t run on the GPU, you can turn on the configuration: --conf spark.rapids.sql.explain=NOT_ON_GPU. A log message will then be output to the Driver log as to why a Spark operation isn’t able to run on the GPU.

About the Authors

	Boris Kovalev Boris Kovalev has worked for the past several years as a Solutions Architect, focusing on NVIDIA Networking/Mellanox technology, and is responsible for complex machine learning, Big Data and advanced VMware-based cloud research and design. Boris previously spent more than 20 years as a senior consultant and solutions architect at multiple companies, most recently at VMware. He has written multiple reference designs covering VMware, machine learning, Kubernetes, and container solutions which are available at the Mellanox Documents website.

Boris Kovalev

Boris Kovalev has worked for the past several years as a Solutions Architect, focusing on NVIDIA Networking/Mellanox technology, and is responsible for complex machine learning, Big Data and advanced VMware-based cloud research and design. Boris previously spent more than 20 years as a senior consultant and solutions architect at multiple companies, most recently at VMware. He has written multiple reference designs covering VMware, machine learning, Kubernetes, and container solutions which are available at the Mellanox Documents website.

	Vitaliy Razinkov Over the past few years, Vitaliy Razinkov has been working as a Solutions Architect on the NVIDIA Networking team, responsible for complex Kubernetes/OpenShift and Microsoft's leading solutions, research and design. He previously spent more than 25 years in senior positions at several companies. Vitaliy has written several reference designs guides on Microsoft technologies, RoCE/RDMA accelerated machine learning in Kubernetes/OpenShift, and container solutions, all of which are available on the NVIDIA Networking Documentation website.

Vitaliy Razinkov

Over the past few years, Vitaliy Razinkov has been working as a Solutions Architect on the NVIDIA Networking team, responsible for complex Kubernetes/OpenShift and Microsoft's leading solutions, research and design. He previously spent more than 25 years in senior positions at several companies. Vitaliy has written several reference designs guides on Microsoft technologies, RoCE/RDMA accelerated machine learning in Kubernetes/OpenShift, and container solutions, all of which are available on the NVIDIA Networking Documentation website.

	Peter Rudenko Peter Rudenko is a software engineer on the NVIDIA High Performance Computing (HPC) team who focuses on accelerating data intensive applications, developing the UCX communication X library, and other big data solutions.

Peter Rudenko

Peter Rudenko is a software engineer on the NVIDIA High Performance Computing (HPC) team who focuses on accelerating data intensive applications, developing the UCX communication X library, and other big data solutions.

Spectrum Ethernet Switches vSphere Ethernet RoCE Spark Cumulus Linux VMware ConnectX ESXi and Photon OS Enterprise Kubernetes

Last updated on Sep 12, 2023.

RDG for Apache Spark 3.0 on Kubernetes Accelerated with RAPIDS over RoCE Network

On This Page

Switch console

Switch console

Server console

Server Console

Server Console

Server Console

Server Console

Server Console

inventory/mycluster/hosts.yaml

Deployment Node Console

Deployment Node Console

K8s Master Node Console

K8s Master Node Console

K8s Master Node Console

K8s Master Node Console

K8s Master Node Console

K8s Master Node Console

K8s Master Node Console

K8s Master Node Console

vars/main.yml

Deployment Node Console

Deployment Node Console

K8s Master Node Console

Worker Node resources

network-attachment-definitions

Shell

Configuration:

Shell

Server Console

Server Console

Node cli

Node cli

Server Console

Server Console

Server Console

Server Console

Server Console

Server Console

Server Console

Server Console

Server Console

Deployment Node Console

Deployment Node Console

Deployment Node Console

Deployment Node Console

Deployment Node Console

Deployment Node Console

Deployment Node Console

Worker Node Console

Deployment Node Console

Deployment Node Console

Deployment Node Console

Deployment Node Console

Deployment Node Console

Deployment Node Console