GPU Sharing Scheduler Extender in Kubernetes

Overview

More and more data scientists run their Nvidia GPU based inference tasks on Kubernetes. Some of these tasks can be run on the same Nvidia GPU device to increase GPU utilization. So one important challenge is how to share GPUs between the pods. The community is also very interested in this topic.

Now there is a GPU sharing solution on native Kubernetes: it is based on scheduler extenders and device plugin mechanism, so you can reuse this solution easily in your own Kubernetes.

Prerequisites

Kubernetes 1.11+
golang 1.10+
NVIDIA drivers ~= 361.93
Nvidia-docker version > 2.0 (see how to install and it's prerequisites)
Docker configured with Nvidia as the default runtime.

Design

For more details about the design of this project, please read this Design document.

Setup

You can follow this Installation Guide. If you are using Alibaba Cloud Kubernetes, please follow this doc to install with Helm Charts.

User Guide

You can check this User Guide.

Developing

Scheduler Extender

git clone https://github.com/AliyunContainerService/gpushare-scheduler-extender.git && cd gpushare-scheduler-extender
docker build -t cheyang/gpushare-scheduler-extender .

Device Plugin

git clone https://github.com/AliyunContainerService/gpushare-device-plugin.git && cd gpushare-device-plugin
docker build -t cheyang/gpushare-device-plugin .

Kubectl Extension

golang > 1.10

mkdir -p $GOPATH/src/github.com/AliyunContainerService
cd $GOPATH/src/github.com/AliyunContainerService
git clone https://github.com/AliyunContainerService/gpushare-device-plugin.git
cd gpushare-device-plugin
go build -o $GOPATH/bin/kubectl-inspect-gpushare-v2 cmd/inspect/*.go

Demo

- Demo 1: Deploy multiple GPU Shared Pods and schedule them on the same GPU device in binpack way

- Demo 2: Avoid GPU memory requests that fit at the node level, but not at the GPU device level

Related Project

gpushare device plugin

Roadmap

Integrate Nvidia MPS as the option for isolation
Automated Deployment for the Kubernetes cluster which is deployed by kubeadm
Scheduler Extener High Availablity
Generic Solution for GPU, RDMA and other devices

Acknowledgments

GPU sharing solution is based on Nvidia Docker2, and their gpu sharing design is our reference. The Nvidia Community is very supportive and We are very grateful.

xjas/gpushare-scheduler-extender