KubeComp

"KubeComp: Resource-centric Composable Container Orchestrator for GPU Pooling" is accepted by UCC 2024!

Introduction

KubeComp is the first solution for Kubernetes to support GPU pooling based on a composable infrastructure. KubeComp makes Kubernetes aware of the underlying composable infrastructure and provides on-demand, optimized, and automated resource management for service deployment.

Deployment

• Cluster Setup
• Resource Pool Simulator
• Disaggregate Device Plugin
• KubeComp Scheduler
• Reconfig Manager

Demo

This is an example yaml file demonstrate how to specify the disaggregated resource and the KubeComp Scheduler.

apiVersion: v1
kind: Pod
metadata:
  name: demo-pod
spec:
  schedulerName: kubecomp-scheduler
  containers:
  - name: demo-pod
    image: "alpine"
    command: ['sh', '-c', 'echo "start" && sleep 3000']
    resources:
      requests:
        memory: "128Mi"
        cpu: "500m"
        falcon.com/gpu: 5
      limits:
        memory: "128Mi"
        cpu: "500m"
        falcon.com/gpu: 5
  restartPolicy: Never

Since the cluster was originally set up with three worker nodes, each with 4 GPUs, the reallocation of 1 GPU will be observed.

Video

https://youtu.be/uqvb6CrrJoU