Enable Enflame GCU sharing

Introduction

HAMi now supports sharing on enflame.com/gcu(i.e S60) by implementing most device-sharing features as NVIDIA GPUs, including:

GCU sharing: Each task can allocate a portion of GCU instead of a whole GCU card, thus GCU can be shared among multiple tasks.

Device Memory and Core Control: GCUs can be allocated with certain percentage of device memory and core, HAMi ensures it does not exceed the boundary.

Device UUID Selection: You can specify which GCU devices to use or exclude using annotations.

Very Easy to use: You don't need to modify your task yaml to use the HAMi scheduler. All your GPU jobs will be automatically supported after installation.

Prerequisites

• Enflame gcushare-device-plugin >= 2.1.6 (please consult your device provider, gcushare has two components: gcushare-scheduler-plugin and gcushare-device-plugin, only gcushare-device-plugin is needed here )
• driver version >= 1.2.3.14
• kubernetes >= 1.24
• enflame-container-toolkit >=2.0.50

Enabling GCU-sharing Support

• Deploy gcushare-device-plugin on enflame nodes (Please consult your device provider to acquire its package and document)

NOTICE: Install only gpushare-device-plugin, don't install gpu-scheduler-plugin package.

NOTE: The default resource names are:

• enflame.com/vgcu for GCU count, only support 1 now.
• enflame.com/vgcu-percentage for the percentage of memory and cores in a gcu slice.

You can customize these names by modifying hami-scheduler-device configMap above.

• Set 'devices.enflame.enabled=true' when deploy HAMi

helm install hami hami-charts/hami --set devices.enflame.enabled=true -n kube-system

Device Granularity

HAMi divides each Enflame GCU into 100 units for resource allocation. When you request a portion of a GPU, you're actually requesting a certain number of these units.

GCU Slice Allocation

• Each unit of enflame.com/vgcu-percentage represents 1% device memory and 1% core
• If you don't specify a memory request, the system will default to using 100% of the available memory
• Memory allocation is enforced with hard limits to ensure tasks don't exceed their allocated memory
• Core allocation is enforced with hard limits to ensure tasks don't exceed their allocated cores

Running Enflame jobs

Enflame GCUs can now be requested by a container using the enflame.com/vgcu and enflame.com/vgcu-percentage resource type:

apiVersion: v1
kind: Pod
metadata:
  name: gcushare-pod-2
  namespace: kube-system
spec:
  terminationGracePeriodSeconds: 0
  containers:
    - name: pod-gcu-example1
      image: ubuntu:18.04
      imagePullPolicy: IfNotPresent
      command:
        - sleep
      args:
        - '100000'
      resources:
        limits:
          enflame.com/vgcu: 1
          enflame.com/vgcu-percentage: 22

NOTICE: You can find more examples in examples folder

Device UUID Selection

You can specify which GPU devices to use or exclude using annotations:

apiVersion: v1
kind: Pod
metadata:
  name: poddemo
  annotations:
    # Use specific GPU devices (comma-separated list)
    enflame.com/use-gpuuuid: "node1-enflame-0,node1-enflame-1"
    # Or exclude specific GPU devices (comma-separated list)
    enflame.com/nouse-gpuuuid: "node1-enflame-2,node1-enflame-3"
spec:
  # ... rest of pod spec

NOTE: The device ID format is {node-name}-enflame-{index}. You can find the available device IDs in the node status.

Finding Device UUIDs

You can find the UUIDs of Enflame GCUs on a node using the following command:

kubectl get pod <pod-name> -o yaml | grep -A 10 "hami.io/<card-type>-devices-allocated"

Or by examining the node annotations:

kubectl get node <node-name> -o yaml | grep -A 10 "hami.io/node-register-<card-type>"

Look for annotations containing device information in the node status.

Notes

1. GCUshare takes effect only for containers that apply for one GCU(i.e enflame.com/vgcu=1 ).

2. Multiple GCU allocation in one container is not supported yet

3. efsmi inside container shows the total device memory, which is NOT a bug, device memory will be properly limited when running tasks.