Enable Iluvatar GCU sharing

Introduction

HAMi now supports iluvatar.ai/gpu(i.e MR-V100、BI-V150、BI-V100) by implementing most device-sharing features as NVIDIA GPUs, including:

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

Device Memory Control: GPUs can be allocated with certain device memory size and have made it that it does not exceed the boundary.

Device Core Control: GPUs can be allocated with limited compute cores and have made it that it does not exceed the boundary.

Device UUID Selection: You can specify which GPU 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

• Iluvatar gpu-manager (please consult your device provider)
• driver version > 3.1.0

Enabling GPU-sharing Support

• Deploy gpu-manager on iluvatar nodes (Please consult your device provider to acquire its package and document)

NOTICE: Install only gpu-manager, don't install gpu-admission package.

• Identify the resource name about core and memory usage(i.e 'iluvatar.ai/vcuda-core', 'iluvatar.ai/vcuda-memory')

• set the 'iluvatarResourceMem' and 'iluvatarResourceCore' parameters when install hami

helm install hami hami-charts/hami --set scheduler.kubeScheduler.imageTag={your kubernetes version} --set iluvatarResourceMem=iluvatar.ai/vcuda-memory --set iluvatarResourceCore=iluvatar.ai/vcuda-core -n kube-system

备注

The default resource names are:

• iluvatar.ai/vgpu for GPU count
• iluvatar.ai/vcuda-memory for memory allocation
• iluvatar.ai/vcuda-core for core allocation

You can customize these names using the parameters above.

Device Granularity

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

Memory Allocation

• Each unit of iluvatar.ai/vcuda-memory represents 256MB of device memory
• 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

• Each unit of iluvatar.ai/vcuda-core represents 1% of the available compute cores
• Core allocation is enforced with hard limits to ensure tasks don't exceed their allocated cores
• When requesting multiple GPUs, the system will automatically set the core resources based on the number of GPUs requested

Running Iluvatar jobs

Iluvatar GPUs can now be requested by a container using the iluvatar.ai/vgpu, iluvatar.ai/vcuda-memory and iluvatar.ai/vcuda-core resource type:

apiVersion: v1
kind: Pod
metadata:
  name: poddemo
spec:
  restartPolicy: Never
  containers:
  - name: poddemo
    image: harbor.4pd.io/vgpu/corex_transformers@sha256:36a01ec452e6ee63c7aa08bfa1fa16d469ad19cc1e6000cf120ada83e4ceec1e
    command:
    - bash
    args:
    - -c
    - |
      set -ex
      echo "export LD_LIBRARY_PATH=/usr/local/corex/lib64:$LD_LIBRARY_PATH">> /root/.bashrc
      cp -f /usr/local/iluvatar/lib64/libcuda.* /usr/local/corex/lib64/
      cp -f /usr/local/iluvatar/lib64/libixml.* /usr/local/corex/lib64/
      source /root/.bashrc
      sleep 360000
    resources:
      requests:
        iluvatar.ai/vgpu: 1
        iluvatar.ai/vcuda-core: 50
        iluvatar.ai/vcuda-memory: 64
      limits:
        iluvatar.ai/vgpu: 1
        iluvatar.ai/vcuda-core: 50
        iluvatar.ai/vcuda-memory: 64

备注

Each unit of vcuda-memory indicates 256M device memory

备注

You can find more examples in the examples/iluvatar folder

Notes

1. You need to set the following prestart command in order for the device-share to work properly

      set -ex
      echo "export LD_LIBRARY_PATH=/usr/local/corex/lib64:$LD_LIBRARY_PATH">> /root/.bashrc
      cp -f /usr/local/iluvatar/lib64/libcuda.* /usr/local/corex/lib64/
      cp -f /usr/local/iluvatar/lib64/libixml.* /usr/local/corex/lib64/
      source /root/.bashrc

2. Virtualization takes effect only for containers that apply for one GPU(i.e iluvatar.ai/vgpu=1 ). When requesting multiple GPUs, the system will automatically set the core resources based on the number of GPUs requested.

3. The iluvatar.ai/vcuda-memory resource is only effective when iluvatar.ai/vgpu=1.

4. Multi-device requests (iluvatar.ai/vgpu > 1) do not support vGPU mode.