Enable Hygon DCU sharing

Introduction

HAMi now supports hygon.com/dcu by implementing most device-sharing features as NVIDIA GPUs, including:

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

Device Memory Control: DCUs can be allocated with certain device memory size on certain type(i.e Z100) and have made it that it does not exceed the boundary.

Device compute core limitation: DCUs can be allocated with certain percentage of device core(i.e hygon.com/dcucores:60 indicate this container uses 60% compute cores of this device)

DCU Type Specification: You can specify which type of DCU to use or to avoid for a certain task, by setting "hygon.com/use-dcutype" or "hygon.com/nouse-dcutype" annotations.

Prerequisites

• dtk driver >= 24.04

Enabling DCU-sharing Support

• Deploy the dcu-vgpu-device-plugin here

Running DCU jobs

Hygon DCUs can now be requested by a container using the hygon.com/dcunum , hygon.com/dcumem and hygon.com/dcucores resource type:

apiVersion: v1
kind: Pod
metadata:
  name: alexnet-tf-gpu-pod-mem
  labels:
    purpose: demo-tf-amdgpu
spec:
  containers:
    - name: alexnet-tf-gpu-container
      image: pytorch:resnet50
      workingDir: /root
      command: ["sleep","infinity"]
      resources:
        limits:
          hygon.com/dcunum: 1 # requesting a GPU
          hygon.com/dcumem: 2000 # each dcu require 2000 MiB device memory
          hygon.com/dcucores: 60 # each dcu use 60% of total compute cores

Enable vDCU inside container

You need to enable vDCU inside container in order to use it.

source /opt/hygondriver/env.sh

check if you have successfully enabled vDCU by using following command

hy-virtual -show-device-info

If you have an output like this, then you have successfully enabled vDCU inside container.

Device 0:
	Actual Device: 0
	Compute units: 60
	Global memory: 2097152000 bytes

Launch your DCU tasks like you usually do

Notes

1. If your image is not a 'dtk-embedded-image', you need to install pciutiils, libelf-dev, kmod after task running, otherwise, dcu-tools like hy-smi or hy-virtual may not work properly

2. DCU-sharing in init container is not supported, pods with "hygon.com/dcumem" in init container will never be scheduled.

3. Only one vdcu can be acquired per container. If you want to mount multiple dcu devices, then you shouldn't set hygon.com/dcumem or hygon.com/dcucores