# cri-virtplus

**Repository Path**: iscas-system/cri-virtplus

## Basic Information

- **Project Name**: cri-virtplus
- **Description**: GPU、NPU/MLU、DCU、FPGA等专用智能处理单元管理。请联系吴恒，邮箱地址：wuheng@otcaix.iscas.ac.cn
- **Primary Language**: Unknown
- **License**: Apache-2.0
- **Default Branch**: refactoring
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2024-04-26
- **Last Updated**: 2026-04-07

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# share

扩展容器支持GPU、MLU、DCU算力。

## 前置要求

glibc<2.34

## 当前支持

|      | Test Environment                                             | Driver & Platform                      | Hardware  |
| ---- | ------------------------------------------------------------ | -------------------------------------- | --------- |
| GPU  | Ubuntu version <= 22.04<br />Kubernetes version 1.23.6<br />Docker version 20.04<br />x86 | driver 11.7<br />CUDAtoolkit 11.7 11.2 | Telsa T4  |
| MLU  | Ubuntu 18.04<br />Kubernetes version 1.23.6<br />Docker version 20.04<br />x86 | driver v5.10.22<br />CNtoolkit 3.7.2   | MLU370-S4 |

## roadmap

- 1.0
  - [X] support local GPU/DCU 
  - [X] support local MLU

- 2.0
  - [ ] support RDMA

## 部署

- 前置要求

  Kubernetes/k3s/cas、Docker/isulad

  GPU Driver、CUDAtoolkit

  Mlu Driver、CNtoolkit
  安装参考: ./docs/env.md

- 构建拦截库，复制到/etc/unishare/hooklib目录

``````
cd ./core
mkdir build
cd build
#容器使用MLU
cmake -DUSE_MLU=ON ..
#容器使用GPU
cmake -DUSE_GPU=ON ..
#容器使用GPU、MLU
cmake ..
make
cp libmylibrary.so /etc/unishare/hooklib
``````

- 复制MLU、GPU库文件至/usr/local/accelerator

  ``````
  cd ./k8sPlugins/deploy
  //修改copy.sh中FILE变量，指向volume.conf的绝对路径
  chmod +x copy.sh
  ``````

- 修改uni-share.yaml文件

  切换至根目录，修改./k8sPlugins/deploy/uni-share.yaml中的url和token

    详情参考https://github.com/kubesys/client-go

- 创建CRD

  ``````
  cd ./k8sPlugins/deploy
  kubectl apply -f crd.yaml
  kubectl apply -f mluCrd.yaml
  ``````

- 构建镜像

  `docker build -t uni-share:v1.0 --load .   `

- 启动daemonset

``````
cd ./k8sPlugins/deploy
kubectl apply -f uni-share.yaml
``````

## pod模板示例

通过resources中的requests和limits指定pod的资源需求，requests和limits须保持一致
core对应GPU的利用率（0-100）
memory对应显存，1点代表256Mi显存（不应分配超过一张卡显存）

``````
apiVersion: v1
kind: Pod
metadata:
  name: testpod-10.2
spec:
  restartPolicy: Never
  containers:
    - name: cuda102
      image: testgpuimg:cuda10.2
      imagePullPolicy: Never
      command: ["sleep"]
      args: ["infinity"]
      resources:
        requests:
          iscas.cn/gpu-core: 30
          #iscas.cn/mlu-core： 30
          iscas.cn/gpu-memory: 20
          #iscas.cn/mlu-memory： 20
        limits:
          iscas.cn/gpu-core: 30
          #iscas.cn/mlu-core： 30
          iscas.cn/gpu-memory: 20
          #iscas.cn/mlu-memory： 20
``````
镜像参考：
  https://hub.docker.com/r/pytorch/pytorch
  https://sdk.cambricon.com/static/PyTorch/MLU370_1.9_v1.17.0_X86_ubuntu18.04_python3.6_docker/
任务测试参考：
  ./docs/test.md