mnist Model Training Example: Run MPIJob on Multi-node Cluster

The below is an mnist example for model training on Amazon EKS.

Build and Store Custom Docker Image

  1. Create a Dockerfile with the below content:

  FROM vault.habana.ai/gaudi-docker/1.8.0/ubuntu20.04/habanalabs/tensorflow-installer-tf-cpu-2.11.0

  RUN git clone -b 1.8.0 https://github.com/HabanaAI/Model-References.git

 FROM vault.habana.ai/gaudi-docker/1.8.0/ubuntu20.04/habanalabs/pytorch-installer-1.13.1:latest

 RUN git clone -b 1.8.0 https://github.com/HabanaAI/Model-References.git

1. Build and push image to AWS’s Elastic Container Registry (ECR) for ease of access on EC2 instances. For further information on how to build and push an image to ECR, refer to Create Elastic Container Registry (ECR) and Upload Images or to the Amazon ECR Getting Started Guide.

Run MPIJob on Multi-node Cluster

  1. Create mpijob-mnist.yaml file. The config file can pull a docker image and set up a container according to habana.ai/gaudi, hugepages-2Mi, memory, etc. These three parameters could be adapted by your task and model.

The following is an example of mpijob-mnist.yaml. Check model code README for details on how to run Multi-Node training:

apiVersion: kubeflow.org/v2beta1
kind: MPIJob
metadata:
  name: mnist-distributed
spec:
  slotsPerWorker: 8
  runPolicy:
    cleanPodPolicy: Running
  mpiReplicaSpecs:
    Launcher:
      replicas: 1
      restartPolicy: Never
      template:
        spec:
          imagePullSecrets:
            - name: private-registry
          terminationGracePeriodSeconds: 0
          containers:
            - image: <Custom Docker Image>
              imagePullPolicy: Always
              name: mnist-launcher
              command:
                - bash
                - -c
                - mpirun --allow-run-as-root --bind-to core -np 16 --map-by ppr:4:socket:PE=6 --merge-stderr-to-stdout
                  --prefix /opt/amazon/openmpi
                  -x PYTHONPATH=/Model-References:/usr/lib/habanalabs
                  -x LD_LIBRARY_PATH=/opt/amazon/openmpi/lib:/opt/amazon/efa/lib/:${LD_LIBRARY_PATH}
                  -x RDMAV_FORK_SAFE=1
                  -x FI_EFA_USE_DEVICE_RDMA=1
                  python3 /Model-References/TensorFlow/examples/hello_world/example_hvd.py
              resources:
                requests:
                  cpu: "100m"
    Worker:
      replicas: 2
      template:
        spec:
          imagePullSecrets:
            - name: private-registry
          terminationGracePeriodSeconds: 0
          containers:
            - image: <Custom Docker Image>
              name: mnist-worker
              securityContext:
                capabilities:
                  add: ["SYS_NICE"]
              resources:
                requests:
                  habana.ai/gaudi: 8
                  hugepages-2Mi: "42000Mi"
                  vpc.amazonaws.com/efa: 4
                  cpu: "90"
                limits:
                  habana.ai/gaudi: 8
                  hugepages-2Mi: "42000Mi"
                  vpc.amazonaws.com/efa: 4
                  cpu: "90"

The following is an example of mpijob-mnist.yaml. Check model code README for details on how to run Multi-Node training:

Note

PyTorch uses shared memory buffers to communicate between processes. By default, Docker containers are allocated 64MB of shared memory. When using > 1 HPU, this can be insufficient. To bypass this issue, specifying hostIPC: true will re-use the host’s shared memory space inside the container.

apiVersion: kubeflow.org/v2beta1
kind: MPIJob
metadata:
  name: mnist-distributed
spec:
  slotsPerWorker: 8
  runPolicy:
    cleanPodPolicy: Running
  mpiReplicaSpecs:
    Launcher:
      replicas: 1
      restartPolicy: Never
      template:
        spec:
          imagePullSecrets:
            - name: private-registry
          terminationGracePeriodSeconds: 0
          containers:
            - image: <Custom Docker Image>
              imagePullPolicy: Always
              name: mnist-launcher
              command:
                - bash
                - -c
                - mpirun --allow-run-as-root --bind-to core -np 16 --map-by ppr:4:socket:PE=6 --merge-stderr-to-stdout
                  --prefix /opt/amazon/openmpi
                  -x PYTHONPATH=/Model-References:/usr/lib/habanalabs
                  -x LD_LIBRARY_PATH=/opt/amazon/openmpi/lib:/opt/amazon/efa/lib/:${LD_LIBRARY_PATH}
                  -x RDMAV_FORK_SAFE=1
                  -x FI_EFA_USE_DEVICE_RDMA=1
                  python3 /Model-References/PyTorch/examples/computer_vision/hello_world/mnist.py
                  --batch-size=64 --epochs=1 --lr=1.0 --gamma=0.7 --hpu --use_lazy_mode
              resources:
                requests:
                  cpu: "100m"
    Worker:
      replicas: 2
      template:
        spec:
          imagePullSecrets:
            - name: private-registry
          terminationGracePeriodSeconds: 0
          hostIPC: true
          containers:
            - image: <Custom Docker Image>
              name: mnist-worker
              securityContext:
                capabilities:
                  add: ["SYS_NICE"]
              resources:
                requests:
                  habana.ai/gaudi: 8
                  hugepages-2Mi: "42000Mi"
                  vpc.amazonaws.com/efa: 4
                  cpu: "90"
                limits:
                  habana.ai/gaudi: 8
                  hugepages-2Mi: "42000Mi"
                  vpc.amazonaws.com/efa: 4
                  cpu: "90"

  1. Update the parameters listed below to run the desired configuration:

    Parameter

    Description

    <Custom Docker Image>

    Image with Resnet dependencies installed

    -np 16

    Number of HPU Cards for training. Should update accordingly to match replicas: 2, the number of DL1’s for training

    replicas: 2

    Number of DL1s for training. Should update accordingly to match -np 16, the number of HPUS for training

  2. Run the job by running the following command:

    kubectl apply -f mpijob-mnist.yaml

  3. Check the job status by running the following command:

    kubectl get pods -A

  4. Retrieve the name of the created pod and run the following command to see the results:

    kubectl logs <pod-name>