Creating a Custom Training Image (MPI + CPU/GPU)

Scenario

In this example, write a Dockerfile to create a custom image on a Linux x86_64 server running Ubuntu 18.04.

Objective: Build and install container images of the following software and use the images and CPUs/GPUs for training jobs on ModelArts.

• ubuntu-18.04
• cuda-11.1
• python-3.7.13
• openmpi-3.0.0

Procedure

Before using a custom image to create a training job, get familiar with Docker and have development experience. The following is the detailed procedure:

1. Prerequisites
2. Step 1 Creating an OBS Bucket and Folder
3. Step 2 Preparing Script Files and Uploading Them to OBS
4. Step 3 Preparing a Host
5. Step 4 Creating a Custom Image
6. Step 5 Uploading an Image to SWR
7. Step 6 Creating a Training Job on ModelArts

Prerequisites

You have registered a Huawei ID and enabled Huawei Cloud services, and the account is not in arrears or frozen.

Step 1 Creating an OBS Bucket and Folder

Create a bucket and the folders in OBS for storing the sample dataset and training code. Table 1 lists the folders to be created. Replace the bucket name test-modelarts and folder name in the example with the actual names.

For details about how to create an OBS bucket and folder, see Creating a Bucket and Creating a Folder.

Ensure that the OBS directory you use and ModelArts are in the same region.

Step 2 Preparing Script Files and Uploading Them to OBS

Prepare the MPI boot script run_mpi.sh and training script mpi-verification.py and upload them to the obs://test-modelarts/mpi/demo-code/ folder of the OBS bucket.

• The content of the MPI boot script run_mpi.sh is as follows:

  #!/bin/bash
  MY_HOME=/home/ma-user
  
  MY_SSHD_PORT=${MY_SSHD_PORT:-"38888"}
  
  MY_TASK_INDEX=${MA_TASK_INDEX:-${VC_TASK_INDEX:-${VK_TASK_INDEX}}}
  
  MY_MPI_SLOTS=${MY_MPI_SLOTS:-"${MA_NUM_GPUS}"}
  
  MY_MPI_TUNE_FILE="${MY_HOME}/env_for_user_process"
  
  if [ -z ${MY_MPI_SLOTS} ]; then
      echo "[run_mpi] MY_MPI_SLOTS is empty, set it be 1"
      MY_MPI_SLOTS="1"
  fi
  
  printf "MY_HOME: ${MY_HOME}\nMY_SSHD_PORT: ${MY_SSHD_PORT}\nMY_MPI_BTL_TCP_IF: ${MY_MPI_BTL_TCP_IF}\nMY_TASK_INDEX: ${MY_TASK_INDEX}\nMY_MPI_SLOTS: ${MY_MPI_SLOTS}\n"
  
  env | grep -E '^MA_|^SHARED_|^S3_|^PATH|^VC_WORKER_|^SCC|^CRED' | grep -v '=$' > ${MY_MPI_TUNE_FILE}
  # add -x to each line
  sed -i 's/^/-x /' ${MY_MPI_TUNE_FILE}
  
  sed -i "s|{{MY_SSHD_PORT}}|${MY_SSHD_PORT}|g" ${MY_HOME}/etc/ssh/sshd_config
  
  # start sshd service
  bash -c "$(which sshd) -f ${MY_HOME}/etc/ssh/sshd_config"
  
  # confirm the sshd is up
  netstat -anp | grep LIS | grep ${MY_SSHD_PORT}
  
  if [ $MY_TASK_INDEX -eq 0 ]; then
      # generate the hostfile of mpi
      for ((i=0; i<$MA_NUM_HOSTS; i++))
      do
          eval hostname=${MA_VJ_NAME}-${MA_TASK_NAME}-${i}.${MA_VJ_NAME}
          echo "[run_mpi] hostname: ${hostname}"
  
          ip=""
          while [ -z "$ip" ]; do
              ip=$(ping -c 1 ${hostname} | grep "PING" | sed -E 's/PING .* .([0-9.]+). .*/\1/g')
              sleep 1
          done
          echo "[run_mpi] resolved ip: ${ip}"
  
          # test the sshd is up
          while :
          do
              if [ cat < /dev/null >/dev/tcp/${ip}/${MY_SSHD_PORT} ]; then
                  break
              fi
              sleep 1
          done
  
          echo "[run_mpi] the sshd of ip ${ip} is up"
  
          echo "${ip} slots=$MY_MPI_SLOTS" >> ${MY_HOME}/hostfile
      done
  
      printf "[run_mpi] hostfile:\n`cat ${MY_HOME}/hostfile`\n"
  fi
  
  RET_CODE=0
  
  if [ $MY_TASK_INDEX -eq 0 ]; then
  
      echo "[run_mpi] start exec command time: "$(date +"%Y-%m-%d-%H:%M:%S")
  
      np=$(( ${MA_NUM_HOSTS} * ${MY_MPI_SLOTS} ))
  
      echo "[run_mpi] command: mpirun -np ${np} -hostfile ${MY_HOME}/hostfile -mca plm_rsh_args \"-p ${MY_SSHD_PORT}\" -tune ${MY_MPI_TUNE_FILE} ... $@"
  
      # execute mpirun at worker-0
      # mpirun
      mpirun \
          -np ${np} \
          -hostfile ${MY_HOME}/hostfile \
          -mca plm_rsh_args "-p ${MY_SSHD_PORT}" \
          -tune ${MY_MPI_TUNE_FILE} \
          -bind-to none -map-by slot \
          -x NCCL_DEBUG -x NCCL_SOCKET_IFNAME -x NCCL_IB_HCA -x NCCL_IB_TIMEOUT -x NCCL_IB_GID_INDEX -x NCCL_IB_TC \
          -x HOROVOD_MPI_THREADS_DISABLE=1 \
          -x PATH -x LD_LIBRARY_PATH \
          -mca pml ob1 -mca btl ^openib -mca plm_rsh_no_tree_spawn true \
          "$@"
  
      RET_CODE=$?
  
      if [ $RET_CODE -ne 0 ]; then
          echo "[run_mpi] exec command failed, exited with $RET_CODE"
      else
          echo "[run_mpi] exec command successfully, exited with $RET_CODE"
      fi
  
      # stop 1...N worker by killing the sleep proc
      sed -i '1d' ${MY_HOME}/hostfile
      if [ `cat ${MY_HOME}/hostfile | wc -l` -ne 0 ]; then
          echo "[run_mpi] stop 1 to (N - 1) worker by killing the sleep proc"
  
          sed -i 's/${MY_MPI_SLOTS}/1/g' ${MY_HOME}/hostfile
          printf "[run_mpi] hostfile:\n`cat ${MY_HOME}/hostfile`\n"
  
          mpirun \
          --hostfile ${MY_HOME}/hostfile \
          --mca plm_rsh_args "-p ${MY_SSHD_PORT}" \
          -x PATH -x LD_LIBRARY_PATH \
          pkill sleep \
          > /dev/null 2>&1
      fi
  
      echo "[run_mpi] exit time: "$(date +"%Y-%m-%d-%H:%M:%S")
  else
      echo "[run_mpi] the training log is in worker-0"
      sleep 365d
      echo "[run_mpi] exit time: "$(date +"%Y-%m-%d-%H:%M:%S")
  fi
  
  exit $RET_CODE

  

  The script run_mpi.sh uses LF line endings. If CRLF line endings are used, executing the training job will fail, and the error "$'\r': command not found" will be displayed in logs.

• The content of the training script mpi-verification.py is as follows:

  import os
  import socket
  
  if __name__ == '__main__':
      print(socket.gethostname())
  
      # https://www.open-mpi.org/faq/?category=running#mpi-environmental-variables
      print('OMPI_COMM_WORLD_SIZE: ' + os.environ['OMPI_COMM_WORLD_SIZE'])
      print('OMPI_COMM_WORLD_RANK: ' + os.environ['OMPI_COMM_WORLD_RANK'])
      print('OMPI_COMM_WORLD_LOCAL_RANK: ' + os.environ['OMPI_COMM_WORLD_LOCAL_RANK'])

Step 3 Preparing a Host

Obtain a Linux x86_64 server running Ubuntu 18.04. Either an ECS or your local PC will do.

For details about how to purchase an ECS, see Purchasing and Logging In to a Linux ECS. Set CPU Architecture to x86 and Image to Public image. Ubuntu 18.04 images are recommended.

Step 4 Creating a Custom Image

Create a container image with the following configurations and use the image to create a training job on ModelArts:

• ubuntu-18.04
• cuda-11.1
• python-3.7.13
• openmpi-3.0.0

The following describes how to create a custom image by writing a Dockerfile.

1. Install Docker.

   The following uses the Linux x86_64 OS as an example to describe how to obtain the Docker installation package. For details about how to install Docker, see official Docker documents. Run the following commands to install Docker:

   curl -fsSL get.docker.com -o get-docker.sh
   sh get-docker.sh

   If the docker images command is executed, Docker has been installed. In this case, skip this step.

2. Check the Docker engine version. Run the following command:

   docker version | grep -A 1 Engine

   The command output is as follows:

    Engine:
     Version:          18.09.0

   

   Use the Docker engine of the preceding version or later to create a custom image.

3. Create a folder named context.

   mkdir -p context

4. Download the Miniconda3 installation file.

   Download the Miniconda3 py37 4.12.0 installation file (Python 3.7.13) from https://repo.anaconda.com/miniconda/Miniconda3-py37_4.12.0-Linux-x86_64.sh.

5. Download the openmpi 3.0.0 installation file.

   Download the openmpi 3.0.0 file edited using Horovod v0.22.1 from https://github.com/horovod/horovod/files/1596799/openmpi-3.0.0-bin.tar.gz.

6. Store the Miniconda3 and openmpi 3.0.0 files in the context folder. The following shows the context folder:

   context
   ├── Miniconda3-py37_4.12.0-Linux-x86_64.sh
   └── openmpi-3.0.0-bin.tar.gz

7. Write the Dockerfile of the container image.
   Create an empty file named Dockerfile in the context folder and copy the following content to the file:

   # The server on which the container image is created must access the Internet.
   
   # Base container image at https://github.com/NVIDIA/nvidia-docker/wiki/CUDA
   #
   # https://docs.docker.com/develop/develop-images/multistage-build/#use-multi-stage-builds
   # require Docker Engine >= 17.05
   #
   # builder stage
   FROM nvidia/cuda:11.1.1-runtime-ubuntu18.04 AS builder
   
   # The default user of the base container image is root.
   # USER root
   
   # Copy the Miniconda3 (Python 3.7.13) installation files to the /tmp directory of the basic container image.
   COPY Miniconda3-py37_4.12.0-Linux-x86_64.sh /tmp
   
   # Install Miniconda3 in the /home/ma-user/miniconda3 directory of the base container image.
   # https://conda.io/projects/conda/en/latest/user-guide/install/linux.html#installing-on-linux
   RUN bash /tmp/Miniconda3-py37_4.12.0-Linux-x86_64.sh -b -p /home/ma-user/miniconda3
   
   # Create the container image.
   FROM nvidia/cuda:11.1.1-runtime-ubuntu18.04
   
   # Install vim, cURL, net-tools, and the SSH tool in Huawei Mirrors.
   RUN cp -a /etc/apt/sources.list /etc/apt/sources.list.bak && \
       sed -i "s@http://.*archive.ubuntu.com@http://repo.huaweicloud.com@g" /etc/apt/sources.list && \
       sed -i "s@http://.*security.ubuntu.com@http://repo.huaweicloud.com@g" /etc/apt/sources.list && \
       echo > /etc/apt/apt.conf.d/00skip-verify-peer.conf "Acquire { https::Verify-Peer false }" && \
       apt-get update && \
       apt-get install -y vim curl net-tools iputils-ping \
       openssh-client openssh-server && \
       ssh -V && \
       mkdir -p /run/sshd && \
       apt-get clean && \
       mv /etc/apt/sources.list.bak /etc/apt/sources.list && \
       rm /etc/apt/apt.conf.d/00skip-verify-peer.conf
   
   # Install the Open MPI 3.0.0 file written using Horovod v0.22.1.
   # https://github.com/horovod/horovod/blob/v0.22.1/docker/horovod/Dockerfile
   # https://github.com/horovod/horovod/files/1596799/openmpi-3.0.0-bin.tar.gz
   COPY openmpi-3.0.0-bin.tar.gz /tmp
   RUN cd /usr/local && \
       tar -zxf /tmp/openmpi-3.0.0-bin.tar.gz && \
       ldconfig && \
       mpirun --version
   
   # Add user ma-user (UID = 1000, GID = 100).
   # A user group whose GID is 100 exists in the base container image. User ma-user can directly run the following command:
   RUN useradd -m -d /home/ma-user -s /bin/bash -g 100 -u 1000 ma-user
   
   # Copy the /home/ma-user/miniconda3 directory from the builder stage to the directory with the same name in the current container image.
   COPY --chown=ma-user:100 --from=builder /home/ma-user/miniconda3 /home/ma-user/miniconda3
   
   # Configure the preset environment variables of the container image.
   # Set PYTHONUNBUFFERED to 1 to prevent log loss.
   ENV PATH=$PATH:/home/ma-user/miniconda3/bin \
       PYTHONUNBUFFERED=1
   
   # Configure the default user and working directory of the container image.
   USER ma-user
   WORKDIR /home/ma-user
   
   # Configure sshd to support SSH password-free login.
   RUN MA_HOME=/home/ma-user && \
       # setup sshd dir
       mkdir -p ${MA_HOME}/etc && \
       ssh-keygen -f ${MA_HOME}/etc/ssh_host_rsa_key -N '' -t rsa  && \
       mkdir -p ${MA_HOME}/etc/ssh ${MA_HOME}/var/run  && \
       # setup sshd config (listen at {{MY_SSHD_PORT}} port)
       echo "Port {{MY_SSHD_PORT}}\n\
   HostKey ${MA_HOME}/etc/ssh_host_rsa_key\n\
   AuthorizedKeysFile ${MA_HOME}/.ssh/authorized_keys\n\
   PidFile ${MA_HOME}/var/run/sshd.pid\n\
   StrictModes no\n\
   UsePAM no" > ${MA_HOME}/etc/ssh/sshd_config && \
       # generate ssh key
       ssh-keygen -t rsa -f ${MA_HOME}/.ssh/id_rsa -P '' && \
       cat ${MA_HOME}/.ssh/id_rsa.pub >> ${MA_HOME}/.ssh/authorized_keys && \
       # disable ssh host key checking for all hosts
       echo "Host *\n\
     StrictHostKeyChecking no" > ${MA_HOME}/.ssh/config

   For details about how to write a Dockerfile, see official Docker documents.

8. Ensure that the Dockerfile has been created. The following shows the context folder:

   context
   ├── Dockerfile
   ├── Miniconda3-py37_4.12.0-Linux-x86_64.sh
   └── openmpi-3.0.0-bin.tar.gz

9. Create the container image. Run the following command in the directory where the Dockerfile is stored to build the container image mpi:3.0.0-cuda11.1:

   1	docker build . -t mpi:3.0.0-cuda11.1

   The following log shows that the image has been created.

   naming to docker.io/library/mpi:3.0.0-cuda11.1

Step 5 Uploading an Image to SWR

1. Log in to the SWR console and select a region. It must share the same region with ModelArts. Otherwise, the image cannot be selected.
2. Click Create Organization in the upper right corner and enter an organization name to create an organization. Customize the organization name. Replace the organization name deep-learning in subsequent commands with the actual organization name.
3. Click Generate Login Command in the upper right corner to obtain the login command. In this example, the temporary login command is copied.
4. Log in to the local environment as user root and enter the copied temporary login command.
5. Upload the image to SWR.
   1. Tag the uploaded image.

      # Replace the region, domain, as well as organization name deep-learning with the actual values.
      sudo docker tag mpi:3.0.0-cuda11.1 swr.cn-north-4.myhuaweicloud.com/deep-learning/mpi:3.0.0-cuda11.1

   2. Upload the image.

      # Replace the region, domain, as well as organization name deep-learning with the actual values.
      sudo docker push swr.cn-north-4.myhuaweicloud.com/deep-learning/mpi:3.0.0-cuda11.1

6. After the image is uploaded, choose My Images in navigation pane on the left of the SWR console to view the uploaded custom images.

   swr.cn-north-4.myhuaweicloud.com/deep-learning/mpi:3.0.0-cuda11.1 is the SWR URL of the custom image.

Step 6 Creating a Training Job on ModelArts

1. Log in to the ModelArts console and check whether access authorization has been configured for your account. If not, configure the authorization by referring to Configuring Agency Authorization for ModelArts with One Click. For users who used AKs for authorization, you are advised to clear the existing authorization and use an agency.
2. In the navigation pane, choose Model Training > Training Jobs.
3. On the Create Training Job page, configure parameters and click Submit.
   • Algorithm Type: Custom algorithm
   • Boot Mode: Custom image
   • Image path: swr.cn-north-4.myhuaweicloud.com/deep-learning/mpi:3.0.0-cuda11.1
   • Code Directory: OBS path to the boot script, for example, obs://test-modelarts/mpi/demo-code/.
   • Boot Command: bash ${MA_JOB_DIR}/demo-code/run_mpi.sh python ${MA_JOB_DIR}/demo-code/mpi-verification.py
   • Environment Variable: Add MY_SSHD_PORT = 38888.
   • Resource Pool: Public resource pools
   • Compute Nodes: Enter 1 or 2.
   • Persistent Log Saving: enabled
   • Job Log Path: OBS path to stored training logs, for example, obs://test-modelarts/mpi/log/.
4. Confirm the configurations of the training job and click Submit.
5. Wait until the training job is created.

   After you submit the job creation request, the system will automatically perform operations on the backend, such as downloading the container image and code directory and running the boot command. A training job requires a certain period of time for running. The duration ranges from dozens of minutes to several hours, depending on the service logic and selected resources. Figure 1 shows the log information after the training job is executed.

   Figure 1 Run logs of worker-0 with one compute node and GPU specifications
   

   Set Compute Nodes to 2 and run the training job. Figure 2 and Figure 3 show the log information.

   Figure 2 Run logs of worker-0 with two compute nodes and GPU specifications
   
   Figure 3 Run logs of worker-1 with two compute nodes and GPU specifications