batch-shipyard/recipes/CNTK-GPU-OpenMPI

CNTK-GPU-OpenMPI

This recipe shows how to run CNTK on GPUs using N-series Azure VM instances in an Azure Batch compute pool.

Please note that CNTK currently uses MPI even for multiple GPUs on a single node.

Configuration

Please see refer to this set of sample configuration files for this recipe.

Pool Configuration

The pool configuration should enable the following properties:

• vm_size must be one of STANDARD_NC6, STANDARD_NC12, STANDARD_NC24, STANDARD_NV6, STANDARD_NV12, STANDARD_NV24. NC VM instances feature K80 GPUs for GPU compute acceleration while NV VM instances feature M60 GPUs for visualization workloads. Because CNTK is a GPU-accelerated compute application, it is best to choose NC VM instances.
• vm_configuration is the VM configuration
  • platform_image specifies to use a platform image
    • publisher should be Canonical or OpenLogic.
    • offer should be UbuntuServer for Canonical or CentOS for OpenLogic.
    • sku should be 16.04-LTS for Ubuntu or 7.3 for CentOS.
• inter_node_communication_enabled must be set to true
• max_tasks_per_node must be set to 1 or omitted

Global Configuration

The global configuration should set the following properties:

• docker_images array must have a reference to a valid CNTK GPU-enabled Docker image. For singlenode (non-MPI) jobs, you can use the official Microsoft CNTK Docker images. For MPI jobs, you will need to use Batch Shipyard compatible Docker images which can be found in the alfpark/cntk repository. Images denoted with refdata tag suffixes found in can be used for this recipe which contains reference data for MNIST and CIFAR-10 examples. If you do not need this reference data then you can use the images without the refdata suffix on the image tag.

Non-MPI Jobs Configuration (SingleNode+SingleGPU)

The jobs configuration should set the following properties within the tasks array which should have a task definition containing:

• image should be the name of the Docker image for this container invocation, e.g., microsoft/cntk:2.1-gpu-python3.5-cuda8.0-cudnn6.0
• command should contain the command to pass to the Docker run invocation. For the microsoft/cntk:2.1-gpu-python3.5-cuda8.0-cudnn6.0 Docker image, and to run the MNIST convolutional example on a single CPU, the command would be: "/bin/bash -c \"source /cntk/activate-cntk && cd /cntk/Examples/Image/DataSets/MNIST && python -u install_mnist.py && cd /cntk/Examples/Image/Classification/ConvNet/Python && python -u ConvNet_MNIST.py\""
• gpu must be set to true. This enables invoking the nvidia-docker wrapper.

MPI Jobs Configuration (SingleNode+MultiGPU, MultiNode+SingleGPU, MultiNode+MultiGPU)

The jobs configuration should set the following properties within the tasks array which should have a task definition containing:

• image should be the name of the Docker image for this container invocation. For this example, this can be alfpark/cntk:2.1-gpu-1bitsgd-py35-cuda8-cudnn6-refdata. Please note that the docker_images in the Global Configuration should match this image name.
• command should contain the command to pass to the Docker run invocation. For this example, we will run the ResNet-20 Distributed training on CIFAR-10 example in the alfpark/cntk:2.1-gpu-1bitsgd-py35-cuda8-cudnn6-refdata Docker image. The application command to run would be: "/cntk/run_cntk.sh -s /cntk/Examples/Image/Classification/ResNet/Python/TrainResNet_CIFAR10_Distributed.py -- --network resnet20 -q 1 -a 0 --datadir /cntk/Examples/Image/DataSets/CIFAR-10 --outputdir $AZ_BATCH_TASK_WORKING_DIR/output"
  • run_cntk.sh has two parameters
    • -s for the Python script to run
    • -w for the working directory (not required for this example to run)
    • -- parameters specified after this are given verbatim to the Python script
• gpu must be set to true. This enables invoking the nvidia-docker wrapper.
• multi_instance property must be defined for multinode executions
  • num_instances should be set to pool_specification_vm_count_dedicated, pool_specification_vm_count_low_priority, pool_current_dedicated, or pool_current_low_priority
  • coordination_command should be unset or null
  • resource_files should be unset or the array can be empty

Dockerfile and supplementary files

The Dockerfile for the Docker image can be found here.

You must agree to the following licenses prior to use:

• CNTK License
• CNTK 1-bit SGD License