HPMLA-CPU-OpenMPI

HPMLA-CPU-OpenMPI

This recipe shows how to run High Performance ML Algorithms (HPMLA) on CPUs across Azure VMs via Open MPI.

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 should be a CPU-only instance, for example, STANDARD_D2_V2.
• 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 HPMLA Docker image that can be run with OpenMPI. The image denoted with 0.0.1 tag found in msmadl/symsgd:0.0.1 is compatible with Azure Batch Shipyard VMs.

MPI Jobs Configuration (MultiNode)

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

• docker_image should be the name of the Docker image for this container invocation. For this example, this should be msmadl/symsgd:0.0.1. 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 HPMLA training example with the msmadl/symsgd:0.0.1 Docker image. The application command to run would be: "/parasail/run_parasail.sh -w /parasail/supersgd -l 1e-4 -k 32 -m 1e-2 -e 10 -r 10 -f $AZ_BATCH_NODE_SHARED_DIR/azblob/<container_name from the data shredding configuration file> -t 1 -g 1 -d $AZ_BATCH_TASK_WORKING_DIR/models -b $AZ_BATCH_NODE_SHARED_DIR/azblob/<container_name from the data shredding configuration file>"

  • run_parasail.sh has these parameters
    • -w the HPMLA superSGD directory
    • -l learning rate
    • -k approximation rank constant
    • -m model combiner convergence threshold
    • -e total epochs
    • -r rounds per epoch
    • -f training file prefix
    • -t number of threads
    • -g log global models every this many epochs
    • -d log global models to this directory at the host"
    • -b location for the algorithm's binary"

• The training data will need to be shredded to match the number of VMs and the thread's count per VM, and then deployed to a mounted Azure blob that the VM docker images have read/write access. A basic python script that can be used to shred and deploy the training data to a blob container, and other data shredding files can be found here.

• shared_data_volumes should contain the shared data volume with an azureblob volume driver as specified in the global configuration file found here.

• multi_instance property must be defined

  • num_instances should be set to pool_current_dedicated, or pool_current_low_priority

Dockerfile and supplementary files

Supplementary files can be found here.

You must agree to the following licenses prior to use:

• High Performance ML Algorithms License
• TPN Ubuntu Container
• Microsoft Third Party Notice