MLPerf Workload Profiles

The following profiles run customer-representative or benchmarking scenarios using the MLPerf workload.

• Workload Details
• MLPerf Training Bert Preprocessing Data

PERF-GPU-MLPERF.json

Runs the MLPerf benchmark workload to test GPU performance.

warning

This workload is supported ONLY for systems that contain Nvidia GPU hardware components. See the documentation above for more specifics.

• Workload Profile

• Supported Platform/Architectures

  • linux-x64

• Supports Disconnected Scenarios

  • No. Internet connection required.

• Dependencies
  The dependencies defined in the 'Dependencies' section of the profile itself are required in order to run the workload operations effectively.

  • Internet connection.
  • The VM must run on hardware containing Nvidia GPU cards/components.

  Additional information on components that exist within the 'Dependencies' section of the profile can be found in the following locations:

  • Installing Dependencies

• Profile Parameters
  The following parameters can be optionally supplied on the command line to modify the behaviors of the workload.

  Parameter	Purpose	Default Value
  Username	Optional. User which needs to be created in container to run MLPerf benchmarks.	testuser
  DiskFilter	Optional. Filter to decide the disk that will be used to download benchmarks. Since benchmarks data is around 800gb, default disk filter is greater than 1000gb.	SizeGreaterThan:1000gb
  LinuxCudaVersion	Optional. Version of CUDA that needs to be installed.	12.4
  LinuxDriverVersion	Optional. Version of GPU driver that needs to be installed.	550
  LinuxLocalRunFile	Optional. Link to download specified CUDA and GPU driver versions.	https://developer.download.nvidia.com/compute/cuda/12.0.0/local_installers/cuda_12.4.0_550.54.14_linux.run
  RequireCustomSystemSupport	Optional. This enables additional A100_PCIe_40GBx8 system support that was not supported by github repo of MLPerf. Ones that are supported by github repo of MLPerf are still supported.	true

• Profile Runtimes
  See the 'Metadata' section of the profile for estimated runtimes. These timings represent the length of time required to run a single round of profile actions. These timings can be used to determine minimum required runtimes for the Virtual Client in order to get results. These are often estimates based on the number of system cores.

• Usage Examples
  The following section provides a few basic examples of how to use the workload profile.

  # Execute the workload profile
  VirtualClient.exe --profile=PERF-GPU-MLPERF.json --system=Demo --timeout=1440 --packageStore="{BlobConnectionString|SAS Uri}"

PERF-GPU-MLPERF-TRAINING-NVIDIA.json

Runs the MLPerf benchmark workload to test GPU performance.

warning

This workload is supported ONLY for systems that contain Nvidia GPU hardware components. See the documentation above for more specifics.

• Workload Profile

• Supported Platform/Architectures

  • linux-x64

• Supports Disconnected Scenarios

  • No. Internet connection required.

• Storage Requirements

  • Min OS Disk Size : 256GB
  • Min Data Disk Size : 8TB (1 Disk)

• Dependencies
  The dependencies defined in the 'Dependencies' section of the profile itself are required in order to run the workload operations effectively.

  • Internet connection.
  • The VM must run on hardware containing Nvidia GPU cards/components.

  Additional information on components that exist within the 'Dependencies' section of the profile can be found in the following locations:

  • Installing Dependencies

• Profile Parameters
  The following parameters can be optionally supplied on the command line to modify the behaviors of the workload.

  Parameter	Purpose	Default Value
  Username	Optional. User which needs to be created in container to run MLPerf benchmarks.
  DiskFilter	Optional. Filter to decide the disk that will be used to download training data.	BiggestSize
  LinuxCudaVersion	Optional. Version of CUDA that needs to be installed.	12.0
  LinuxDriverVersion	Optional. Version of GPU driver that needs to be installed.	525
  LinuxLocalRunFile	Optional. Link to download specified CUDA and GPU driver versions.	https://developer.download.nvidia.com/compute/cuda/12.0.0/local_installers/cuda_12.0.0_525.60.13_linux.run

• MLPerfTrainingExecutor Parameters

  Parameter	Purpose	Default Value
  Model	Optional. The type of model to run for training supported models mentioned in mlperf.md	bert
  Username	Optional. Username for running the docker. On null it selects logged in user.	null
  BatchSize	Optional. BatchSize for the datachunks in training model.	40
  Implementation	Optional. Implementation present for a given model/benchmark. Example for bert link	pytorch-22.09
  ContainerName	Optional. Name for docker model.	language_model
  DataPath	Optional. Folder name for training data. /mlperftraining0/{DataPath}	mlperf-training-data-bert.1.0.0
  GPUNum	Optional. Number of GPUs to stress.	8
  ConfigFile	Optional. Configuration for running workload. Visit the implementation for a model for all supported config files. link.	config_DGXA100_1x8x56x1.sh
  PackageName	Required. Packname for mlperf training.

• MLPERF Trainging Changes

  • Changed "nvidia-docker" to "docker" inside run_with_docker.sh file.

• Profile Runtimes
  See the 'Metadata' section of the profile for estimated runtimes. These timings represent the length of time required to run a single round of profile actions. These timings can be used to determine minimum required runtimes for the Virtual Client in order to get results. These are often estimates based on the number of system cores.

• Usage Examples
  The following section provides a few basic examples of how to use the workload profile.

  # Execute the workload profile
  sudo ./VirtualClient --profile=PERF-GPU-MLPERF-TRAINING.json --system=Demo --timeout=1440 --packageStore="{BlobConnectionString|SAS Uri}"