TransferBench

TransferBench is a simple utility capable of benchmarking simultaneous copies between user-specified CPU and GPU devices.

Documentation for TransferBench is available at https://rocm.docs.amd.com/projects/TransferBench/en/latest/index.html.

Requirements

• You must have a ROCm stack installed on your system (HIP runtime)
• You must have libnuma installed on your system
• AMD IOMMU must be enabled and set to passthrough for AMD Instinct cards

Documentation

To build documentation locally, use the following code:

cd docs

pip3 install -r .sphinx/requirements.txt

python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html

Building TransferBench

You can build TransferBench using Makefile or CMake.

• Makefile:

  make

• CMake:

  mkdir build
  cd build
  CXX=/opt/rocm/bin/hipcc cmake ..
  make

  If ROCm is not installed in /opt/rocm/, you must set ROCM_PATH to the correct location.

NVIDIA platform support

You can build TransferBench to run on NVIDIA platforms via HIP or native NVCC.

Use the following code to build with HIP for NVIDIA (note that you must have a HIP-compatible CUDA version installed, e.g., CUDA 11.5):

CUDA_PATH=<path_to_CUDA> HIP_PLATFORM=nvidia make`

Use the following code to build with native NVCC (builds TransferBenchCuda):

make

Things to note

• Running TransferBench with no arguments displays usage instructions and detected topology information

• You can use several preset configurations instead of a configuration file:

  • a2a : All-to-all benchmark test
  • cmdline: Take in Transfers to run from command-line instead of via file
  • p2p : Peer-to-peer benchmark test
  • pcopy : Benchmark parallel copies from a single GPU to other GPUs
  • rsweep : Random sweep across possible sets of transfers
  • rwrite : Benchmarks parallel remote writes from a single GPU to other GPUs
  • scaling: GPU subexecutor scaling tests
  • schmoo : Local/Remote read/write/copy between two GPUs
  • sweep : Sweep across possible sets of transfers

• When using the same GPU executor in multiple simultaneous transfers on separate streams (USE_SINGLE_STREAM=0), performance may be serialized due to the maximum number of hardware queues available

  • The number of maximum hardware queues can be adjusted via GPU_MAX_HW_QUEUES