Installing TransferBench

Installing TransferBench#

This topic describes how to build TransferBench.

Prerequisite#

  • Install ROCm stack on the system to obtain HIP runtime

  • Install libnuma on the system

  • Enable AMD IOMMU and set to passthrough for AMD Instinct cards

Building TransferBench#

Here are the steps to build TransferBench:

  1. Download the latest version of TransferBench from the git repository.

    git clone https://github.com/ROCm/TransferBench.git
cd TransferBench

  2. Build TransferBench using Makefile or CMake.

    To build using Makefile, use:

    make

    To build using CMake, use:

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

Note

If ROCm is installed in a folder other than /opt/rocm/, set ROCM_PATH appropriately.

NIC executor support will be enabled if IBVerbs is detected and if infiniband/verbs.h is found in the default include path. NIC executor support can be disabled explicitly by setting DISABLE_NIC_EXEC=1

MPI support will be enabled if mpi.h is found in MPI_PATH/include/ MPI executor support can be disabled explicitly by setting DISABLE_MPI_COMM=1

Building documentation#

To build documentation locally, use:

cd docs
pip3 install -r ./sphinx/requirements.txt
python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html

NVIDIA platform support#

You can build TransferBench to run on NVIDIA platforms using native NVIDIA CUDA Compiler Driver (NVCC).

To build with native NVCC, use:

make

TransferBench looks for NVCC in /usr/local/cuda by default. To modify the location of NVCC, use environment variable CUDA_PATH:

CUDA_PATH=/usr/local/cuda make
Contents