dmabuf Support¶

oneCCL provides experimental support for data transfers between Intel GPU memory and NIC using Linux dmabuf, which is exposed through OFI API for verbs provider.

Requirements¶

Linux kernel version >= 5.12
RDMA core version >= 34.0
level-zero-devel package

Usage¶

oneCCL, OFI and OFI/verbs from Intel® oneAPI Base Toolkit support device memory transfers. Refer to Run instructions for usage.

If you want to build software components from sources, refer to Build instructions.

Build instructions¶

OFI¶

git clone --single-branch --branch v1.13.2 https://github.com/ofiwg/libfabric.git
cd libfabric
./autogen.sh
./configure --prefix=<ofi_install_dir> --enable-verbs=<rdma_core_install_dir> --with-ze=<level_zero_install_dir> --enable-ze-dlopen=yes
make -j install

Note

You may also get OFI release package directly from here. No need to run autogen.sh if using the release package.

oneCCL¶

cmake -DCMAKE_INSTALL_PREFIX=<ccl_install_dir> -DLIBFABRIC_DIR=<ofi_install_dir> -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=dpcpp -DCOMPUTE_BACKEND=dpcpp -DENABLE_OFI_HMEM=1 ..
make -j install

Run instructions¶

Set the environment.

If Intel® oneAPI Base Toolkit is used:

source <toolkit_install_dir>/setvars.sh

If software components are built from sources:

source <ccl_install_dir>/env/setvars.sh
export LD_LIBRARY_PATH=<ofi_install_path>/lib:${LD_LIBRARY_PATH}

Run allreduce test with ring algorithm and SYCL USM device buffers:

export CCL_ATL_TRANSPORT=ofi
export CCL_ATL_HMEM=1
export CCL_ALLREDUCE=ring
export FI_PROVIDER=verbs
mpiexec -n 2 <ccl_install_dir>/examples/sycl/sycl_allreduce_usm_test gpu device

oneCCL documentation