oneAPI Collective Communications Library (oneCCL)

oneAPI Collective Communications Library (oneCCL) provides an efficient implementation of communication patterns used in deep learning.

oneCCL is integrated into:

• Horovod* (distributed training framework). Refer to Horovod with oneCCL for details.
• PyTorch* (machine learning framework). Refer to PyTorch bindings for oneCCL for details.

Release notes available by link.

Prerequisites

Ubuntu* 18 GNU*: C, C++ 4.8.5 or higher.

SYCL support

Intel(R) oneAPI DPC++ Compiler with L0 v1.0 support

Installation

General installation scenario

cd oneccl
mkdir build
cd build
cmake ..
make -j install

If a "clear" build is needed, then one should create a new build directory and invoke cmake within it.

Specify installation directory

Modify cmake command as follow:

cmake .. -DCMAKE_INSTALL_PREFIX=/path/to/installation/directory

If no -DCMAKE_INSTALL_PREFIX is specified then ccl will be installed into _install subdirectory of the current build directory, e.g. ccl/build/_install

Specify compiler

Modify cmake command as follow:

cmake .. -DCMAKE_C_COMPILER=your_c_compiler -DCMAKE_CXX_COMPILER=your_cxx_compiler

Specify SYCL cross-platform abstraction level

If your CXX compiler requires SYCL, it is possible to specify it (CodePlay ComputeCpp and DPC++ are available for now). Modify cmake command as follows:

cmake .. -DCMAKE_C_COMPILER=your_c_compiler -DCMAKE_CXX_COMPILER=compute++ -DCOMPUTE_RUNTIME=computecpp
cmake .. -DCMAKE_C_COMPILER=your_c_compiler -DCMAKE_CXX_COMPILER=dpcpp -DCOMPUTE_RUNTIME=dpcpp

OpenCL search location path hint can be specified by using standard environment OPENCLROOT additionally:

OPENCLROOT=your_opencl_location cmake .. -DCMAKE_C_COMPILER=your_c_compiler -DCMAKE_CXX_COMPILER=compute++ -DCOMPUTE_RUNTIME=computecpp

Specify build type

Modify cmake command as follow:

cmake .. -DCMAKE_BUILD_TYPE=[Debug|Release|RelWithDebInfo|MinSizeRel]

Enable make verbose output

Modify make command as follow to see all parameters used by make during compilation and linkage:

make -j VERBOSE=1

Build with address sanitizer

Modify cmake command as follow:

cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_ASAN=true

Note: address sanitizer only works in Debug build

Make sure that libasan.so exists.

Usage

Launching Example Application

Use the command:

$ source <install_dir>/env/setvars.sh
$ cd <install_dir>/examples
$ mpirun -n 2 ./benchmark/benchmark

Setting workers affinity

There are two ways to set workers threads affinity - explicit and automatic

Setting affinity explicitly

1. Set environment variable CCL_WORKER_COUNT with desired number of workers threads
2. Set environment variable CCL_WORKER_AFFINITY with IDs of cores to be bound to

Example:

export CCL_WORKER_COUNT=4
export CCL_WORKER_AFFINITY=3,4,5,6

With variables above CCL will create 4 threads and pin them to cores with numbers 3,4,5 and 6 accordingly

Setting affinity automatically

NOTE: automatic pinning only works if application has been launched using mpirun provided by CCL distribution package.

1. Set environment variable CCL_WORKER_COUNT with desired number of workers threads
2. Set environment variable CCL_WORKER_AFFINITY with value auto

Example:

export CCL_WORKER_COUNT=4
export CCL_WORKER_AFFINITY=auto

With variables above CCL will create 4 threads and pin them to the last 4 cores available for the launched process.

The exact IDs of CPU cores depend on parameters passed to mpirun

FAQ

When do I need a clean build? When should I remove my favorite build directory?

In the most cases there is no need in removal of the current build directory. Just run make to compile and link changed files. Only if one sees some suspicious build errors after significant change in the code (e.g. after rebase or change of branch) then it is a hint to clean build directory.