Aluminum provides a generic interface to high-performance communication libraries, with a focus on allreduce algorithms. Blocking and non-blocking algorithms and GPU-aware algorithms are supported. Aluminum also contains custom implementations of select algorithms to optimize for certain situations.

• Blocking and non-blocking algorithms
• GPU-aware algorithms
• Implementations/interfaces:
  • MPI: MPI and custom algorithms implemented on top of MPI
  • NCCL: Interface to Nvidia's NCCL 2 library
  • MPI-CUDA: Custom GPU-aware algorithms

• A compiler that supports at least C++11
• MPI (at least MPI 3.0)
• CUDA (optional if no GPU support is needed)
• NCCL2 (optional if no NCCL support is needed)

CMake 3.9 or newer is required. An out-of-source build is required:

mkdir build && cd build
cmake <options> /path/to/aluminum/source

The required packages are MPI, OpenMP, and HWLOC. MPI and OpenMP use the standard CMake packages and can be manipulated in the standard way. HWLOC, if installed in a nonstandard location, may require HWLOC_DIR to be set to the appropriate installation prefix.

The CUDA-based backends assume CUDA is a first-class language in CMake. An alternative CUDA compiler can be selected using

-DCMAKE_CUDA_COMPILER=/path/to/my/nvcc

If the NCCL backend is used, the NCCL_DIR variable may be used to point CMake to a nonstandard installation prefix.

For the NCCL backend:

-DALUMINUM_ENABLE_NCCL=ON

For the MPI-CUDA backend:

-DALUMINUM_ENABLE_MPI_CUDA=ON

The NCCL and MPI-CUDA backends can be combined.

Here is a complete example:

CMAKE_PREFIX_PATH=/path/to/your/MPI:$CMAKE_PREFIX_PATH cmake -D ALUMINUM_ENABLE_NCCL=YES -D ALUMINUM_ENABLE_MPI_CUDA=YES -D NCCL_DIR=/path/to/NCCL ..

The test_correctness binary will check the correctness of Aluminum's allreduce implementations. The usage is

test_correctness [Al backend: MPI, NCCL, MPI-CUDA]

For example, to test the MPI backend:

mpirun -n 128 ./test_correctness

To test the NCCL backend, instead:

mpirun -n 128 ./test_correctness NCCL

The benchmark_allreduce benchmark can be run similarly, and will report runtimes for different allreduce algorithms.

Coming soon...

• Nikoli Dryden
• Naoya Maruyama
• Andy Yoo
• Tom Benson

See also contributors.

Aluminum is licensed under the Apache License 2.0. See LICENSE for details.