First released in September 2018, 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.
Improvements included in this release:
- Host-transfer implementations of standard collectives in the
MPI-CUDAbackend - Experimental RMA Put/Get operations
- Improved algorithm specification, point-to-point operations, testing, and benchmarks
Learn more:
