One copy instead of 16. Utilize multiple endpoints. Saturate the fabric. Reduce memory usage per MPI rank. Multi-thread MPI performance comparable to single threaded options. How can you achieve this parallelism with your code and more? This short webinar delivers the goods.
Learn how the Intel® MPI Library implementation using multiple endpoints (Multi-EP) enables a single multi-threaded rank per node, and more than one MPI thread per rank, requiring fewer MPI ranks. Multi-EP is especially effective for hybrid MPI + OpenMP* threaded programs, enabling multiple threads to be active in the MPI library at the same time without requiring extra synchronization. This allows a single MPI rank using multiple threads to saturate the network bandwidth, namely Intel® Omni-Path Architecture, eliminating the need for multiple ranks per node. Watch now to learn:
- How a Hybrid MPI + OpenMP program with 16 cores can be laid out in 5 different ways.
- Three ways to do MPI_reduce
- MPI shared memory and subcommunicators strategies
- How Multi-EP allows multiple threads per rank to do MPI communications, allowing the user to choose the optimal number of MPI ranks/node (as few as one).
Download Intel MPI Library from this page, and click here for more information about Multi-EP.