Shared Memory-based Algorithms for MPI Irregular Collective Operations
Alexey Romanyuta, Mikhail G. Kurnosov · 2022 IEEE International Multi-Conference on Engineering, Computer and Information Sciences (SIBIRCON) · 2022
Algorithms for MPI_Scatterv, MPI_Gatherv and MPI_Allgatherv collective operations using the shared memory of SMP/NUMA machine are proposed. The algorithms use an approach based on copying message fragments through queues in a shared memory segment (Copy-In-Copy-Out). The implementation is based on Open MPI and on average reduces the execution time by 20-40% for the MPI_Scatterv, MPI_Gatherv, MPI_Allgatherv operations compared to the coll/tuned component of the Open MPI library.