Algorithms for MPI all-to-all exchange using shared memory

Proposed algorithms for implementing MPI_Alltoall and MPI_Alltoallv collective operations using shared memory of multiprocessor servers and the Linux kernel CMA mechanism. The algorithms use either a queue system or CMA system calls for message exchange. The software implementation is based on the Open MPI library as a separate component coll/sharm. In experiments conducted on a server with x86-64 architecture, the MPI_Alltoall operation achieved the greatest time reduction by 1.48 times and MPI_Alltoallv by 1.37 times compared to implementations in the coll/tuned and coll/ucc components of the Open MPI library. Recommendations for using the algorithms for different message sizes are proposed