Improvement and performance analysis of MPI_ALLGATHER algorithm
Qing Jiao · Computer Engineering and Applications Journal · 2007
We discuss issues related to the high performance implementation of collective communication operations in MPICH for clusters in this paper.The Ring and Neighbor Exchange algorithms have the best communication locality property,and they have more communication steps than others.The Recursive Doubling and Bruck algorithms have less communication steps,and they have worse communication locality property than others.Based on the discussion,we present a new MPI_ALLGATHER algorithm that combines the Neighbor Exchange and Recursive Doubling algorithms.The new algorithm has less data transfer times than the Neighbor Exchange algorithm and better communication locality property than the Recursive Doubling algorithm.We evaluate the new algorithm in Myrinet cluster and find that the new algorithm performs better than the Neighbor Exchange algorithm and sometimes performs the best for medium-size messages among all the algorithms.The new algorithm also performs better for long-size messages than the Recursive Doubling and Bruck algorithms.