Collective operations

Message Passing Parallel Programming with MPI

Published in Vivek Kale, Parallel Computing Architectures and APIs, 2019

Collective operations are blocking. Collective communication routines do not take message tag arguments. Collective operations within subsets of processes are accomplished by first partitioning the subsets into new groups and then attaching the new groups to new communicators. Finally, users should work with MPI defined datatypes—not with derived types.

MPI Parallel Implementation for Pseudo-Spectral Simulations for Turbulent Channel Flow

View Article

Journal Information

Published in International Journal of Computational Fluid Dynamics, 2020

Oh-Kyoung Kwon, Jin Lee, Junghoon Lee, Ji-Hoon Kang, Jung-Il Choi

The strong scalability test for the L550 case of grid points indicated that the non-blocking collective operations with 256 nodes using 16,384 MPI ranks provided 15.76 times performance improvement over 8 nodes using 512 MPI ranks. In addition, we evaluated which transposition algorithm is appropriate for the present study. Applying the non-blocking collective operations leads to 1.76 and 3.55 times better computation in terms of performance for the S180 and L550 cases compared to the baseline, respectively. Moreover, a weak scaling test confirmed that the non-blocking collective operations can be scaled up to 512 nodes using a 32,768 MPI ranks. This is because the non-blocking collective operations make performance be improved due to the latency mitigation by overlapping computation and communication.

Collective operations

Explore chapters and articles related to this topic

Message Passing Parallel Programming with MPI

MPI Parallel Implementation for Pseudo-Spectral Simulations for Turbulent Channel Flow