-
Notifications
You must be signed in to change notification settings - Fork 50
Benchmark Grid Gather Scatter performance
Sam Reeve edited this page Nov 2, 2023
·
2 revisions
The plots below show performance of grid halo on the ORNL Frontier supercomputer. Both CPU and GPU performance are compared as a function of total grid points per rank (with eight MPI ranks).
Creation refers to construction of the Halo
object (building the communication steering vectors) and gather
and scatter
refer to executing the communication (copying ghosted grid points on neighboring ranks). Each point represents a single integer halo width. The lowest, outlier lines represent the smallest halo width (=1).
Default parameters with the commandline "large" setting were used for these results.
Cabana - A Co-Designed Library for Exascale Particle Simulations