Using MPI's Non-Blocking Allreduce for Health Checks in Dynamic LBM Simulations

Abstract

In this work, we add a MPI_Iallreduce in dynamic, distributed simulations utilizing our explicit time integration solvers in the APES simulation framework to replace its blocking variant for healthchecks during run time. The use of a non-blocking reduction operation allows for less strict synchronization between the computing processes in the distributed computation by shifting the communication to the back-ground and thereby enables a more robust scalability of the applicationinto the region where only little computation per process remains.