Dynamic Load Balancing of Plasma and Flow Simulations

Abstract

Extracting performance from simulations with complex information dependencies on massively parallel computers requires the computational work to be evenly distributed across the processing resources while maintaining low communication costs. Plasma simulations using a particle-in-cell method and computational fluid dynamics using unstructured mesh-based finite element and volume methods present three distinct distribution requirements. To meet these needs, we present EnGPar’s diffusive partition improvement method. An initial demonstration of EnGPar’s particle distribution improvement is provided along with fluid dynamics mesh partition improvement results on up to 512Ki processes on an IBM BlueGene/Q.