Characterizing the parallel performance of a large‐scale, particle‐in‐cell plasma simulation code

Abstract

AbstractThe efficient implementation of particle‐in‐cell (PIC) plasma simulation codes on distributed memory concurrent computers is made difficult by the conflicting aims of balancing the computational load and minimizing interprocessor communication. This paper surveys previous work on PIC plasma simulation codes on advanced architecture computers, identifies the main issues in parallelizing such codes, and discusses different decomposition schemes. For MIMD concurrent computers the adaptive Eulerian (AE) decomposition scheme is attractive because it seeks to maintain approximate load balance dynamically while avoiding excessive non‐local communication. The load balance and communication characteristics of a large‐scale AE/PIC code on the hypercube are investigated by simulating the behavior of the parallel code on a Cray‐2. The results show that for the three‐dimensional problems studied, in which the particle distribution is highly inhomogeneous, the communication of data between the particle and mesh decompositions dominates the communication overhead. Performing the load balancing may also make a significant contribution to the concurrent overhead if the grain size is sufficiently small. The advantages of the simulation approach in investigating the behavior of concurrent large‐scale applications are discussed, together with portability and software engineering issues.