Performance modeling of the finite-difference time-domain method on parallel systems

As high-performance parallel codes are developed or ported to new architectures, it is often difficult to quantify the the causes of performance problems. Models of program performance can provide users with insight into the effect of system and program parameters on performance, can help programmers tune applications, and can help programmers make decisions about processor allocation. This paper introduces a modeling technique applied to the Finite-Difference Time-Domain (FDTD) algorithm. The technique models the performance of an existing application in terms of the size of the problem being solved and the number of processors. The models show that for sufficiently large problem sizes the algorithm performs well. However, for smaller problem sizes or when too many processors are used, the models show that parallel overheads become significant.