A Study of FDS Computational Performance in Heterogeneous Hardware Architectures -Applied for grassland fires

Abstract

Fire Dynamic Simulator (FDS), a fire simulation program, applies Message Passing Interface (MPI) and Open Multi Processing (OpenMP) libraries for large-scale simulation. FDS can be executed by dividing simulation problems in a computing cluster using MPI. The main point is to divide the entire domain to be interpreted into several sub-domains and allow each sub-domain to be calculated by an individual computer with an individual processor. When performing parallel computation, FDS first decomposes each sub-domain, then supports two-step parallelization in which multi-threading is applied within each sub-domain, and uses the OpenMP library to implement multi-threading. In this study, OpenACC, a parallelization technique capable of using heterogeneous hardware architectures, was partially applied to FDS. As an application problem, the calculation performance is evaluated through CSIRO Grassland Fires, a verification case of FDS. The hardware for evaluation was a personal computer consisting of dual Xeon 2678-V3 and GeForce GTX 1070. The FDS source code applies OpenACC using PGI Fortran as a compiler in Linux environments. In calculation performance, calculations using CPU and GPU together show 1.89 times faster performance than calculations using a single CPU. In case of using 1 GPU and 16 CPUs (MPP + OpenACC), the analysis result is 21 times faster. In this regard, analysis of grassland fire of WFDS was performed.