A Performance Model for GPU-Accelerated FDTD Applications

In this work we develop, validate and use a performance model for a Finite-Difference Time-Domain (FDTD) application which is parallelized on multiple GPUs. FDTD is a method for simulating electrodynamic interaction and is applied in a number of research and engineering areas. In this work we focus on a particular implementation called B-CALM (Belgium-California Light Machine). We adopt a simple, semi-empirical modelling approach to design a model which we validate for different hardware architectures. Using the model allows making implementation decisions and exploring the architectural design space with the goal of optimizing HPC systems for this application.