Strategies to parallelize a finite element mesh truncation technique on multi-core and many-core architectures

Abstract Achieving maximum parallel performance on multi-core CPUs and many-core GPUs is a challenging task depending multiple factors. These include, for example, the number granularity of computations or use memories devices. In this paper, we assess those factors by evaluating comparing different parallelizations same problem multiprocessor containing CPU with 40 cores four P100 Pascal architecture. We use, as study case, convolutional operation behind non-standard finite element mesh truncation technique in context open region electromagnetic wave propagation problems. A total six algorithms implemented using OpenMP CUDA have been used to carry out comparison leveraging levels parallelism both types platforms. Three are presented first time including multi-GPU method, two others improved versions previously developed some authors. This paper presents thorough experimental evaluation radar cross-sectional prediction problem. Results show that obtained GPU clearly overcomes CPU, much more so if distribute data computations. Accelerations close 30 while version accelerations larger than 250 achieved.