Hybrid Parallelism in Finite Volume Based Algorithms in Application to Two-Dimensional Scattering Problem Setting

• Published in: Computational mathematics and modeling Vol. 31; no. 3; pp. 355 - 363
• Main Authors: Zagidullin, R. R., Smirnov, A. P., Matveev, S. A., Shestopalov, Y. V., Rykovanov, S. G.
• Format: Journal Article
• Language: English
• Published: New York Springer US 2020; Springer Nature B.V
• Subjects: Algorithms; Applications of Mathematics; Central processing units; Computational Mathematics and Numerical Analysis; CPUs; Electrodynamics; Finite volume method; GPU optimization; Graphics processing units; Hybrid parallelism; Mathematical Modeling and Industrial Mathematics; Mathematics; Mathematics and Statistics; Optimization; Parallel processing; Supercomputers; Two dimensional analysis; hybrid parallelism; finite volume method; GPU optimization
• ISSN: 1046-283X; 1573-837X; 1573-837X
• DOI: 10.1007/s10598-020-09496-6

We analyze the parallelism particularities of finite volume based computational algorithms. We use a problem set from electrodynamics and parallelize its solution. It serves as an example and illustration of our findings. The tests are carried out on the Zhores supercomputer. The hardware used to run parallel algorithms includes Nvidia Tesla V100 GPUs and Intel Xeon CPUs. We find that finite volume method discretizations work well with parallelization on GPUs. The speedup can be increased by proper CUDA optimizations. However, it is very resource consuming to perform all computations on supercomputer GPUs if the problem size is exceptionally big. Hence, it is necessary to harness every processing power that a supercomputer can offer including CPUs. We offer ways to parallelize the posed problem in a hybrid fashion.