Towards a finite volume model on a many-core platform
Pablo Igounet, Pablo Alfaro, Gabriel Usera, Pablo Ezzatti · International Journal of High Performance Systems Architecture · 2012
This article presents a study of porting the finite volume model, caffa3d.MB, to GPU in order to improve its computational efficiency. In particular, we evaluated all the stages of the aforementioned model and implemented on GPU the most costly one, the linear systems resolution by means of the strongly implicit procedure (SIP) method. The experimental evaluation of the proposed implementation of the solver demonstrates that a significant runtime reduction can be attained (acceleration values up to 18×) when compared with a CPU version, and this improvement significantly reduces the total runtime of the model. This results evidence a promising prospect for a full GPU-based implementation of finite volume flow solvers like caffa3d.MB.