{Seminar} @ CDS: 13th January : “An Incompressible CFD Solver for Multi-GPU Platforms”

The Graphical Processing Units (GPUs) help in considerably reducing the computational time with reduced energy costs. GPUs provide high number of FLOPS and they potentially offer the most efficient architecture to carry out large-scale calculations for scientific problems. However, in order to get desired scalability in GPU system, programmer needs to efficiently utilize different on-chip and off-chip memory components. Furthermore, execution efficiency depends on issues like choosing optimum block size, eliminating warp divergence etc. Therefore, efficient CUDA programming require an in-depth knowledge of GPU hardware, whereas programming standards like OpenACC can help in obtaining optimized parallelization of legacy codes through incremental changes. I will discuss the steps for optimization of an incompressible computational fluid dynamics (CFD) solver for multi-GPU platforms using OpenACC. An immersed boundary technique will be demonstrated with fully parallelized search algorithm in connection with flow over moving and deformable bodies. Application problems in biofluidics and micro-flyers will be shown in which around 100 times speed-up is obtained through GPU acceleration.

Host Faculty: Prof. Sathish Vadhiyar