Multi-GPU/CPU deflated preconditioned conjugate gradient for bubbly flow solver
Rohit Gupta, Martin B. van Gijzen, C. Vuik · High Performance Computing Symposium · 2014
We present a Multi-GPU/CPU implementation of Deflated Preconditioned Conjugate Gradient (DPCG) to solve a highly ill-conditioned linear system arising from the discretized Pressure-correction equation on GPUs and CPUs. We discuss the challenges and choices in such an implementation with respect to communication and data layout. We present results of our implementation for systems having up to 16 million unknowns. Across 8 GPUs (on distinct nodes connected via MPI) our code achieves atleast 2 times speedup compared to 32 cores (across 4 distinct nodes connected via MPI). Comparing with 64 CPU cores across 8 nodes the same GPU version proves to be comparable in terms of wall-clock time.