Implementation and evaluation of a communication avoiding Krylov subspace method, CA-GMRES, on HA-PACS/TCA

Matsumoto, Kazuya*; Idomura, Yasuhiro   ; Ina, Takuya*; Mayumi, Akie ; Yamada, Susumu  

Communication avoiding (CA) Krylov methods are promising solutions for communication bottlenecks on supercomputers based on many core processors or accelerators. In this work, we implemented the CA-GMRES method on a GPU cluster, the HA-PACS, and evaluated its performance on a non-symmetric matrix solver from a nuclear CFD code. The result shows that the CA-GMRES method is significantly faster than the conventional Krylov methods such as the GMRES method and the GCR method.