The International Conference for High Performance Computing, Networking, Storage and Analysis
Speedup and Numerical Evaluation of Multiple-Precision Krylov Subspace Method Using GPU Cluster for Large-Sparse Linear System.
Authors: Yuta Hirokawa (Tokyo University of Technology), Taku Itoh (Tokyo University of Technology), Hiroto Tadano (University of Tsukuba), Soichiro Ikuno (Tokyo University of Technology)
Abstract: Implementing multiple-precision Krylov subspace method on GPU cluster for a large-sparse linear system is investigated, and the method is numerically evaluated. It is well known that a number of iteration of Krylov subspace method depend on accumulation errors, and the error may affect the calculation results. In order to settle these issues, the multiple precision operation Krylov subspace method is implemented on GPU cluster using GNU Multiple Precision Arithmetic Library (GMP) and CUDA Multiple Precision Arithmetic Library (CUMP), and the method is parallelized to get high performance. The result of computation shows that the variable preconditioned Bi-CGSTAB on GPU cluster is up to 16.38 times faster than that of CPU with OpenMP.