Parallel Matrix-Matrix Multiplication Based on HPL with a GPU-Accelerated PC Cluster
Authors:
- Qin WangQWQin Wang
- Junichi OhmuraJOJunichi Ohmura
- Shan AxidaSAShan Axida
- Takefumi MiyoshiTMTakefumi Miyoshi
- Hidetsugu IrieHIHidetsugu Irie
- Tsutomu YoshinagaTYTsutomu Yoshinaga
- 1 November 2010
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 243-248
Abstract
In this paper, we propose an approach for significantly improving the performance of parallel matrix-matrix multiplication using a GPU-accelerated cluster. For one node, we implement a CPUs-GPU parallel double-precision general matrix-matrix multiplication (dgemm) operation and achieve a performance improvement of 32% as compared to the GPU-only case and 56% as compared to the CPUs-only case. For the entire cluster, we use the overlap GPU acceleration solution to high-performance Linpack (HPL), which eliminates the close dependency between the LU decomposition and the dgemm operation, and achieve a performance improvement of 5.72% as compared to the flat GPU acceleration case.Keywords
This publication has 6 references indexed in Scilit:
- MPI Collectives on Modern Multicore Clusters: Performance Optimizations and Communication Characteristics2008
- Parallel Processing of Matrix Multiplication in a CPU and GPU Heterogeneous Environment2007
- Improving Linpack Performance on SMP Clusters with Asynchronous MPI ProgrammingIPSJ Digital Courier, 2006
- Sparse matrix solvers on the GPUACM Transactions on Graphics, 2003
- Building a high-performance collective communication library2002
- Exploiting fast matrix multiplication within the level 3 BLASACM Transactions on Mathematical Software, 1990