Vector and Parallel Tuning of Solid Earth Simulation Codes - GeoFEM and Householder QR Decomposition -

In this paper, we discuss vector and parallel tuning of GeoFEM and the Householder QR decomposition process being solid earth simulation codes. Process of GeoFEM code can be roughly divided into two parts, Matrix Assemble part and Solver part. Currently, GeoFEM's parallel iterative solver has attained good parallel and vector performance, and GeoFEM's Matrix Assemble part has attained good parallel performance too. The remaining problem is vector performance of Matrix Assemble part. In this study, we mainly focused on vector optimization of Matrix Assemble Part and vector performance of whole GeoFEM Test Code. In vector performance of whole GeoFEM Test Code, finally, 45% for peak performance (SX-4 /peak : 2GFlops) was obtained by changing loop structure. In this paper, we describe evaluation of vector/parallel performance of Householder QR decomposition processes in a new geodetic inversion method. Finally, 9.23GFlops was obtained for total performance by 8PE(SX-4/ peak : 16Gflops).