超並列環境向きの固有値計算アルゴリズムと自動チューニング(<特集>数値計算のための自動チューニング)

概要

論文の詳細を見る
On the emerging era for many-core computing environment, there are three topics to be overcome; i) highly parallel algorithm, ii) resource conflict avoiding algorithms and implementation technique, iii) flexible algorithm and implementation being aware of heterogeneity. The selection of an appropriate resource from a large amount of resource is a sort of optimization problem, and automatic tuning technique plays a big role here. It is needless to say that significance of numerical linear libraries rises in the field of the scientific simulations in the future. However, we must argue whether the existing libraries and algorithms are suitable for a present multicore and a massively parallel computer system. This paper surveys the trend for the existing software, especially the numerical linear algebra library for multicore processors. The investigation focuses on the standard eigenvalue problem of dense real symmetrical matrices, and suitable algorithm for the parallel computing environment and the automatic tuning technique are reported. Finally, it touches about the next generation eigenvalue solver developed on the next generation supercomputer system.
2010-09-24