非均質環境分散アルゴリズムの性能予測法と利用支援環境の構築に関する研究

摘要

The environment of parallel computing such as grids and clusters is widespread, but that configuration is not always homogeneous. In general, the execution performance of parallel programs greatly depends on algorithms and hardware configuration ,etc, and the performance is not improved simply in proportion to the number of PU(Processing Unit)'s. Then, it is useful to estimate beforehand the execution time of parallel programs for the optimization of the tuning of algorithm , arrangement of data,and etc. In our study, to support the development of parallel programs using MPI(Message Passing Interface), we improve the method of performance estimation based on simple trial execution and we propose a method that is effective for program in heterogenous environment. Moreover, the support system for development of parallel program is constructed.%近年，クラスタ環境の普及により，MPI（Message Passing Interface）を用いた並列計算のニーズが 高まってきている．一般に並列計算のプログラム は，PU（Processing Unit）の数の増加にともなって実 行性能が向上するが，ある一定数を越えると逆に低 下することがある．そのため，事前に並列プログラ ムの実行時間を予測し，最適なPU数を見積もる，あ るいは実行時間予測の情報を元に最適なPU構成を 求める手法が提案されている.