ホーム 研究室メンバー 研究内容 教育活動 アクセス リンク 学生の方はこちら VRミニ博物館 PHOTO GALLERY


Lecture 2

Exploiting Parallelism in Finite Element Flow Simulations

Flow simulation carries with it significant computing costs; computational analysis of complex flow systems becomes a reasonable alternative to experiments only if current high-performance computing resources can be effectively employed. The increasing interest in optimization and optimal control only increases these computational requirements. The Moore's law has long ceased to guarantee rapid performance improvements of single CPU cores, but continues to deliver multi-core CPUs with great potential for higher performance. The use of parallel computing techniques, which were previously necessary to exploit the largest and most expensive machines, is thus now a requirement even for single CPU use. The two dominant parallel programming models, MPI-based distributed-memory model and OpenMP-based shared-memory model, compete for attention of computational scientists, but may be in fact necessary to apply in tandem on emerging hardware. To complicate matters further, the fast-evolving field of accelerators such as GPUs leads to additional programming standards, with unclear return on investment. Several aspects of parallel implementation of finite element flow simulation codes are discussed, including iterative solution update techniques and communication kernels. Performance statistics of a parallel flow simulation code on the Blue Gene architecture are presented.

計算力学研究室 〒112-8551 東京都文京区春日1-13-27 2号館3階2307・2315号室 . 03-3817-1815
Copyright (C) Computational Mechanics Laboratory 2008 All Rights Reserved.