2007年12月7日
OhHelp: a simple but efficient space-partitioned dynamic load balancing for particle simulations (ハイパフォーマンスコンピューティング)
情報処理学会研究報告ハイパフォーマンスコンピューティング(HPC)
- ,
- ,
- 巻
- 2007
- 号
- 122
- 開始ページ
- 25
- 終了ページ
- 30
- 記述言語
- 英語
- 掲載種別
- 出版者・発行元
- 一般社団法人情報処理学会
This paper proposes a new dynamic load balancing method for parallel particle simulations named OhHelp for the efficient execution on distributed memory systems. This method simply and equally partitions the space where charged particles reside but effectively balances the loads i.e. numbers of particles among computation nodes each of which is assigned a partitioned subspace as its primarily responsible one. The load balancing is achieved by letting a node help another node which gives a part of its overflown load to the helper node.Therefore the helper node is also responsible another subspace as its secondary one. The method makes the parallel simulation efficient not only because of the balanced number of particles but also thanks to the balanced computational load for subspaces and the simple boundary data exchange between adjacent subspaces. In fact our preliminary evaluation of an MPI implementation on the PrimePower HPC2500 supercomputer exhibits linear or super-linear speedup with up to 128 processors.This paper proposes a new dynamic load balancing method for parallel particle simulations, named OhHelp, for the efficient execution on distributed memory systems. This method simply and equally partitions the space where charged particles reside, but effectively balances the loads, i.e., numbers of particles, among computation nodes each of which is assigned a partitioned subspace as its primarily responsible one. The load balancing is achieved by letting a node help another node which gives a part of its overflown load to the helper node.Therefore, the helper node is also responsible another subspace as its secondary one. The method makes the parallel simulation efficient not only because of the balanced number of particles, but also thanks to the balanced computational load for subspaces and the simple boundary data exchange between adjacent subspaces. In fact, our preliminary evaluation of an MPI implementation on the PrimePower HPC2500 supercomputer exhibits linear or super-linear speedup with up to 128 processors.
- リンク情報
- ID情報
-
- ISSN : 0919-6072
- CiNii Articles ID : 110006549612
- CiNii Books ID : AN10463942