実行時間予測モデルの構築法の改善 Improvements of Execution-time Estimation Model Construction

この論文にアクセスする

この論文をさがす

著者

    • 河合 裕 KAWAI YUU
    • 豊橋技術科学大学知識情報工学系 Department of Knowledge-based Information Engineering, Toyohashi University of Technology
    • 市川 周一 ICHIKAWA SHUICHI
    • 豊橋技術科学大学知識情報工学系 Department of Knowledge-based Information Engineering, Toyohashi University of Technology

抄録

既存の並列応用の多くは均一並列環境を前提としているため,不均一クラスタ上で実行すると負荷不均衡により性能が低下する.高速PE上に複数のプロセス起動することで負荷の均衡化が望めるが,最適に負荷を分散することは難しい.高橋と市川は実行時間予測モデルを構築し,4つの科学技術応用(HPL,CFD,FEM,FFT)について(準)最適構成を予測できることを示した.しかしモデル構築のために均一なサブクラスタを必要とする上,大きく精度が低下する場合があった.本研究では,不均一クラスタ自体からモデルを構築する方法と,精度低下を防ぐモデル構築手法を検討する.評価の結果,従来と同等以上のモデルを構築し,より高い精度で(準)最適構成を予測することができた.The performance of a parallel application for homogeneous environment is degraded by load imbalance on heterogeneous clusters. Such load imbalance can be alleviated by invoking multiple processes on fast PEs in a heterogeneous cluster. Takahashi and Ichikawa constructed execution-time estimation models for four scientific applications, with which the sub-optimal configurations of heterogeneous clusters were estimated. However, their models have to be extracted from homogeneous subclusters, and the precisions of models are very sensitive to the fluctuations in measurement results. This study examines a method to construct models from heterogeneous cluster itself, and a method to construct more robust models. The derived models were shown to be superior to the previous models.

The performiince of a parallel application for homogeneous environment is degraded by load imbalance on heterogeneous clusters. Such load imbalance can be alleviated by invoking multiple processes on fast PEs in a heterogeneous cluster. Takahashi and Ichikawa constructed execution-time estimation models for four scientific applications, with which the sub-optimal configiirations of heterogeneous clusters were estimated. However, their models have to be extracted from homogeneous subclusters, and the precisions of models are very sensitive to the fluctuations in measurement results. This study examines a method to construct models from heterogeneous cluster itself, and a method to construct more robust models. The derived models were shown to be superior to the previous models.

収録刊行物

  • 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)

    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) 2007(17(2007-HPC-109)), 79-84, 2007-03-01

    一般社団法人情報処理学会

各種コード

  • NII論文ID(NAID)
    110006248422
  • NII書誌ID(NCID)
    AN10463942
  • 本文言語コード
    JPN
  • 資料種別
    Technical Report
  • ISSN
    09196072
  • NDL 記事登録ID
    8704712
  • NDL 雑誌分類
    ZM13(科学技術--科学技術一般--データ処理・計算機)
  • NDL 請求記号
    Z14-1121
  • データ提供元
    NDL  NII-ELS  IPSJ 
ページトップへ