時系列データの統計解析によるPCクラスタシステム解析手法の提案  [in Japanese] System Analysis Method by Statistical Analysis on Time-series Data for a PC Cluster System  [in Japanese]

Access this Article

Search this Article

Author(s)

Abstract

本論文では,PC クラスタシステムにおける新たなシステム解析手法を提案する.本手法の特長は,時系列データを入力とし,統計解析の1 手法であるクラスタ分析を時間方向とノード方向の両方に対して行う点にある.本手法により,PC クラスタシステム上で採取した大量の測定データ(関数プロファイル情報およびCPI 等のCPU アーキテクチャ性能情報)から性能ボトルネックとなりうるシステムの挙動変化が,いつ・どのノードで発生したかを迅速に検出し,その原因を特定できる.我々は,本手法を実装した統合解析ツールiScopes(Integrated Software for Comprehensive Performance Evaluation System)を開発し,8 ノード(16CPU)で構成されるPC クラスタシステムへの実適用を行った.その結果,異常ノードの発見やアプリケーション分析に対する本手法の有用性を確認できた.In this paper, we propose a novel system analysis method for a PC cluster system. The feature of our method is to use cluster analysis (one of the statistical analysis methods) on time-series data for both directions of time and multiple nodes. Our method can quickly identify the performance bottleneck which can impact the behavior of the PC cluster system from a large amount of measurement data. Our target measurement data includes functional profiling data and CPU performance information from the architectural viewpoint. We developed iScopes (Integrated Software for Comprehensive Performance Evaluation System) implementing our method. Our evaluation result on the 8 nodes (16 CPUs) PC cluster system shows its effectiveness to identify the irregular node and analyze a parallel application behavior.

In this paper, we propose a novel system analysis method for a PC cluster system. The feature of our method is to use cluster analysis (one of the statistical analysis methods) on time-series data for both directions of time and multiple nodes. Our method can quickly identify the performance bottleneck which can impact the behavior of the PC cluster system from a large amount of measurement data. Our target measurement data includes functional profiling data and CPU performance information from the architectural viewpoint. We developed iScopes (Integrated Software for Comprehensive Performance Evaluation System) implementing our method. Our evaluation result on the 8 nodes (16 CPUs) PC cluster system shows its effectiveness to identify the irregular node and analyze a parallel application behavior.

Journal

  • 情報処理学会論文誌コンピューティングシステム(ACS)

    情報処理学会論文誌コンピューティングシステム(ACS) 47(SIG12(ACS15)), 250-261, 2006-09-15

    Information Processing Society of Japan (IPSJ)

References:  21

Codes

  • NII Article ID (NAID)
    110004782244
  • NII NACSIS-CAT ID (NCID)
    AA11833852
  • Text Lang
    JPN
  • Article Type
    Article
  • ISSN
    1882-7829
  • NDL Article ID
    8515843
  • NDL Call No.
    Z74-C192
  • Data Source
    CJP  NDL  NII-ELS  IPSJ 
Page Top