メニーコア向けNUMA最適並列分散I/Oの予備検証  [in Japanese] Preliminary Validation of NUMA aware Parallelized I/O for Manycore  [in Japanese]

Access this Article

Search this Article

Author(s)

Abstract

将来の High Performance Computing (HPC) システムではメニーコア化・NUMA 化が進み,これらを考慮したシステムソフトウェアが重要になる.本稿では NUMA 型のメニーコアシステムにおける並列分散 I/O の実現に向けた予備検証の結果と考察について述べる.予備検証では,NUMA 環境での性能律速要素と律速時の性能影響を確認するため,既存の NUMA 環境でキャッシュ I/O における read/write の I/O 性能を測定した.本稿では予備検証で得た結果をもとに,将来 HPC での最適化された並列分散 I/O を実現する上での最適な分散ポリシーについて議論する.Future High Performance Computing (HPC) systems will have more manycore processors with NUMA. This paper shows results of preliminary validation of NUMA aware Parallelized I/O for Manycore and delivers consideration for it. In the preliminary validation, we measured I/O bandwidth of read /write using cache I/O on existing NUMA machine to check rate-limiting factors and impact of the effect. This paper discusses I/O distribution policies to optimize parallelized I/O for future HPC system.

Future High Performance Computing (HPC) systems will have more manycore processors with NUMA. This paper shows results of preliminary validation of NUMA aware Parallelized I/O for Manycore and delivers consideration for it. In the preliminary validation, we measured I/O bandwidth of read /write using cache I/O on existing NUMA machine to check rate-limiting factors and impact of the effect. This paper discusses I/O distribution policies to optimize parallelized I/O for future HPC system.

Journal

  • IPSJ SIG Notes

    IPSJ SIG Notes 2013-OS-125(1), 1-7, 2013-04-18

    Information Processing Society of Japan (IPSJ)

Keywords

Codes

  • NII Article ID (NAID)
    110009579587
  • NII NACSIS-CAT ID (NCID)
    AN10444176
  • Text Lang
    JPN
  • Article Type
    Technical Report
  • Data Source
    NII-ELS  IPSJ 
Page Top