同期操作に対する投機的メモリ・アクセス機構specMEMの改良  [in Japanese] Improvement of the Speculative Memory Access Mechanism : specMEM  [in Japanese]

Access this Article

Search this Article

Author(s)

Abstract

共有メモリ型並列計算機における同期処理オーバヘッドを削減する手法として,我々は同期操作に後続するメモリアクセスを同期成立確認以前に実行する機構specMEMを提案してきた.この機構の特徴は,投機失敗の検出やそれに伴う計算状態の復元を,機能メモリを用いたコヒーレント・キャッシュへの簡単な拡張により実現することにある.これまでの評価では,負荷の変動によって同期区間が伸縮するようなプログラムに対してspecMEMが有効であることが確かめられている.しかし同時に,投機によりキャッシュ・ミスペナルティが増加し,プログラムによっては性能が低下してしまうことも明らかになっている.そこで本報告ではspecMEMの改良方式として,投機的更新を示す新たな状態の追加と,通常のメモリで構成される2次キャッシュの導入を提案する.SPLASH-2ベンチマークを用いた評価を行なった結果,Radix Sortで見られた投機的実行による悪影響を60%程度削減できることが明らかになった.また2次キャッシュの導入はspecMEMの効果を高め,LU分解の性能向上率が22%から25%に増加することも明らかになった.In order to reduce the overhead of synchronizing operations of shared memory multiprocessors, we have proposed a mechanism named specMEM to excute memory accesses following a synchronizing operation speculatively before the completion of the synchronization is confirmed. A unique feature of our mechanism is that the detection of speculation failure and the restoration of computational state on the failure are implemented by a small extension of coherent cache with a simple functional memory. We showed that specMEM is effective to programs in which computational loads flactuate. We also observed, however, that the speculation increases cache miss penalty not only limlitting the efficiency of the specMEM but also degrading the performance of programs with load concentration. In this paper, we propose two techniques to reduce the cache miss penalty; adding one more cache state for speculation; and attaching a secondary cache using non-functional ordinary memory. The evaluation result with SPLASH-2 shows that the performance degradation factor of Radix Sort is reduced by 60%. It is also shown that secondary cache effectively improves the performance of specMEM. For example, the speed-up of LU decomposition is 25% with secondary cache while 22% without that.

In order to reduce the overhead of synchronizing operations of shared memory multiprocessors, we have proposed a mechanism named specMEM to excute memory accesses following a synchronizing operation speculatively before the completion of the synchronization is confirmed. A unique feature of our mechanism is that the detection of speculation failure and the restoration of computational state on the failure are implemented by a small extension of coherent cache with a simple functional memory. We showed that specMEM is effective to programs in which computational loads flactuate. We also observed, however, that the speculation increases cache miss penalty not only limitting the efficiency of the specMEM but also degrading the performance of programs with load concentration. In this paper, we propose two techniques to reduce the cache miss penalty;adding one more cache state for speculation;and attaching a secondary cache using non-functional ordinary memory. The evaluation result with SPLASH-2 shows that the performance degradation factor of Radix Sort is reduced by 60%. It is also shown that secondary cache effectively improves the performance of specMEM. For example, the speed-up of LU decomposition is 25% with secondary cache while 22% without that.

Journal

  • IPSJ SIG Notes

    IPSJ SIG Notes 2000(74(2000-ARC-139)), 13-18, 2000-08-03

    Information Processing Society of Japan (IPSJ)

References:  3

Codes

  • NII Article ID (NAID)
    110002774851
  • NII NACSIS-CAT ID (NCID)
    AN10096105
  • Text Lang
    JPN
  • Article Type
    Technical Report
  • ISSN
    09196072
  • NDL Article ID
    5740845
  • NDL Call No.
    Z14-1121
  • Data Source
    CJP  NDL  NII-ELS  IPSJ 
Page Top