Multilayer cache partitioning for multiprogram workloads

Mahmut Kandemir, Ramya Prabhakar, Mustafa Karakoy, Yuanrui Zhang

Research output: Chapter in Book/Report/Conference proceedingConference contribution

2 Scopus citations

Abstract

We present a fully-automated, model based, multilayer cache partitioning scheme for multiprogram workloads running on multicore machines. As opposed to prior efforts, this scheme partitions shared caches at multiple layers simultaneously in a coordinated fashion. This scheme tries to achieve two objectives. First, it tries to satisfy the specified quality of service (QoS) values for all applications by partitioning the shared cache hierarchy across them, and second, it distributes the remaining excess cache capacity (if any) across applications such that a global performance metric is maximized. Our experimental analysis shows that the proposed multilayer partitioning scheme generates, on average, 33.1% improvement (on the weighted speedup metric) over the next best-performing scheme and is very successful in satisfying the QoS requirements of applications. Also, we show that partitioning each layer in isolation cannot generate the benefits obtained through our coordinated partitioning scheme. In addition, we observed that the difference between our scheme and an optimal scheme (that derives best dynamic partitions) was less than 15% for all the workloads tested and 6.6% on average.

Original languageEnglish (US)
Title of host publicationEuro-Par 2011 Parallel Processing - 17th International Conference, Proceedings
Pages130-141
Number of pages12
EditionPART 1
DOIs
StatePublished - Sep 8 2011
Event17th International Conference on Parallel Processing, Euro-Par 2011 - Bordeaux, France
Duration: Aug 29 2011Sep 2 2011

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
NumberPART 1
Volume6852 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Other

Other17th International Conference on Parallel Processing, Euro-Par 2011
CountryFrance
CityBordeaux
Period8/29/119/2/11

All Science Journal Classification (ASJC) codes

  • Theoretical Computer Science
  • Computer Science(all)

Fingerprint Dive into the research topics of 'Multilayer cache partitioning for multiprogram workloads'. Together they form a unique fingerprint.

Cite this