Multilayer cache partitioning for multiprogram workloads

Mahmut Kandemir, Ramya Prabhakar, Mustafa Karakoy, Yuanrui Zhang

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

2 Citations (Scopus)

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

Fingerprint

Cache
Workload
Multilayer
Partitioning
Multilayers
Quality of service
Quality of Service
Partition
Experimental Analysis
Performance Metrics
Isolation
Excess
Speedup
Model-based
Metric
Requirements

All Science Journal Classification (ASJC) codes

  • Theoretical Computer Science
  • Computer Science(all)

Cite this

Kandemir, M., Prabhakar, R., Karakoy, M., & Zhang, Y. (2011). Multilayer cache partitioning for multiprogram workloads. In Euro-Par 2011 Parallel Processing - 17th International Conference, Proceedings (PART 1 ed., pp. 130-141). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 6852 LNCS, No. PART 1). https://doi.org/10.1007/978-3-642-23400-2_13
Kandemir, Mahmut ; Prabhakar, Ramya ; Karakoy, Mustafa ; Zhang, Yuanrui. / Multilayer cache partitioning for multiprogram workloads. Euro-Par 2011 Parallel Processing - 17th International Conference, Proceedings. PART 1. ed. 2011. pp. 130-141 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); PART 1).
@inproceedings{c634a4c32ce84317b738c27dc1fed633,
title = "Multilayer cache partitioning for multiprogram workloads",
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.",
author = "Mahmut Kandemir and Ramya Prabhakar and Mustafa Karakoy and Yuanrui Zhang",
year = "2011",
month = "9",
day = "8",
doi = "10.1007/978-3-642-23400-2_13",
language = "English (US)",
isbn = "9783642233999",
series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
number = "PART 1",
pages = "130--141",
booktitle = "Euro-Par 2011 Parallel Processing - 17th International Conference, Proceedings",
edition = "PART 1",

}

Kandemir, M, Prabhakar, R, Karakoy, M & Zhang, Y 2011, Multilayer cache partitioning for multiprogram workloads. in Euro-Par 2011 Parallel Processing - 17th International Conference, Proceedings. PART 1 edn, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), no. PART 1, vol. 6852 LNCS, pp. 130-141, 17th International Conference on Parallel Processing, Euro-Par 2011, Bordeaux, France, 8/29/11. https://doi.org/10.1007/978-3-642-23400-2_13

Multilayer cache partitioning for multiprogram workloads. / Kandemir, Mahmut; Prabhakar, Ramya; Karakoy, Mustafa; Zhang, Yuanrui.

Euro-Par 2011 Parallel Processing - 17th International Conference, Proceedings. PART 1. ed. 2011. p. 130-141 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 6852 LNCS, No. PART 1).

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

TY - GEN

T1 - Multilayer cache partitioning for multiprogram workloads

AU - Kandemir, Mahmut

AU - Prabhakar, Ramya

AU - Karakoy, Mustafa

AU - Zhang, Yuanrui

PY - 2011/9/8

Y1 - 2011/9/8

N2 - 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.

AB - 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.

UR - http://www.scopus.com/inward/record.url?scp=80052386002&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=80052386002&partnerID=8YFLogxK

U2 - 10.1007/978-3-642-23400-2_13

DO - 10.1007/978-3-642-23400-2_13

M3 - Conference contribution

AN - SCOPUS:80052386002

SN - 9783642233999

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 130

EP - 141

BT - Euro-Par 2011 Parallel Processing - 17th International Conference, Proceedings

ER -

Kandemir M, Prabhakar R, Karakoy M, Zhang Y. Multilayer cache partitioning for multiprogram workloads. In Euro-Par 2011 Parallel Processing - 17th International Conference, Proceedings. PART 1 ed. 2011. p. 130-141. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); PART 1). https://doi.org/10.1007/978-3-642-23400-2_13