Distributed in-GPU data cache for document-oriented data store via PCIe over 10 Gbit ethernet

Shin Morishima, Hiroki Matsutani

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

1 Citation (Scopus)

Abstract

As one of NOSQL data stores, a document-oriented data store manages data as documents in a scheme-less manner. Various string match queries, such as a perfect match, begins-with (prefix) match, partial match, and regular expression based match, are performed for the documents. To accelerate such string match queries, we propose DistGPU Cache (Distributed In-GPU Data Cache), in which data store server and GPU devices are connected via a PCI-Express (PCIe) over 10 Gbit Ethernet (10 GbE), so that GPU devices that store and search documents can be added and removed dynamically. We also propose a partitioning method that distributes ranges of cached documents to GPU devices based on a hash function. The distributed cache over GPU devices can be dynamically divided and merged when the GPU devices are added and removed, respectively. We evaluate the proposed DistGPU Cache in terms of regular expression match query throughput with up to three NVIDIA GeForce GTX 980 devices connected to a host via PCIe over 10 GbE. We demonstrate that the communication overhead of remote GPU devices is small and can be compensated by a great flexibility to add more GPU devices via a network. We also show that DistGPU Cache with the remote GPU devices significantly outperforms the original data store.

Original languageEnglish
Title of host publicationEuro-Par 2016
Subtitle of host publicationParallel Processing Workshops - Euro-Par 2016 International Workshops, Revised Selected Papers
PublisherSpringer Verlag
Pages41-55
Number of pages15
Volume10104 LNCS
ISBN (Print)9783319589428
DOIs
Publication statusPublished - 2017
Event22nd International Conference on Parallel and Distributed Computing, Euro-Par 2016 - Grenoble, France
Duration: 2016 Aug 242016 Aug 26

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume10104 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Other

Other22nd International Conference on Parallel and Distributed Computing, Euro-Par 2016
CountryFrance
CityGrenoble
Period16/8/2416/8/26

Fingerprint

Ethernet
Cache
Express
Regular Expressions
Query
Strings
Graphics processing unit
Hash functions
Hash Function
Prefix
Accelerate
Partitioning
Throughput
Servers
Server
Flexibility
Partial
Evaluate
Communication

ASJC Scopus subject areas

  • Theoretical Computer Science
  • Computer Science(all)

Cite this

Morishima, S., & Matsutani, H. (2017). Distributed in-GPU data cache for document-oriented data store via PCIe over 10 Gbit ethernet. In Euro-Par 2016: Parallel Processing Workshops - Euro-Par 2016 International Workshops, Revised Selected Papers (Vol. 10104 LNCS, pp. 41-55). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 10104 LNCS). Springer Verlag. https://doi.org/10.1007/978-3-319-58943-5_4

Distributed in-GPU data cache for document-oriented data store via PCIe over 10 Gbit ethernet. / Morishima, Shin; Matsutani, Hiroki.

Euro-Par 2016: Parallel Processing Workshops - Euro-Par 2016 International Workshops, Revised Selected Papers. Vol. 10104 LNCS Springer Verlag, 2017. p. 41-55 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 10104 LNCS).

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

Morishima, S & Matsutani, H 2017, Distributed in-GPU data cache for document-oriented data store via PCIe over 10 Gbit ethernet. in Euro-Par 2016: Parallel Processing Workshops - Euro-Par 2016 International Workshops, Revised Selected Papers. vol. 10104 LNCS, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 10104 LNCS, Springer Verlag, pp. 41-55, 22nd International Conference on Parallel and Distributed Computing, Euro-Par 2016, Grenoble, France, 16/8/24. https://doi.org/10.1007/978-3-319-58943-5_4
Morishima S, Matsutani H. Distributed in-GPU data cache for document-oriented data store via PCIe over 10 Gbit ethernet. In Euro-Par 2016: Parallel Processing Workshops - Euro-Par 2016 International Workshops, Revised Selected Papers. Vol. 10104 LNCS. Springer Verlag. 2017. p. 41-55. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). https://doi.org/10.1007/978-3-319-58943-5_4
Morishima, Shin ; Matsutani, Hiroki. / Distributed in-GPU data cache for document-oriented data store via PCIe over 10 Gbit ethernet. Euro-Par 2016: Parallel Processing Workshops - Euro-Par 2016 International Workshops, Revised Selected Papers. Vol. 10104 LNCS Springer Verlag, 2017. pp. 41-55 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).
@inproceedings{325e15dcea374b75a96aeab6c02eaeea,
title = "Distributed in-GPU data cache for document-oriented data store via PCIe over 10 Gbit ethernet",
abstract = "As one of NOSQL data stores, a document-oriented data store manages data as documents in a scheme-less manner. Various string match queries, such as a perfect match, begins-with (prefix) match, partial match, and regular expression based match, are performed for the documents. To accelerate such string match queries, we propose DistGPU Cache (Distributed In-GPU Data Cache), in which data store server and GPU devices are connected via a PCI-Express (PCIe) over 10 Gbit Ethernet (10 GbE), so that GPU devices that store and search documents can be added and removed dynamically. We also propose a partitioning method that distributes ranges of cached documents to GPU devices based on a hash function. The distributed cache over GPU devices can be dynamically divided and merged when the GPU devices are added and removed, respectively. We evaluate the proposed DistGPU Cache in terms of regular expression match query throughput with up to three NVIDIA GeForce GTX 980 devices connected to a host via PCIe over 10 GbE. We demonstrate that the communication overhead of remote GPU devices is small and can be compensated by a great flexibility to add more GPU devices via a network. We also show that DistGPU Cache with the remote GPU devices significantly outperforms the original data store.",
author = "Shin Morishima and Hiroki Matsutani",
year = "2017",
doi = "10.1007/978-3-319-58943-5_4",
language = "English",
isbn = "9783319589428",
volume = "10104 LNCS",
series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
publisher = "Springer Verlag",
pages = "41--55",
booktitle = "Euro-Par 2016",
address = "Germany",

}

TY - GEN

T1 - Distributed in-GPU data cache for document-oriented data store via PCIe over 10 Gbit ethernet

AU - Morishima, Shin

AU - Matsutani, Hiroki

PY - 2017

Y1 - 2017

N2 - As one of NOSQL data stores, a document-oriented data store manages data as documents in a scheme-less manner. Various string match queries, such as a perfect match, begins-with (prefix) match, partial match, and regular expression based match, are performed for the documents. To accelerate such string match queries, we propose DistGPU Cache (Distributed In-GPU Data Cache), in which data store server and GPU devices are connected via a PCI-Express (PCIe) over 10 Gbit Ethernet (10 GbE), so that GPU devices that store and search documents can be added and removed dynamically. We also propose a partitioning method that distributes ranges of cached documents to GPU devices based on a hash function. The distributed cache over GPU devices can be dynamically divided and merged when the GPU devices are added and removed, respectively. We evaluate the proposed DistGPU Cache in terms of regular expression match query throughput with up to three NVIDIA GeForce GTX 980 devices connected to a host via PCIe over 10 GbE. We demonstrate that the communication overhead of remote GPU devices is small and can be compensated by a great flexibility to add more GPU devices via a network. We also show that DistGPU Cache with the remote GPU devices significantly outperforms the original data store.

AB - As one of NOSQL data stores, a document-oriented data store manages data as documents in a scheme-less manner. Various string match queries, such as a perfect match, begins-with (prefix) match, partial match, and regular expression based match, are performed for the documents. To accelerate such string match queries, we propose DistGPU Cache (Distributed In-GPU Data Cache), in which data store server and GPU devices are connected via a PCI-Express (PCIe) over 10 Gbit Ethernet (10 GbE), so that GPU devices that store and search documents can be added and removed dynamically. We also propose a partitioning method that distributes ranges of cached documents to GPU devices based on a hash function. The distributed cache over GPU devices can be dynamically divided and merged when the GPU devices are added and removed, respectively. We evaluate the proposed DistGPU Cache in terms of regular expression match query throughput with up to three NVIDIA GeForce GTX 980 devices connected to a host via PCIe over 10 GbE. We demonstrate that the communication overhead of remote GPU devices is small and can be compensated by a great flexibility to add more GPU devices via a network. We also show that DistGPU Cache with the remote GPU devices significantly outperforms the original data store.

UR - http://www.scopus.com/inward/record.url?scp=85020405798&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85020405798&partnerID=8YFLogxK

U2 - 10.1007/978-3-319-58943-5_4

DO - 10.1007/978-3-319-58943-5_4

M3 - Conference contribution

AN - SCOPUS:85020405798

SN - 9783319589428

VL - 10104 LNCS

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

SP - 41

EP - 55

BT - Euro-Par 2016

PB - Springer Verlag

ER -