WIP: TRLAN SVD iterative solver with fixed-size reverse communication for in-database processing of sparse datasets

Piotr Luszczek; Ichitaro Yamazaki; Jakub Kurzak; Jack Dongarra

WIP: TRLAN SVD iterative solver with fixed-size reverse communication for in-database processing of sparse datasets

Piotr Luszczek, Ichitaro Yamazaki, Jakub Kurzak, Jack Dongarra

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

In this paper, we present that an implementation of an iterative TRLAN (Thick Restart Lanczos) SVD (Singular Value Decomposition) algorithm that has an important property of not requiring any ownership of neither the matrix nor the vector data. We call this new implementation a fixed-size reverse communication scheme. One crucial consequence is that we achieved a communication bound limited by the iteration count. Also, the amount of communicated data is independent of the input problem size. Our performance results show that this affords very good scalability even in the case when data is stored inside a database system with all the associated overheads of failover and resilience.

Original language	English
Title of host publication	Simulation Series
Editors	Hamdi Kavak, Omer F. Keskin
Publisher	The Society for Modeling and Simulation International
Edition	9
ISBN (Electronic)	9781510860186
ISBN (Print)	9781510860131, 9781510860148, 9781510860155, 9781510860162, 9781510860179, 9781510860186, 9781510860186, 9781510860209
State	Published - 2018
Externally published	Yes
Event	Works in Progress Symposium, WIP 2018, Part of the 2018 Spring Simulation Multiconference, SpringSim 2018 - Baltimore, United States Duration: Apr 15 2018 → Apr 18 2018

Publication series

Name	Simulation Series
Number	9
Volume	50
ISSN (Print)	0735-9276

Conference

Conference	Works in Progress Symposium, WIP 2018, Part of the 2018 Spring Simulation Multiconference, SpringSim 2018
Country/Territory	United States
City	Baltimore
Period	04/15/18 → 04/18/18

Funding

JACK DONGARRA holds appointments at the University of Tennessee Knoxville, Oak Ridge National Laboratory, and the Manchester University. He specializes in numerical algorithms in linear algebra, parallel computing, use of advanced computer architectures, programming methodology, and tools for parallel computers. Dongarra was awarded the IEEE Sid Fern-bach Award in 2004, received the first IEEE Medal of Excellence in Scalable Computing in 2008, the first SIAM Special Interest Group on Supercomputing’s award for Career Achievement in 2010, and the IEEE IPDPS 2011 Charles Babbage Award. He’s a Fellow of the AAAS, ACM, IEEE, and SIAM, and a member of the National Academy of Engineering. His email address is [email protected].

Funders	Funder number
IEEE Foundation

Keywords

Array database storage
Reverse communication iteration
Thick restart lanczos singular valued decomposition

Cite this

@inproceedings{4385606acb594b658abdb5183953bc48,

title = "WIP: TRLAN SVD iterative solver with fixed-size reverse communication for in-database processing of sparse datasets",

abstract = "In this paper, we present that an implementation of an iterative TRLAN (Thick Restart Lanczos) SVD (Singular Value Decomposition) algorithm that has an important property of not requiring any ownership of neither the matrix nor the vector data. We call this new implementation a fixed-size reverse communication scheme. One crucial consequence is that we achieved a communication bound limited by the iteration count. Also, the amount of communicated data is independent of the input problem size. Our performance results show that this affords very good scalability even in the case when data is stored inside a database system with all the associated overheads of failover and resilience.",

keywords = "Array database storage, Reverse communication iteration, Thick restart lanczos singular valued decomposition",

author = "Piotr Luszczek and Ichitaro Yamazaki and Jakub Kurzak and Jack Dongarra",

note = "Publisher Copyright: {\textcopyright} 2018 Society for Modeling and Simulation (SCS) International.; Works in Progress Symposium, WIP 2018, Part of the 2018 Spring Simulation Multiconference, SpringSim 2018 ; Conference date: 15-04-2018 Through 18-04-2018",

year = "2018",

language = "English",

isbn = "9781510860131",

series = "Simulation Series",

publisher = "The Society for Modeling and Simulation International",

number = "9",

editor = "Hamdi Kavak and Keskin, {Omer F.}",

booktitle = "Simulation Series",

edition = "9",

}

Luszczek, P, Yamazaki, I, Kurzak, J & Dongarra, J 2018, WIP: TRLAN SVD iterative solver with fixed-size reverse communication for in-database processing of sparse datasets. in H Kavak & OF Keskin (eds), Simulation Series. 9 edn, Simulation Series, no. 9, vol. 50, The Society for Modeling and Simulation International, Works in Progress Symposium, WIP 2018, Part of the 2018 Spring Simulation Multiconference, SpringSim 2018, Baltimore, United States, 04/15/18.

WIP: TRLAN SVD iterative solver with fixed-size reverse communication for in-database processing of sparse datasets. / Luszczek, Piotr; Yamazaki, Ichitaro; Kurzak, Jakub et al.
Simulation Series. ed. / Hamdi Kavak; Omer F. Keskin. 9. ed. The Society for Modeling and Simulation International, 2018. (Simulation Series; Vol. 50, No. 9).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - WIP

T2 - Works in Progress Symposium, WIP 2018, Part of the 2018 Spring Simulation Multiconference, SpringSim 2018

AU - Luszczek, Piotr

AU - Yamazaki, Ichitaro

AU - Kurzak, Jakub

AU - Dongarra, Jack

PY - 2018

Y1 - 2018

N2 - In this paper, we present that an implementation of an iterative TRLAN (Thick Restart Lanczos) SVD (Singular Value Decomposition) algorithm that has an important property of not requiring any ownership of neither the matrix nor the vector data. We call this new implementation a fixed-size reverse communication scheme. One crucial consequence is that we achieved a communication bound limited by the iteration count. Also, the amount of communicated data is independent of the input problem size. Our performance results show that this affords very good scalability even in the case when data is stored inside a database system with all the associated overheads of failover and resilience.

AB - In this paper, we present that an implementation of an iterative TRLAN (Thick Restart Lanczos) SVD (Singular Value Decomposition) algorithm that has an important property of not requiring any ownership of neither the matrix nor the vector data. We call this new implementation a fixed-size reverse communication scheme. One crucial consequence is that we achieved a communication bound limited by the iteration count. Also, the amount of communicated data is independent of the input problem size. Our performance results show that this affords very good scalability even in the case when data is stored inside a database system with all the associated overheads of failover and resilience.

KW - Array database storage

KW - Reverse communication iteration

KW - Thick restart lanczos singular valued decomposition

UR - http://www.scopus.com/inward/record.url?scp=85055053901&partnerID=8YFLogxK

M3 - Conference contribution

AN - SCOPUS:85055053901

SN - 9781510860131

SN - 9781510860148

SN - 9781510860155

SN - 9781510860162

SN - 9781510860179

SN - 9781510860186

SN - 9781510860209

T3 - Simulation Series

BT - Simulation Series

A2 - Kavak, Hamdi

A2 - Keskin, Omer F.

PB - The Society for Modeling and Simulation International

Y2 - 15 April 2018 through 18 April 2018

ER -

WIP: TRLAN SVD iterative solver with fixed-size reverse communication for in-database processing of sparse datasets

Abstract

Publication series

Conference

Funding

Keywords

Other files and links

Fingerprint

Cite this