A scalable approach for LRT computation in GPGPU environments

Linsey Xiaolin Pang, Sanjay Chawla, Bernhard Scholz, Georgina Wilcox

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

4 Citations (Scopus)

Abstract

In this paper we propose new algorithmic techniques for massively data parallel computation of the Likelihood Ratio Test (LRT) on a large spatial data grid. LRT is the state-of-the-art method for identifying hotspots or anomalous regions in spatially referenced data. LRT is highly adaptable permitting the use of a large class of statistical distributions to model the data. However, standard sequential implementations of LRT may take several days on modern machines to identify anomalous regions even for moderately sized spatial grids. This work claims three novel contributions. First, we devise a dynamic program with a pre-processing step of O(n2) that allows us to compute the statistic for any given region in O(1), where n is the length of the grid. Second, we propose a scheme to accelerate the likelihood computation of a complement region using a bounding technique. Third, we provide a parallelization strategy for the LRT computation on GPGPUs. In concert all three contributions result in a speed up of nearly four hundred times reducing the LRT computation time of large spatial grids from several days to minutes.

Original languageEnglish
Title of host publicationLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Pages595-608
Number of pages14
Volume7808 LNCS
DOIs
Publication statusPublished - 2013
Externally publishedYes
Event15th Asia-Pacific Web Conference on Web Technologies and Applications, APWeb 2013 - Sydney, NSW
Duration: 4 Apr 20136 Apr 2013

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume7808 LNCS
ISSN (Print)03029743
ISSN (Electronic)16113349

Other

Other15th Asia-Pacific Web Conference on Web Technologies and Applications, APWeb 2013
CitySydney, NSW
Period4/4/136/4/13

Fingerprint

GPGPU
Likelihood Ratio Test
Grid
Anomalous
Data Grid
Statistical Distribution
Large Data
Statistics
Parallel Computation
Spatial Data
Hot Spot
Parallelization
Accelerate
Statistic
Preprocessing
Likelihood
Speedup
Complement
Processing

Keywords

  • 1EXP
  • GPGPUs
  • LRT
  • Spatial outlier
  • upper-bounding

ASJC Scopus subject areas

  • Computer Science(all)
  • Theoretical Computer Science

Cite this

Pang, L. X., Chawla, S., Scholz, B., & Wilcox, G. (2013). A scalable approach for LRT computation in GPGPU environments. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 7808 LNCS, pp. 595-608). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 7808 LNCS). https://doi.org/10.1007/978-3-642-37401-2_58

A scalable approach for LRT computation in GPGPU environments. / Pang, Linsey Xiaolin; Chawla, Sanjay; Scholz, Bernhard; Wilcox, Georgina.

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Vol. 7808 LNCS 2013. p. 595-608 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 7808 LNCS).

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

Pang, LX, Chawla, S, Scholz, B & Wilcox, G 2013, A scalable approach for LRT computation in GPGPU environments. in Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). vol. 7808 LNCS, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 7808 LNCS, pp. 595-608, 15th Asia-Pacific Web Conference on Web Technologies and Applications, APWeb 2013, Sydney, NSW, 4/4/13. https://doi.org/10.1007/978-3-642-37401-2_58
Pang LX, Chawla S, Scholz B, Wilcox G. A scalable approach for LRT computation in GPGPU environments. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Vol. 7808 LNCS. 2013. p. 595-608. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). https://doi.org/10.1007/978-3-642-37401-2_58
Pang, Linsey Xiaolin ; Chawla, Sanjay ; Scholz, Bernhard ; Wilcox, Georgina. / A scalable approach for LRT computation in GPGPU environments. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Vol. 7808 LNCS 2013. pp. 595-608 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).
@inproceedings{9496c1280cfa4d77826b4996673bb328,
title = "A scalable approach for LRT computation in GPGPU environments",
abstract = "In this paper we propose new algorithmic techniques for massively data parallel computation of the Likelihood Ratio Test (LRT) on a large spatial data grid. LRT is the state-of-the-art method for identifying hotspots or anomalous regions in spatially referenced data. LRT is highly adaptable permitting the use of a large class of statistical distributions to model the data. However, standard sequential implementations of LRT may take several days on modern machines to identify anomalous regions even for moderately sized spatial grids. This work claims three novel contributions. First, we devise a dynamic program with a pre-processing step of O(n2) that allows us to compute the statistic for any given region in O(1), where n is the length of the grid. Second, we propose a scheme to accelerate the likelihood computation of a complement region using a bounding technique. Third, we provide a parallelization strategy for the LRT computation on GPGPUs. In concert all three contributions result in a speed up of nearly four hundred times reducing the LRT computation time of large spatial grids from several days to minutes.",
keywords = "1EXP, GPGPUs, LRT, Spatial outlier, upper-bounding",
author = "Pang, {Linsey Xiaolin} and Sanjay Chawla and Bernhard Scholz and Georgina Wilcox",
year = "2013",
doi = "10.1007/978-3-642-37401-2_58",
language = "English",
isbn = "9783642374005",
volume = "7808 LNCS",
series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
pages = "595--608",
booktitle = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",

}

TY - GEN

T1 - A scalable approach for LRT computation in GPGPU environments

AU - Pang, Linsey Xiaolin

AU - Chawla, Sanjay

AU - Scholz, Bernhard

AU - Wilcox, Georgina

PY - 2013

Y1 - 2013

N2 - In this paper we propose new algorithmic techniques for massively data parallel computation of the Likelihood Ratio Test (LRT) on a large spatial data grid. LRT is the state-of-the-art method for identifying hotspots or anomalous regions in spatially referenced data. LRT is highly adaptable permitting the use of a large class of statistical distributions to model the data. However, standard sequential implementations of LRT may take several days on modern machines to identify anomalous regions even for moderately sized spatial grids. This work claims three novel contributions. First, we devise a dynamic program with a pre-processing step of O(n2) that allows us to compute the statistic for any given region in O(1), where n is the length of the grid. Second, we propose a scheme to accelerate the likelihood computation of a complement region using a bounding technique. Third, we provide a parallelization strategy for the LRT computation on GPGPUs. In concert all three contributions result in a speed up of nearly four hundred times reducing the LRT computation time of large spatial grids from several days to minutes.

AB - In this paper we propose new algorithmic techniques for massively data parallel computation of the Likelihood Ratio Test (LRT) on a large spatial data grid. LRT is the state-of-the-art method for identifying hotspots or anomalous regions in spatially referenced data. LRT is highly adaptable permitting the use of a large class of statistical distributions to model the data. However, standard sequential implementations of LRT may take several days on modern machines to identify anomalous regions even for moderately sized spatial grids. This work claims three novel contributions. First, we devise a dynamic program with a pre-processing step of O(n2) that allows us to compute the statistic for any given region in O(1), where n is the length of the grid. Second, we propose a scheme to accelerate the likelihood computation of a complement region using a bounding technique. Third, we provide a parallelization strategy for the LRT computation on GPGPUs. In concert all three contributions result in a speed up of nearly four hundred times reducing the LRT computation time of large spatial grids from several days to minutes.

KW - 1EXP

KW - GPGPUs

KW - LRT

KW - Spatial outlier

KW - upper-bounding

UR - http://www.scopus.com/inward/record.url?scp=84875850803&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84875850803&partnerID=8YFLogxK

U2 - 10.1007/978-3-642-37401-2_58

DO - 10.1007/978-3-642-37401-2_58

M3 - Conference contribution

SN - 9783642374005

VL - 7808 LNCS

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

SP - 595

EP - 608

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

ER -