Performance and Accuracy Predictions of Approximation Methods for Shortest-Path Algorithms on Gpus

dc.contributor.author Aktılav, Busenur
dc.contributor.author Öz, Işıl
dc.date.accessioned 2022-08-11T10:55:44Z
dc.date.available 2022-08-11T10:55:44Z
dc.date.issued 2022
dc.description This work was supported by the Scientific and Technological Research Council of Turkey, Grant No: 119E011 en_US
dc.description.abstract Approximate computing techniques, where less-than-perfect solutions are acceptable, present performance-accuracy trade-offs by performing inexact computations. Moreover, heterogeneous architectures, a combination of miscellaneous compute units, offer high performance as well as energy efficiency. Graph algorithms utilize the parallel computation units of heterogeneous GPU architectures as well as performance improvements offered by approximation methods. Since different approximations yield different speedup and accuracy loss for the target execution, it becomes impractical to test all methods with various parameters. In this work, we perform approximate computations for the three shortest-path graph algorithms and propose a machine learning framework to predict the impact of the approximations on program performance and output accuracy. We evaluate random predictions for both synthetic and real road-network graphs, and predictions of the large graph cases from small graph instances. We achieve less than 5% prediction error rates for speedup and inaccuracy values. en_US
dc.identifier.doi 10.1016/j.parco.2022.102942
dc.identifier.issn 0167-8191 en_US
dc.identifier.issn 0167-8191
dc.identifier.scopus 2-s2.0-85133193847
dc.identifier.uri https://doi.org/10.1016/j.parco.2022.102942
dc.identifier.uri https://hdl.handle.net/11147/12307
dc.language.iso en en_US
dc.publisher Elsevier en_US
dc.relation.ispartof Parallel Computing en_US
dc.rights info:eu-repo/semantics/embargoedAccess en_US
dc.subject Approximate computing en_US
dc.subject GPU computing en_US
dc.subject Machine learning en_US
dc.title Performance and Accuracy Predictions of Approximation Methods for Shortest-Path Algorithms on Gpus en_US
dc.type Article en_US
dspace.entity.type Publication
gdc.author.id 0000-0002-8310-1143
gdc.author.id 0000-0002-8310-1143 en_US
gdc.author.institutional Aktılav, Busenur
gdc.author.institutional Öz, Işıl
gdc.bip.impulseclass C5
gdc.bip.influenceclass C5
gdc.bip.popularityclass C5
gdc.coar.access embargoed access
gdc.coar.type text::journal::journal article
gdc.collaboration.industrial false
gdc.contributor.affiliation 01. Izmir Institute of Technology en_US
gdc.contributor.affiliation 01. Izmir Institute of Technology en_US
gdc.description.department İzmir Institute of Technology. Computer Engineering en_US
gdc.description.publicationcategory Makale - Uluslararası Hakemli Dergi - Kurum Öğretim Elemanı en_US
gdc.description.scopusquality Q3
gdc.description.volume 112 en_US
gdc.description.wosquality Q2
gdc.identifier.openalex W4281794385
gdc.identifier.wos WOS:000833419300002
gdc.index.type WoS
gdc.index.type Scopus
gdc.oaire.diamondjournal false
gdc.oaire.impulse 0.0
gdc.oaire.influence 2.635068E-9
gdc.oaire.isgreen false
gdc.oaire.popularity 2.2369273E-9
gdc.oaire.publicfunded false
gdc.oaire.sciencefields 0103 physical sciences
gdc.oaire.sciencefields 0202 electrical engineering, electronic engineering, information engineering
gdc.oaire.sciencefields 02 engineering and technology
gdc.oaire.sciencefields 01 natural sciences
gdc.openalex.collaboration National
gdc.openalex.fwci 0.0
gdc.openalex.normalizedpercentile 0.04
gdc.opencitations.count 0
gdc.plumx.mendeley 5
gdc.plumx.scopuscites 0
gdc.scopus.citedcount 0
gdc.wos.citedcount 0
relation.isAuthorOfPublication.latestForDiscovery e0de33d0-b187-47e9-bae7-9b17aaabeb67
relation.isOrgUnitOfPublication.latestForDiscovery 9af2b05f-28ac-4014-8abe-a4dfe192da5e

Files

Original bundle

Now showing 1 - 1 of 1
Loading...
Name:
1-s2.0-S0167819122000400-main.pdf
Size:
2.64 MB
Format:
Adobe Portable Document Format
Description:
Article (Makale)

License bundle

Now showing 1 - 1 of 1
Loading...
Name:
license.txt
Size:
3.2 KB
Format:
Item-specific license agreed upon to submission
Description: