Technology-Driven, Highly-Scalable Dragonfly Topology

Evolving technology and increasing pin-bandwidth motivate the use of high-radix routers to reduce the diameter, latency, and cost of interconnection networks. High-radix networks, however, require longer cables than their low-radix counterparts. Because cables dominate network cost, the number of ca...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Kim, John, Dally, Wiliam J., Scott, Steve, Abts, Dennis
Format: Tagungsbericht
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page 88
container_issue
container_start_page 77
container_title
container_volume
creator Kim, John
Dally, Wiliam J.
Scott, Steve
Abts, Dennis
description Evolving technology and increasing pin-bandwidth motivate the use of high-radix routers to reduce the diameter, latency, and cost of interconnection networks. High-radix networks, however, require longer cables than their low-radix counterparts. Because cables dominate network cost, the number of cables, and particularly the number of long, global cables should be minimized to realize an efficient network. In this paper, we introduce the dragonfly topology which uses a group of high-radix routers as a virtual router to increase the effective radix of the network. With this organization, each minimally routed packet traverses at most one global channel. By reducing global channels, a dragonfly reduces cost by 20% compared to a flattened butterfly and by 52% compared to a folded Clos network in configurations with ≥ 16K nodes.We also introduce two new variants of global adaptive routing that enable load-balanced routing in the dragonfly. Each router in a dragonfly must make an adaptive routing decision based on the state of a global channel connected to a different router. Because of the indirect nature of this routing decision, conventional adaptive routing algorithms give degraded performance. We introduce the use of selective virtual-channel discrimination and the use of credit round-trip latency to both sense and signal channel congestion. The combination of these two methods gives throughput and latency that approaches that of an ideal adaptive routing algorithm.
doi_str_mv 10.1109/ISCA.2008.19
format Conference Proceeding
fullrecord <record><control><sourceid>acm_6IE</sourceid><recordid>TN_cdi_ieee_primary_4556717</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><ieee_id>4556717</ieee_id><sourcerecordid>acm_books_10_1109_ISCA_2008_19</sourcerecordid><originalsourceid>FETCH-LOGICAL-a284t-4ef5aa72f5248ac0d46105779f3a7b034f5f982434cfcd1e5c0494bb33dc32723</originalsourceid><addsrcrecordid>eNqNkDtPwzAYRS0eElHpxsaShQkcPr9ie6xaoJUqMbRIbJbj2GnATaoEIeXfk1J-ANMd7tHV1UHohkBGCOjH1WY-yyiAyog-QwkVUmBJ2Ps5mmqpQOZaMCK5ukAJgZzhXGl5haZ9_wEAROcjyxIktt7tmja21YAXXf3tm4d0WVe7OOCNs9EW0aeLzlZtE-KQbtvDL3qNLoONvZ_-5QS9PT9t50u8fn1ZzWdrbKniX5j7IKyVNAjKlXVQ8pyAkFIHZmUBjAcRtKKccRdcSbxwwDUvCsZKx6ikbIJuT7u1994cunpvu8FwIcb3cmzvT611e1O07WdvCJijHHOUY45yDNGm6GofRvruPzT7AbZQXys</addsrcrecordid><sourcetype>Publisher</sourcetype><iscdi>true</iscdi><recordtype>conference_proceeding</recordtype></control><display><type>conference_proceeding</type><title>Technology-Driven, Highly-Scalable Dragonfly Topology</title><source>IEEE Electronic Library (IEL) Conference Proceedings</source><creator>Kim, John ; Dally, Wiliam J. ; Scott, Steve ; Abts, Dennis</creator><creatorcontrib>Kim, John ; Dally, Wiliam J. ; Scott, Steve ; Abts, Dennis</creatorcontrib><description>Evolving technology and increasing pin-bandwidth motivate the use of high-radix routers to reduce the diameter, latency, and cost of interconnection networks. High-radix networks, however, require longer cables than their low-radix counterparts. Because cables dominate network cost, the number of cables, and particularly the number of long, global cables should be minimized to realize an efficient network. In this paper, we introduce the dragonfly topology which uses a group of high-radix routers as a virtual router to increase the effective radix of the network. With this organization, each minimally routed packet traverses at most one global channel. By reducing global channels, a dragonfly reduces cost by 20% compared to a flattened butterfly and by 52% compared to a folded Clos network in configurations with ≥ 16K nodes.We also introduce two new variants of global adaptive routing that enable load-balanced routing in the dragonfly. Each router in a dragonfly must make an adaptive routing decision based on the state of a global channel connected to a different router. Because of the indirect nature of this routing decision, conventional adaptive routing algorithms give degraded performance. We introduce the use of selective virtual-channel discrimination and the use of credit round-trip latency to both sense and signal channel congestion. The combination of these two methods gives throughput and latency that approaches that of an ideal adaptive routing algorithm.</description><identifier>ISSN: 1063-6897</identifier><identifier>ISBN: 9780769531748</identifier><identifier>ISBN: 0769531741</identifier><identifier>EISSN: 2575-713X</identifier><identifier>DOI: 10.1109/ISCA.2008.19</identifier><language>eng</language><publisher>Washington, DC, USA: IEEE Computer Society</publisher><subject>Bandwidth ; Cables ; Computer architecture ; Costs ; Delay ; dragonfly ; Hardware ; Hardware -- Hardware validation ; Hardware -- Integrated circuits -- Interconnect ; Hardware -- Integrated circuits -- Interconnect -- Input -- output circuits ; interconnection networks ; Multiprocessor interconnection networks ; Network topology ; Optical network units ; Optical sensors ; Routing ; topology</subject><ispartof>2008 International Symposium on Computer Architecture, 2008, p.77-88</ispartof><woscitedreferencessubscribed>false</woscitedreferencessubscribed><citedby>FETCH-LOGICAL-a284t-4ef5aa72f5248ac0d46105779f3a7b034f5f982434cfcd1e5c0494bb33dc32723</citedby></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><linktohtml>$$Uhttps://ieeexplore.ieee.org/document/4556717$$EHTML$$P50$$Gieee$$H</linktohtml><link.rule.ids>309,310,780,784,789,790,2058,27925,54920</link.rule.ids><linktorsrc>$$Uhttps://ieeexplore.ieee.org/document/4556717$$EView_record_in_IEEE$$FView_record_in_$$GIEEE</linktorsrc></links><search><creatorcontrib>Kim, John</creatorcontrib><creatorcontrib>Dally, Wiliam J.</creatorcontrib><creatorcontrib>Scott, Steve</creatorcontrib><creatorcontrib>Abts, Dennis</creatorcontrib><title>Technology-Driven, Highly-Scalable Dragonfly Topology</title><title>2008 International Symposium on Computer Architecture</title><addtitle>ISCA</addtitle><description>Evolving technology and increasing pin-bandwidth motivate the use of high-radix routers to reduce the diameter, latency, and cost of interconnection networks. High-radix networks, however, require longer cables than their low-radix counterparts. Because cables dominate network cost, the number of cables, and particularly the number of long, global cables should be minimized to realize an efficient network. In this paper, we introduce the dragonfly topology which uses a group of high-radix routers as a virtual router to increase the effective radix of the network. With this organization, each minimally routed packet traverses at most one global channel. By reducing global channels, a dragonfly reduces cost by 20% compared to a flattened butterfly and by 52% compared to a folded Clos network in configurations with ≥ 16K nodes.We also introduce two new variants of global adaptive routing that enable load-balanced routing in the dragonfly. Each router in a dragonfly must make an adaptive routing decision based on the state of a global channel connected to a different router. Because of the indirect nature of this routing decision, conventional adaptive routing algorithms give degraded performance. We introduce the use of selective virtual-channel discrimination and the use of credit round-trip latency to both sense and signal channel congestion. The combination of these two methods gives throughput and latency that approaches that of an ideal adaptive routing algorithm.</description><subject>Bandwidth</subject><subject>Cables</subject><subject>Computer architecture</subject><subject>Costs</subject><subject>Delay</subject><subject>dragonfly</subject><subject>Hardware</subject><subject>Hardware -- Hardware validation</subject><subject>Hardware -- Integrated circuits -- Interconnect</subject><subject>Hardware -- Integrated circuits -- Interconnect -- Input -- output circuits</subject><subject>interconnection networks</subject><subject>Multiprocessor interconnection networks</subject><subject>Network topology</subject><subject>Optical network units</subject><subject>Optical sensors</subject><subject>Routing</subject><subject>topology</subject><issn>1063-6897</issn><issn>2575-713X</issn><isbn>9780769531748</isbn><isbn>0769531741</isbn><fulltext>true</fulltext><rsrctype>conference_proceeding</rsrctype><creationdate>2008</creationdate><recordtype>conference_proceeding</recordtype><sourceid>6IE</sourceid><sourceid>RIE</sourceid><recordid>eNqNkDtPwzAYRS0eElHpxsaShQkcPr9ie6xaoJUqMbRIbJbj2GnATaoEIeXfk1J-ANMd7tHV1UHohkBGCOjH1WY-yyiAyog-QwkVUmBJ2Ps5mmqpQOZaMCK5ukAJgZzhXGl5haZ9_wEAROcjyxIktt7tmja21YAXXf3tm4d0WVe7OOCNs9EW0aeLzlZtE-KQbtvDL3qNLoONvZ_-5QS9PT9t50u8fn1ZzWdrbKniX5j7IKyVNAjKlXVQ8pyAkFIHZmUBjAcRtKKccRdcSbxwwDUvCsZKx6ikbIJuT7u1994cunpvu8FwIcb3cmzvT611e1O07WdvCJijHHOUY45yDNGm6GofRvruPzT7AbZQXys</recordid><startdate>20080601</startdate><enddate>20080601</enddate><creator>Kim, John</creator><creator>Dally, Wiliam J.</creator><creator>Scott, Steve</creator><creator>Abts, Dennis</creator><general>IEEE Computer Society</general><general>IEEE</general><scope>6IE</scope><scope>6IH</scope><scope>CBEJK</scope><scope>RIE</scope><scope>RIO</scope></search><sort><creationdate>20080601</creationdate><title>Technology-Driven, Highly-Scalable Dragonfly Topology</title><author>Kim, John ; Dally, Wiliam J. ; Scott, Steve ; Abts, Dennis</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-LOGICAL-a284t-4ef5aa72f5248ac0d46105779f3a7b034f5f982434cfcd1e5c0494bb33dc32723</frbrgroupid><rsrctype>conference_proceedings</rsrctype><prefilter>conference_proceedings</prefilter><language>eng</language><creationdate>2008</creationdate><topic>Bandwidth</topic><topic>Cables</topic><topic>Computer architecture</topic><topic>Costs</topic><topic>Delay</topic><topic>dragonfly</topic><topic>Hardware</topic><topic>Hardware -- Hardware validation</topic><topic>Hardware -- Integrated circuits -- Interconnect</topic><topic>Hardware -- Integrated circuits -- Interconnect -- Input -- output circuits</topic><topic>interconnection networks</topic><topic>Multiprocessor interconnection networks</topic><topic>Network topology</topic><topic>Optical network units</topic><topic>Optical sensors</topic><topic>Routing</topic><topic>topology</topic><toplevel>online_resources</toplevel><creatorcontrib>Kim, John</creatorcontrib><creatorcontrib>Dally, Wiliam J.</creatorcontrib><creatorcontrib>Scott, Steve</creatorcontrib><creatorcontrib>Abts, Dennis</creatorcontrib><collection>IEEE Electronic Library (IEL) Conference Proceedings</collection><collection>IEEE Proceedings Order Plan (POP) 1998-present by volume</collection><collection>IEEE Xplore All Conference Proceedings</collection><collection>IEEE Electronic Library (IEL)</collection><collection>IEEE Proceedings Order Plans (POP) 1998-present</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext_linktorsrc</fulltext></delivery><addata><au>Kim, John</au><au>Dally, Wiliam J.</au><au>Scott, Steve</au><au>Abts, Dennis</au><format>book</format><genre>proceeding</genre><ristype>CONF</ristype><atitle>Technology-Driven, Highly-Scalable Dragonfly Topology</atitle><btitle>2008 International Symposium on Computer Architecture</btitle><stitle>ISCA</stitle><date>2008-06-01</date><risdate>2008</risdate><spage>77</spage><epage>88</epage><pages>77-88</pages><issn>1063-6897</issn><eissn>2575-713X</eissn><isbn>9780769531748</isbn><isbn>0769531741</isbn><abstract>Evolving technology and increasing pin-bandwidth motivate the use of high-radix routers to reduce the diameter, latency, and cost of interconnection networks. High-radix networks, however, require longer cables than their low-radix counterparts. Because cables dominate network cost, the number of cables, and particularly the number of long, global cables should be minimized to realize an efficient network. In this paper, we introduce the dragonfly topology which uses a group of high-radix routers as a virtual router to increase the effective radix of the network. With this organization, each minimally routed packet traverses at most one global channel. By reducing global channels, a dragonfly reduces cost by 20% compared to a flattened butterfly and by 52% compared to a folded Clos network in configurations with ≥ 16K nodes.We also introduce two new variants of global adaptive routing that enable load-balanced routing in the dragonfly. Each router in a dragonfly must make an adaptive routing decision based on the state of a global channel connected to a different router. Because of the indirect nature of this routing decision, conventional adaptive routing algorithms give degraded performance. We introduce the use of selective virtual-channel discrimination and the use of credit round-trip latency to both sense and signal channel congestion. The combination of these two methods gives throughput and latency that approaches that of an ideal adaptive routing algorithm.</abstract><cop>Washington, DC, USA</cop><pub>IEEE Computer Society</pub><doi>10.1109/ISCA.2008.19</doi><tpages>12</tpages></addata></record>
fulltext fulltext_linktorsrc
identifier ISSN: 1063-6897
ispartof 2008 International Symposium on Computer Architecture, 2008, p.77-88
issn 1063-6897
2575-713X
language eng
recordid cdi_ieee_primary_4556717
source IEEE Electronic Library (IEL) Conference Proceedings
subjects Bandwidth
Cables
Computer architecture
Costs
Delay
dragonfly
Hardware
Hardware -- Hardware validation
Hardware -- Integrated circuits -- Interconnect
Hardware -- Integrated circuits -- Interconnect -- Input -- output circuits
interconnection networks
Multiprocessor interconnection networks
Network topology
Optical network units
Optical sensors
Routing
topology
title Technology-Driven, Highly-Scalable Dragonfly Topology
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2024-12-24T15%3A00%3A39IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-acm_6IE&rft_val_fmt=info:ofi/fmt:kev:mtx:book&rft.genre=proceeding&rft.atitle=Technology-Driven,%20Highly-Scalable%20Dragonfly%20Topology&rft.btitle=2008%20International%20Symposium%20on%20Computer%20Architecture&rft.au=Kim,%20John&rft.date=2008-06-01&rft.spage=77&rft.epage=88&rft.pages=77-88&rft.issn=1063-6897&rft.eissn=2575-713X&rft.isbn=9780769531748&rft.isbn_list=0769531741&rft_id=info:doi/10.1109/ISCA.2008.19&rft_dat=%3Cacm_6IE%3Eacm_books_10_1109_ISCA_2008_19%3C/acm_6IE%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_id=info:pmid/&rft_ieee_id=4556717&rfr_iscdi=true