STrack: A Reliable Multipath Transport for AI/ML Clusters

Emerging artificial intelligence (AI) and machine learning (ML) workloads present new challenges of managing the collective communication used in distributed training across hundreds or even thousands of GPUs. This paper presents STrack, a novel hardware-offloaded reliable transport protocol aimed a...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Le, Yanfang, Pan, Rong, Newman, Peter, Blendin, Jeremias, Kabbani, Abdul, Jain, Vipin, Sivaramu, Raghava, Matus, Francis
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page
container_issue
container_start_page
container_title
container_volume
creator Le, Yanfang
Pan, Rong
Newman, Peter
Blendin, Jeremias
Kabbani, Abdul
Jain, Vipin
Sivaramu, Raghava
Matus, Francis
description Emerging artificial intelligence (AI) and machine learning (ML) workloads present new challenges of managing the collective communication used in distributed training across hundreds or even thousands of GPUs. This paper presents STrack, a novel hardware-offloaded reliable transport protocol aimed at improving the performance of AI /ML workloads by rethinking key aspects of the transport layer. STrack optimizes congestion control and load balancing in tandem: it incorporates an adaptive load balancing algorithm leveraging ECN, while adopts RTT as multi-bit congestion indicators for precise congestion window adjustment. Additionally, STrack facilitates out-of-order delivery, selective retransmission, and swift loss recovery in hardware for multipath environment. The extensive simulation comparing STrack and RoCEv2 demonstrates that STrack outperforms RoCEv2 by up to 6X with synthetic workloads and by 27.4% with collective workloads, even with the optimized RoCEv2 system setup.
doi_str_mv 10.48550/arxiv.2407.15266
format Article
fullrecord <record><control><sourceid>arxiv_GOX</sourceid><recordid>TN_cdi_arxiv_primary_2407_15266</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2407_15266</sourcerecordid><originalsourceid>FETCH-arxiv_primary_2407_152663</originalsourceid><addsrcrecordid>eNpjYJA0NNAzsTA1NdBPLKrILNMzMjEw1zM0NTIz42SwDA4pSkzOtlJwVAhKzclMTMpJVfAtzSnJLEgsyVAAyuUVF-QXlSik5RcpOHrq-_ooOOeUFpekFhXzMLCmJeYUp_JCaW4GeTfXEGcPXbAl8QVFmbmJRZXxIMviwZYZE1YBAMuDMwY</addsrcrecordid><sourcetype>Open Access Repository</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype></control><display><type>article</type><title>STrack: A Reliable Multipath Transport for AI/ML Clusters</title><source>arXiv.org</source><creator>Le, Yanfang ; Pan, Rong ; Newman, Peter ; Blendin, Jeremias ; Kabbani, Abdul ; Jain, Vipin ; Sivaramu, Raghava ; Matus, Francis</creator><creatorcontrib>Le, Yanfang ; Pan, Rong ; Newman, Peter ; Blendin, Jeremias ; Kabbani, Abdul ; Jain, Vipin ; Sivaramu, Raghava ; Matus, Francis</creatorcontrib><description>Emerging artificial intelligence (AI) and machine learning (ML) workloads present new challenges of managing the collective communication used in distributed training across hundreds or even thousands of GPUs. This paper presents STrack, a novel hardware-offloaded reliable transport protocol aimed at improving the performance of AI /ML workloads by rethinking key aspects of the transport layer. STrack optimizes congestion control and load balancing in tandem: it incorporates an adaptive load balancing algorithm leveraging ECN, while adopts RTT as multi-bit congestion indicators for precise congestion window adjustment. Additionally, STrack facilitates out-of-order delivery, selective retransmission, and swift loss recovery in hardware for multipath environment. The extensive simulation comparing STrack and RoCEv2 demonstrates that STrack outperforms RoCEv2 by up to 6X with synthetic workloads and by 27.4% with collective workloads, even with the optimized RoCEv2 system setup.</description><identifier>DOI: 10.48550/arxiv.2407.15266</identifier><language>eng</language><subject>Computer Science - Networking and Internet Architecture</subject><creationdate>2024-07</creationdate><rights>http://creativecommons.org/licenses/by/4.0</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>228,230,780,885</link.rule.ids><linktorsrc>$$Uhttps://arxiv.org/abs/2407.15266$$EView_record_in_Cornell_University$$FView_record_in_$$GCornell_University$$Hfree_for_read</linktorsrc><backlink>$$Uhttps://doi.org/10.48550/arXiv.2407.15266$$DView paper in arXiv$$Hfree_for_read</backlink></links><search><creatorcontrib>Le, Yanfang</creatorcontrib><creatorcontrib>Pan, Rong</creatorcontrib><creatorcontrib>Newman, Peter</creatorcontrib><creatorcontrib>Blendin, Jeremias</creatorcontrib><creatorcontrib>Kabbani, Abdul</creatorcontrib><creatorcontrib>Jain, Vipin</creatorcontrib><creatorcontrib>Sivaramu, Raghava</creatorcontrib><creatorcontrib>Matus, Francis</creatorcontrib><title>STrack: A Reliable Multipath Transport for AI/ML Clusters</title><description>Emerging artificial intelligence (AI) and machine learning (ML) workloads present new challenges of managing the collective communication used in distributed training across hundreds or even thousands of GPUs. This paper presents STrack, a novel hardware-offloaded reliable transport protocol aimed at improving the performance of AI /ML workloads by rethinking key aspects of the transport layer. STrack optimizes congestion control and load balancing in tandem: it incorporates an adaptive load balancing algorithm leveraging ECN, while adopts RTT as multi-bit congestion indicators for precise congestion window adjustment. Additionally, STrack facilitates out-of-order delivery, selective retransmission, and swift loss recovery in hardware for multipath environment. The extensive simulation comparing STrack and RoCEv2 demonstrates that STrack outperforms RoCEv2 by up to 6X with synthetic workloads and by 27.4% with collective workloads, even with the optimized RoCEv2 system setup.</description><subject>Computer Science - Networking and Internet Architecture</subject><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2024</creationdate><recordtype>article</recordtype><sourceid>GOX</sourceid><recordid>eNpjYJA0NNAzsTA1NdBPLKrILNMzMjEw1zM0NTIz42SwDA4pSkzOtlJwVAhKzclMTMpJVfAtzSnJLEgsyVAAyuUVF-QXlSik5RcpOHrq-_ooOOeUFpekFhXzMLCmJeYUp_JCaW4GeTfXEGcPXbAl8QVFmbmJRZXxIMviwZYZE1YBAMuDMwY</recordid><startdate>20240721</startdate><enddate>20240721</enddate><creator>Le, Yanfang</creator><creator>Pan, Rong</creator><creator>Newman, Peter</creator><creator>Blendin, Jeremias</creator><creator>Kabbani, Abdul</creator><creator>Jain, Vipin</creator><creator>Sivaramu, Raghava</creator><creator>Matus, Francis</creator><scope>AKY</scope><scope>GOX</scope></search><sort><creationdate>20240721</creationdate><title>STrack: A Reliable Multipath Transport for AI/ML Clusters</title><author>Le, Yanfang ; Pan, Rong ; Newman, Peter ; Blendin, Jeremias ; Kabbani, Abdul ; Jain, Vipin ; Sivaramu, Raghava ; Matus, Francis</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-arxiv_primary_2407_152663</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2024</creationdate><topic>Computer Science - Networking and Internet Architecture</topic><toplevel>online_resources</toplevel><creatorcontrib>Le, Yanfang</creatorcontrib><creatorcontrib>Pan, Rong</creatorcontrib><creatorcontrib>Newman, Peter</creatorcontrib><creatorcontrib>Blendin, Jeremias</creatorcontrib><creatorcontrib>Kabbani, Abdul</creatorcontrib><creatorcontrib>Jain, Vipin</creatorcontrib><creatorcontrib>Sivaramu, Raghava</creatorcontrib><creatorcontrib>Matus, Francis</creatorcontrib><collection>arXiv Computer Science</collection><collection>arXiv.org</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext_linktorsrc</fulltext></delivery><addata><au>Le, Yanfang</au><au>Pan, Rong</au><au>Newman, Peter</au><au>Blendin, Jeremias</au><au>Kabbani, Abdul</au><au>Jain, Vipin</au><au>Sivaramu, Raghava</au><au>Matus, Francis</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>STrack: A Reliable Multipath Transport for AI/ML Clusters</atitle><date>2024-07-21</date><risdate>2024</risdate><abstract>Emerging artificial intelligence (AI) and machine learning (ML) workloads present new challenges of managing the collective communication used in distributed training across hundreds or even thousands of GPUs. This paper presents STrack, a novel hardware-offloaded reliable transport protocol aimed at improving the performance of AI /ML workloads by rethinking key aspects of the transport layer. STrack optimizes congestion control and load balancing in tandem: it incorporates an adaptive load balancing algorithm leveraging ECN, while adopts RTT as multi-bit congestion indicators for precise congestion window adjustment. Additionally, STrack facilitates out-of-order delivery, selective retransmission, and swift loss recovery in hardware for multipath environment. The extensive simulation comparing STrack and RoCEv2 demonstrates that STrack outperforms RoCEv2 by up to 6X with synthetic workloads and by 27.4% with collective workloads, even with the optimized RoCEv2 system setup.</abstract><doi>10.48550/arxiv.2407.15266</doi><oa>free_for_read</oa></addata></record>
fulltext fulltext_linktorsrc
identifier DOI: 10.48550/arxiv.2407.15266
ispartof
issn
language eng
recordid cdi_arxiv_primary_2407_15266
source arXiv.org
subjects Computer Science - Networking and Internet Architecture
title STrack: A Reliable Multipath Transport for AI/ML Clusters
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2024-12-23T06%3A01%3A51IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-arxiv_GOX&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=STrack:%20A%20Reliable%20Multipath%20Transport%20for%20AI/ML%20Clusters&rft.au=Le,%20Yanfang&rft.date=2024-07-21&rft_id=info:doi/10.48550/arxiv.2407.15266&rft_dat=%3Carxiv_GOX%3E2407_15266%3C/arxiv_GOX%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_id=info:pmid/&rfr_iscdi=true