Network-Offloaded Bandwidth-Optimal Broadcast and Allgather for Distributed AI

In the Fully Sharded Data Parallel (FSDP) training pipeline, collective operations can be interleaved to maximize the communication/computation overlap. In this scenario, outstanding operations such as Allgather and Reduce-Scatter can compete for the injection bandwidth and create pipeline bubbles....

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:SC24: International Conference for High Performance Computing, Networking, Storage and Analysis s. 1 - 17
Hlavní autoři: Khalilov, Mikhail, Girolamo, Salvatore Di, Chrapek, Marcin, Nudelman, Rami, Bloch, Gil, Hoefler, Torsten
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 17.11.2024
Témata:
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Abstract In the Fully Sharded Data Parallel (FSDP) training pipeline, collective operations can be interleaved to maximize the communication/computation overlap. In this scenario, outstanding operations such as Allgather and Reduce-Scatter can compete for the injection bandwidth and create pipeline bubbles. To address this problem, we propose a novel bandwidth-optimal Allgather collective algorithm that leverages hardware multicast. We use multicast to build a constant-time reliable Broadcast protocol, a building block for constructing an optimal Allgather schedule. Our Allgather algorithm achieves 2 \times traffic reduction on a 188 -node testbed. To free the host side from running the protocol, we employ SmartNIC offloading. We extract the parallelism in our Allgather algorithm and map it to a SmartNIC specialized for hiding the cost of data movement. We show that our SmartNIC-offloaded collective progress engine can scale to the next generation of 1.6 Tbit/s links.
AbstractList In the Fully Sharded Data Parallel (FSDP) training pipeline, collective operations can be interleaved to maximize the communication/computation overlap. In this scenario, outstanding operations such as Allgather and Reduce-Scatter can compete for the injection bandwidth and create pipeline bubbles. To address this problem, we propose a novel bandwidth-optimal Allgather collective algorithm that leverages hardware multicast. We use multicast to build a constant-time reliable Broadcast protocol, a building block for constructing an optimal Allgather schedule. Our Allgather algorithm achieves 2 \times traffic reduction on a 188 -node testbed. To free the host side from running the protocol, we employ SmartNIC offloading. We extract the parallelism in our Allgather algorithm and map it to a SmartNIC specialized for hiding the cost of data movement. We show that our SmartNIC-offloaded collective progress engine can scale to the next generation of 1.6 Tbit/s links.
Author Bloch, Gil
Khalilov, Mikhail
Nudelman, Rami
Hoefler, Torsten
Girolamo, Salvatore Di
Chrapek, Marcin
Author_xml – sequence: 1
  givenname: Mikhail
  surname: Khalilov
  fullname: Khalilov, Mikhail
  email: mikhail.khalilov@inf.ethz.ch
  organization: ETH Zurich,Department of Computer Science,Zurich,Switzerland
– sequence: 2
  givenname: Salvatore Di
  surname: Girolamo
  fullname: Girolamo, Salvatore Di
  email: sdigirolamo@nvidia.com
  organization: NVIDIA Corporation,Zurich,Switzerland
– sequence: 3
  givenname: Marcin
  surname: Chrapek
  fullname: Chrapek, Marcin
  email: marcin.chrapek@inf.ethz.ch
  organization: ETH Zurich,Department of Computer Science,Zurich,Switzerland
– sequence: 4
  givenname: Rami
  surname: Nudelman
  fullname: Nudelman, Rami
  email: ramin@nvidia.com
  organization: NVIDIA Corporation,Santa Clara,United States
– sequence: 5
  givenname: Gil
  surname: Bloch
  fullname: Bloch, Gil
  email: gil@nvidia.com
  organization: NVIDIA Corporation,Yokne'am Illit,Israel
– sequence: 6
  givenname: Torsten
  surname: Hoefler
  fullname: Hoefler, Torsten
  email: torsten.hoefler@inf.ethz.ch
  organization: ETH Zurich,Department of Computer Science,Zurich,Switzerland
BookMark eNotjs9KxDAYxCMoqGtfQDzkBVq_JG3SHLv138KyPajnJWm-uMXaLmlk8e0N6GmG-THDXJPzaZ6QkFsGBWOg71_bkpUgCw68LABSdEYyrXQtKhAV10xdkmxZBguVUkIJEFdkt8N4msNn3nk_zsaho2szudPg4iHvjnH4MiNdh0R6s0SaEG3G8cPEAwbq50AfhiWGwX7H1Gw2N-TCm3HB7F9X5P3p8a19ybfd86ZttrnhVRlzC70FkKXTvXIieXSMaVtZI7Wrsao9Sl8jFzW3DJUGLzhoSDUm0afrK3L3tzsg4v4Y0s3ws2egtAAJ4henU090
CODEN IEEPAD
ContentType Conference Proceeding
DBID 6IE
6IL
CBEJK
RIE
RIL
DOI 10.1109/SC41406.2024.00109
DatabaseName IEEE Electronic Library (IEL) Conference Proceedings
IEEE Xplore POP ALL
IEEE Xplore All Conference Proceedings
IEEE Xplore Digital Library
IEEE Proceedings Order Plans (POP All) 1998-Present
DatabaseTitleList
Database_xml – sequence: 1
  dbid: RIE
  name: IEEE Electronic Library (IEL)
  url: https://ieeexplore.ieee.org/
  sourceTypes: Publisher
DeliveryMethod fulltext_linktorsrc
EISBN 9798350352917
EndPage 17
ExternalDocumentID 10793060
Genre orig-research
GroupedDBID 6IE
6IL
ACM
ALMA_UNASSIGNED_HOLDINGS
APO
CBEJK
LHSKQ
RIE
RIL
ID FETCH-LOGICAL-a254t-b0cb0064d9c7d3cb0ed119b5ba69d8e58fe6f8e2382b1e790f32090b0c16ef373
IEDL.DBID RIE
ISICitedReferencesCount 1
ISICitedReferencesURI http://www.webofscience.com/api/gateway?GWVersion=2&SrcApp=Summon&SrcAuth=ProQuest&DestLinkType=CitingArticles&DestApp=WOS_CPL&KeyUT=001414891300002&url=https%3A%2F%2Fcvtisr.summon.serialssolutions.com%2F%23%21%2Fsearch%3Fho%3Df%26include.ft.matches%3Dt%26l%3Dnull%26q%3D
IngestDate Wed Jan 01 06:01:57 EST 2025
IsPeerReviewed false
IsScholarly false
Language English
LinkModel DirectLink
MergedId FETCHMERGED-LOGICAL-a254t-b0cb0064d9c7d3cb0ed119b5ba69d8e58fe6f8e2382b1e790f32090b0c16ef373
PageCount 17
ParticipantIDs ieee_primary_10793060
PublicationCentury 2000
PublicationDate 2024-Nov.-17
PublicationDateYYYYMMDD 2024-11-17
PublicationDate_xml – month: 11
  year: 2024
  text: 2024-Nov.-17
  day: 17
PublicationDecade 2020
PublicationTitle SC24: International Conference for High Performance Computing, Networking, Storage and Analysis
PublicationTitleAbbrev SC
PublicationYear 2024
Publisher IEEE
Publisher_xml – name: IEEE
SSID ssib057737303
Score 1.9334532
Snippet In the Fully Sharded Data Parallel (FSDP) training pipeline, collective operations can be interleaved to maximize the communication/computation overlap. In...
SourceID ieee
SourceType Publisher
StartPage 1
SubjectTerms AI accelerators
Artificial intelligence
Clusters
Engines
Multicast algorithms
Networking
Next generation networking
Pipelines
Protocols
Substrates
Supercomputers
Training
Title Network-Offloaded Bandwidth-Optimal Broadcast and Allgather for Distributed AI
URI https://ieeexplore.ieee.org/document/10793060
WOSCitedRecordID wos001414891300002&url=https%3A%2F%2Fcvtisr.summon.serialssolutions.com%2F%23%21%2Fsearch%3Fho%3Df%26include.ft.matches%3Dt%26l%3Dnull%26q%3D
hasFullText 1
inHoldings 1
isFullTextHit
isPrint
link http://cvtisr.summon.serialssolutions.com/2.0.0/link/0/eLvHCXMwlV1LTwMhECbaePCkxhq1ajh4RffB8jjWaqOXtYma9NbAMpgm69a0W_37DrRVLx68EQghGQZmPphvhpBLG7wGrgUrvM4ZV5ljaEZy5sCiyoDIVOVisQlZlmo81qM1WT1yYQAgBp_BVWjGv3w3q5bhqQxPOGpTIhChb0spVmStjfIUUuaorfmGGJPo66cBR_gQ4hCykCI7Bh3-KqESLchw759r75PuDxePjr6tzAHZguaQlOUqeps9el_PjANHb0zjPqcuvFTiLfBmaooI27jKLFqKQ7Rf16_R26PoptLbkC83lLrCmf2HLnkZ3j0P7tm6MgIzCOhaZpMqHBfudCVdjm1waaptYY3QTkGhPAivAM1xZlOQOvF5lugEp6UCPMrpiHSaWQPHhFqjeSpN5qFA14gb4xXXRnJhEEuJ1J2QbhDG5H2V_GKykcPpH_09shvkHeh6qTwjnXa-hHOyU32008X8Im7ZF38blzs
linkProvider IEEE
linkToHtml http://cvtisr.summon.serialssolutions.com/2.0.0/link/0/eLvHCXMwlV1NSwMxEA1SBT2pWPHbHLxG9yObbI61Wlqsa8EKvZXsZiKFdSvtVv--k7RVLx68hSxhYXayMy-ZN4-Qq9xlDVwJllgVM55GhmEYiZmBHF0GRJQWxotNyCxLRyM1WJHVPRcGAHzxGVy7ob_LN9Ni4Y7KcIejNwUCEfqmk85KlnSttfskUsbor_GaGhOom-c2RwDhKhEi1yTblx3-ElHxMaSz-8-375HmDxuPDr7jzD7ZgOqAZNmyfps9WVtOtQFDb3VlPifGnVXif-BNlxQxtjaFntcUH9FWWb76fI9iokrvXMdcJ3aFK1u9Jnnp3A_bXbbSRmAaIV3N8qBwG4YbVUgT4xhMGKo8ybVQJoUktSBsChiQozwEqQIbR4EKcFkowKKdDkmjmlZwRGiuFQ-ljiwkmBxxrW3KlZZcaERTIjTHpOmMMX5ftr8Yr-1w8sf8JdnuDh_7434vezglO872jrwXyjPSqGcLOCdbxUc9mc8u_Of7ApS5moY
openUrl ctx_ver=Z39.88-2004&ctx_enc=info%3Aofi%2Fenc%3AUTF-8&rfr_id=info%3Asid%2Fsummon.serialssolutions.com&rft_val_fmt=info%3Aofi%2Ffmt%3Akev%3Amtx%3Abook&rft.genre=proceeding&rft.title=SC24%3A+International+Conference+for+High+Performance+Computing%2C+Networking%2C+Storage+and+Analysis&rft.atitle=Network-Offloaded+Bandwidth-Optimal+Broadcast+and+Allgather+for+Distributed+AI&rft.au=Khalilov%2C+Mikhail&rft.au=Girolamo%2C+Salvatore+Di&rft.au=Chrapek%2C+Marcin&rft.au=Nudelman%2C+Rami&rft.date=2024-11-17&rft.pub=IEEE&rft.spage=1&rft.epage=17&rft_id=info:doi/10.1109%2FSC41406.2024.00109&rft.externalDocID=10793060