Modern distributed machine learning (ML) training workloads benefit significantly from leveraging GPUs. However, significant contention ensues when multiple such workloads are run atop a shared cluster of GPUs. A key question is how to fairly apportion GPUs across workloads. We find that established cluster scheduling disciplines are a poor fit because of ML workloads\u2019 unique attributes: ML jobs have long-running tasks that need to be gang-scheduled, and their performance is sensitive to tasks\u2019 relative placement. <\/p>\n
We propose THEMIS, a new scheduling framework for ML training workloads. It\u2019s GPU allocation policy enforces that ML workloads complete in a finish-time fair manner, a new notion we introduce. To capture placement sensitivity and ensure efficiency, THEMIS uses a two-level scheduling architecture where ML workloads bid on available resources that are offered in an auction run by a central arbiter. Our auction design allocates GPUs to winning bids by trading off efficiency for fairness in the short term, but ensuring finish-time fairness in the long term. Our evaluation on a production trace shows that THEMIS can improve fairness by more than 2.25X and is 5% to 250% more cluster efficient in comparison to state-of-the-art schedulers.<\/p>\n","protected":false},"excerpt":{"rendered":"
Modern distributed machine learning (ML) training workloads benefit significantly from leveraging GPUs. However, significant contention ensues when multiple such workloads are run atop a shared cluster of GPUs. A key question is how to fairly apportion GPUs across workloads. We find that established cluster scheduling disciplines are a poor fit because of ML workloads\u2019 unique […]<\/p>\n","protected":false},"featured_media":0,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","footnotes":""},"msr-content-type":[3],"msr-research-highlight":[],"research-area":[13556,13547],"msr-publication-type":[193716],"msr-product-type":[],"msr-focus-area":[],"msr-platform":[],"msr-download-source":[],"msr-locale":[268875],"msr-post-option":[],"msr-field-of-study":[],"msr-conference":[],"msr-journal":[],"msr-impact-theme":[],"msr-pillar":[],"class_list":["post-627351","msr-research-item","type-msr-research-item","status-publish","hentry","msr-research-area-artificial-intelligence","msr-research-area-systems-and-networking","msr-locale-en_us"],"msr_publishername":"","msr_edition":"","msr_affiliation":"","msr_published_date":"2020-2-27","msr_host":"","msr_duration":"","msr_version":"","msr_speaker":"","msr_other_contributors":"","msr_booktitle":"","msr_pages_string":"","msr_chapter":"","msr_isbn":"","msr_journal":"","msr_volume":"","msr_number":"","msr_editors":"","msr_series":"","msr_issue":"","msr_organization":"","msr_how_published":"","msr_notes":"","msr_highlight_text":"","msr_release_tracker_id":"","msr_original_fields_of_study":"","msr_download_urls":"","msr_external_url":"","msr_secondary_video_url":"","msr_longbiography":"","msr_microsoftintellectualproperty":0,"msr_main_download":"","msr_publicationurl":"","msr_doi":"","msr_publication_uploader":[{"type":"url","viewUrl":"false","id":"false","title":"https:\/\/arxiv.org\/pdf\/1907.01484.pdf","label_id":"243109","label":0}],"msr_related_uploader":"","msr_attachments":[],"msr-author-ordering":[{"type":"text","value":"Kshiteej Mahajan","user_id":0,"rest_url":false},{"type":"text","value":"Arjun Balasubramanian","user_id":0,"rest_url":false},{"type":"text","value":"Arjun Singhvi","user_id":0,"rest_url":false},{"type":"text","value":"Shivaram Venkataraman","user_id":0,"rest_url":false},{"type":"text","value":"Aditya Akella","user_id":0,"rest_url":false},{"type":"user_nicename","value":"Amar Phanishayee","user_id":30975,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Amar Phanishayee"},{"type":"text","value":"Shuchi Chawla","user_id":0,"rest_url":false}],"msr_impact_theme":[],"msr_research_lab":[199565],"msr_event":[],"msr_group":[144927],"msr_project":[472845],"publication":[],"video":[],"download":[],"msr_publication_type":"inproceedings","related_content":{"projects":[{"ID":472845,"post_title":"Project Fiddle","post_name":"fiddle","post_type":"msr-project","post_date":"2018-03-20 10:20:05","post_modified":"2025-02-06 11:06:39","post_status":"publish","permalink":"https:\/\/www.microsoft.com\/en-us\/research\/project\/fiddle\/","post_excerpt":"The goal of Project Fiddle is to build efficient systems infrastructure for very fast distributed DNN training. Our goal is to support 100x more efficient training. Our innovations cut across the systems stack: the memory subsystem, structuring parallel computation across GPUs and machines, and interconnects between GPUs and across machines.","_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-project\/472845"}]}}]},"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/627351","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-research-item"}],"version-history":[{"count":1,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/627351\/revisions"}],"predecessor-version":[{"id":627357,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/627351\/revisions\/627357"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=627351"}],"wp:term":[{"taxonomy":"msr-content-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-content-type?post=627351"},{"taxonomy":"msr-research-highlight","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-highlight?post=627351"},{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=627351"},{"taxonomy":"msr-publication-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-type?post=627351"},{"taxonomy":"msr-product-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-product-type?post=627351"},{"taxonomy":"msr-focus-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-focus-area?post=627351"},{"taxonomy":"msr-platform","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-platform?post=627351"},{"taxonomy":"msr-download-source","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-download-source?post=627351"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=627351"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=627351"},{"taxonomy":"msr-field-of-study","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-field-of-study?post=627351"},{"taxonomy":"msr-conference","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-conference?post=627351"},{"taxonomy":"msr-journal","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-journal?post=627351"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=627351"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=627351"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}