Large ML models and datasets have necessitated the use of multi-GPU systems for distributed model training. To harness the power offered by multi-GPU systems, it is critical to eliminate bottlenecks in inter-GPU communication – a problem made challenging by the heterogeneous nature of interconnects. In this work, we present TACCL, a synthesizer for collective communication primitives for large-scale multi-GPU systems. TACCL encodes a profiled topology and input size into a synthesis problem to generate optimized communication algorithms. TACCL is built on top of the standard NVIDIA Collective Communication Library (NCCL), allowing it to be a drop-in replacement for GPU communication in frameworks like PyTorch with minimal changes. TACCL generates algorithms for communication primitives like Allgather, Alltoall, and Allreduce that are up to $3\\times$ faster than NCCL. Using TACCL’s algorithms speeds up the end-to-end training of an internal mixture of experts model by $17\\%$. By decomposing the optimization problem into parts and leveraging the symmetry in multi-GPU topologies, TACCL synthesizes collectives for up to 80-GPUs in less than 3 minutes, at least two orders of magnitude faster than other synthesis-based state-of-the-art collective communication libraries.<\/p>\n","protected":false},"excerpt":{"rendered":"
Large ML models and datasets have necessitated the use of multi-GPU systems for distributed model training. To harness the power offered by multi-GPU systems, it is critical to eliminate bottlenecks in inter-GPU communication – a problem made challenging by the heterogeneous nature of interconnects. In this work, we present TACCL, a synthesizer for collective communication […]<\/p>\n","protected":false},"featured_media":0,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"footnotes":""},"msr-content-type":[3],"msr-research-highlight":[],"research-area":[13560,13547],"msr-publication-type":[193724],"msr-product-type":[],"msr-focus-area":[],"msr-platform":[],"msr-download-source":[],"msr-locale":[268875],"msr-field-of-study":[246904,263812,246691,263809,263806,249661,258049,250669,263803,263800],"msr-conference":[],"msr-journal":[],"msr-impact-theme":[],"msr-pillar":[],"class_list":["post-829342","msr-research-item","type-msr-research-item","status-publish","hentry","msr-research-area-programming-languages-software-engineering","msr-research-area-systems-and-networking","msr-locale-en_us","msr-field-of-study-algorithm","msr-field-of-study-collective-communication","msr-field-of-study-computer-science","msr-field-of-study-distributed-element-model","msr-field-of-study-heterogeneous-network","msr-field-of-study-mixture-of-experts","msr-field-of-study-network-topology","msr-field-of-study-optimization-problem","msr-field-of-study-power-physics","msr-field-of-study-topology-electrical-circuits"],"msr_publishername":"","msr_edition":"","msr_affiliation":"","msr_published_date":"2021-11-7","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":"arXiv preprint","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":1,"msr_main_download":"","msr_publicationurl":"","msr_doi":"","msr_publication_uploader":[{"type":"url","viewUrl":"false","id":"false","title":"http:\/\/arxiv-export-lb.library.cornell.edu\/abs\/2111.04867v2","label_id":"243109","label":0},{"type":"url","viewUrl":"false","id":"false","title":"https:\/\/dblp.uni-trier.de\/db\/journals\/corr\/corr2111.html#abs-2111-04867","label_id":"243109","label":0}],"msr_related_uploader":"","msr_attachments":[],"msr-author-ordering":[{"type":"user_nicename","value":"Aashaka Shah","user_id":43056,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Aashaka Shah"},{"type":"guest","value":"vijay-chidambaram","user_id":829348,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=vijay-chidambaram"},{"type":"user_nicename","value":"Meghan Cowan","user_id":41524,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Meghan Cowan"},{"type":"user_nicename","value":"Saeed Maleki","user_id":36131,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Saeed Maleki"},{"type":"user_nicename","value":"Madan Musuvathi","user_id":32766,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Madan Musuvathi"},{"type":"user_nicename","value":"Todd Mytkowicz","user_id":34235,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Todd Mytkowicz"},{"type":"user_nicename","value":"Jacob Nelson","user_id":36275,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Jacob Nelson"},{"type":"user_nicename","value":"Olli Saarikivi","user_id":37700,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Olli Saarikivi"},{"type":"user_nicename","value":"Rachee Singh","user_id":39952,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Rachee Singh"}],"msr_impact_theme":[],"msr_research_lab":[],"msr_event":[],"msr_group":[],"msr_project":[],"publication":[],"video":[],"download":[],"msr_publication_type":"miscellaneous","_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/829342"}],"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":2,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/829342\/revisions"}],"predecessor-version":[{"id":990768,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/829342\/revisions\/990768"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=829342"}],"wp:term":[{"taxonomy":"msr-content-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-content-type?post=829342"},{"taxonomy":"msr-research-highlight","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-highlight?post=829342"},{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=829342"},{"taxonomy":"msr-publication-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-type?post=829342"},{"taxonomy":"msr-product-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-product-type?post=829342"},{"taxonomy":"msr-focus-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-focus-area?post=829342"},{"taxonomy":"msr-platform","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-platform?post=829342"},{"taxonomy":"msr-download-source","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-download-source?post=829342"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=829342"},{"taxonomy":"msr-field-of-study","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-field-of-study?post=829342"},{"taxonomy":"msr-conference","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-conference?post=829342"},{"taxonomy":"msr-journal","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-journal?post=829342"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=829342"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=829342"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}