[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-117753-en":3,"doc-seo-117753-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":27,"seo_description":14,"update_tm":28,"read_time":29},117753,1099513958607,"Jiven","https://ap-avatar.wpscdn.com/avatar/100002390cf8733938c?x-image-process=image/resize,m_fixed,w_180,h_180&k=1778829742770036399",8,"Research & Report","Baechi - Fast Device Placement of Machine Learning Graphs","Baechi targets the placement challenge in distributed machine learning when GPUs or devices have limited memory and training graphs must be split across devices. Existing learning-based model-parallelism can produce fast training step times, but it often requires hours or days to generate an operator placement plan. Baechi introduces an algorithmic placement system for small memory-constrained clusters and integrates with TensorFlow and PyTorch. Experiments on GPUs show placement plan generation is 654×–206K× faster than prior learning-based methods, with training step time close to expert placements and within 6.2% in TensorFlow.","arXiv :2301 .08695v1 [ cs .DC] 20 Jan 2023  \nBaechi: Fast Device Placement of Machine Learning Graphs  \nBEOMYEOL JEON, University of Illinois at Urbana-Champaign, USA LINDA CAI, Princeton University, USA  \nCHIRAG SHETTY, University of Illinois at Urbana-Champaign, USA PALLAVI SRIVASTAVA, University of Illinois at Urbana-Champaign*, USA JINTAO JIANG, University of Illinois at Urbana-Champaign*, USA XIAOLAN KE, University of Illinois at Urbana-Champaign*, USA YITAO MENG, University of Illinois at Urbana-Champaign*, USA CONG XIE, University of Illinois at Urbana-Champaign*, USA INDRANIL GUPTA, University of Illinois at Urbana-Champaign, USA  \nMachine Learning graphs (or models) can be challenging or impossible to train when either devices have limited memory, or models are large. To split the model across devices, learning-based approaches are still popular. While these result in model placements that train fast on data (i.e., low step times), learning-based model-parallelism is time-consuming, taking many hours or days to create a placement plan of operators on devices. We present the Baechi system, the first to adopt an algorithmic approach to the placement problem for running machine learning training graphs on small clusters of memory-constrained devices. We integrate our implementation of Baechi into two popular open-source learning frameworks: TensorFlow and PyTorch. Our experimental results using GPUs show that: (i) Baechi generates placement plans 654×–206K × faster than state-of-the-art learning-based approaches, and (ii) Baechi-placed model’s step (training) time is comparable to expert placements in PyTorch, and only up to 6.2% worse than expert placements in TensorFlow. We prove mathematically that our two algorithms are within a constant factor of the optimal. Our work shows that compared to learning-based approaches, algorithmic approaches can face different challenges for adaptation to Machine learning systems, but also they offer proven bounds, and significant performance benefits.  \nCCS Concepts: • Computer systems organization → Cloud computing.  \nAdditional Key Words and Phrases: Machine Learning Systems, Placement Algorithms, Constrained Memory, TensorFlow, PyTorch, Distributed Systems  \nThis submission is an extended version of \"Baechi: Fast Device Placement of Machine Learning Graphs - Beomyeol Jeon, Linda Cai, Pallavi Srivastava, Jintao Jiang, Xiaolan Ke, Yitao Meng, Cong Xie, and Indranil Gupta. In Proceedings of the 11th ACM Symposium on Cloud Computing (Virtual Event, USA) (SoCC’20). Association for Computing Machinery, New York, NY, USA, 416–430. [https://doi.org/10.1145/3419111.3421302](https://doi.org/10.1145/3419111.3421302)\". Document detailing the additional contributions has been attached as a supplementary material  \n*Work done while the authors were at University of Illinois at Urbana-Champaign, USA  \nAuthors’ addresses: Beomyeol Jeon, University of Illinois at Urbana-Champaign, Urbana, USA, [bj2@illinois.edu](bj2@illinois.edu); Linda Cai, Princeton University,  \nUSA, [tcai@princeton.edu](tcai@princeton.edu); Chirag Shetty, University of Illinois at Urbana-Champaign, USA, [cshetty2@illinois.edu](cshetty2@illinois.edu); Pallavi Srivastava, University ofIlli  \nnois at Urbana-Champaign*, USA; Jintao Jiang, University of Illinois at Urbana-Champaign*, USA; Xiaolan Ke, University of Illinois at Urbana-Champaign*, USA; Yitao Meng, University of Illinois at Urbana-Champaign*, USA; Cong Xie, University of Illinois at Urbana-Champaign*, USA; Indranil Gupta, University of Illinois at Urbana-Champaign, USA, [indy@illinois.edu](indy@illinois.edu).  \nPermission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Ab","cbCaitjtcR3K64XW","https://ap.wps.com/l/cbCaitjtcR3K64XW","pdf",1646796,1,37,"English","en",105,"# Introduction\n## Motivation: memory limits in large ML models\n## Small clusters of memory-constrained devices\n# Baechi system overview\n## Algorithmic placement for training graphs\n## Integration with TensorFlow and PyTorch\n# Experimental results\n## Placement plan generation speed\n## Training step time comparison with expert placements\n# Theoretical analysis\n## Constant-factor optimality bounds","[{\"question\":\"为什么需要对机器学习训练图做设备放置（device placement）？\",\"answer\":\"当设备内存受限或模型很大时，模型可能无法在单个设备上完整训练，必须把训练图切分到多台设备上并决定每个算子放在哪个设备上。\"},{\"question\":\"Baechi与现有基于学习的放置方法相比有什么优势？\",\"answer\":\"Baechi采用算法化方案生成放置计划，实验表明其生成速度比现有学习式方法快654×到206K×，且训练步时间接近专家放置。\"},{\"question\":\"Baechi的结果在TensorFlow和PyTorch上分别如何表现？\",\"answer\":\"文中指出，在PyTorch中，Baechi放置后的step（训练）时间与专家放置可比；在TensorFlow中，仅比专家放置差最多6.2%。\"}]","Baechi - Fast Device Placement of Machine Learning Graphs | PDF",1785679367,93,{"code":4,"msg":31,"data":32},"ok",{"site_id":24,"language":23,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"baechi-fast-device-placement-of-machine-learning-graphs","",{"@graph":36,"@context":85},[37,54,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":53},"https://docshare.wps.com/document/baechi-fast-device-placement-of-machine-learning-graphs/117753/",4,{"url":52,"name":13,"@type":55,"author":56,"headline":13,"publisher":58,"fileFormat":61,"inLanguage":23,"description":14,"dateModified":62,"datePublished":62,"encodingFormat":61,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":57},"Person",{"url":41,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-08-02",true,{"@type":65,"interactionType":66,"userInteractionCount":4},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"为什么需要对机器学习训练图做设备放置（device placement）？","Question",{"text":75,"@type":76},"当设备内存受限或模型很大时，模型可能无法在单个设备上完整训练，必须把训练图切分到多台设备上并决定每个算子放在哪个设备上。","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"Baechi与现有基于学习的放置方法相比有什么优势？",{"text":80,"@type":76},"Baechi采用算法化方案生成放置计划，实验表明其生成速度比现有学习式方法快654×到206K×，且训练步时间接近专家放置。",{"name":82,"@type":73,"acceptedAnswer":83},"Baechi的结果在TensorFlow和PyTorch上分别如何表现？",{"text":84,"@type":76},"文中指出，在PyTorch中，Baechi放置后的step（训练）时间与专家放置可比；在TensorFlow中，仅比专家放置差最多6.2%。","https://schema.org",{"og:url":52,"og:type":87,"og:title":13,"og:site_name":59,"og:description":14},"article",{"robots":89,"canonical":52},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":53,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":46,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]