[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84790-en":3,"doc-seo-84790-105":29,"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":20,"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":13,"seo_description":14,"update_tm":27,"read_time":28},84790,5909877438554,"Maeve","https://ap-avatar.wpscdn.com/avatar/5600025385ad2bf12a7?_k=1778553567797529272",8,"Research & Report","FAST: A Holistic Framework for Optimizing Memory-I O, Computation, and Sampling in Temporal GNN Training","Temporal Graph Neural Networks (TGNNs) are used to learn from dynamic graphs in recommendation, social network analysis, and traffic forecasting, yet scaling training on large graphs is limited by three coupled bottlenecks: memory I/O, irregular GPU computation, and costly temporal neighbor sampling. FAST provides an end-to-end optimization framework that jointly accelerates sampling, memory I/O, and computation. It introduces SlimCache for within-batch compression and cross-batch caching, and thread-efficient temporal sparse graph operators to improve GPU locality and reduce aggregation and edge softmax latency, plus a topology-aware sampling strategy. Experiments on large real-world datasets show 2.1× average speedup (up to 4.7×) without accuracy loss.","arXiv :2607 .05095v 1 [ cs .LG] 6 Jul 2026  \nFAST: A Holistic Framework for Optimizing Memory-I/O, Computation, and Sampling in Temporal GNN Training  \nYushu Cai  \nXidian University Guangzhou, Guangdong, China [yushucai@stu.xidian.edu.cn](yushucai@stu.xidian.edu.cn)  \nQingrui Zhu  \nXidian University Guangzhou, Guangdong, China [zhuqingrui@stu.xidian.edu.cn](zhuqingrui@stu.xidian.edu.cn)  \nLei Liu  \nXidian University Guangzhou, Guangdong, China [liulei2303@stu.xidian.edu.cn](liulei2303@stu.xidian.edu.cn)  \nKai Sheng  \nXidian University Guangzhou, Guangdong, China [kaisheng@xidian.edu.cn](kaisheng@xidian.edu.cn)  \nHao Chen  \nHunan University Changsha, Hunan, China [haochen@hnu.edu.cn](haochen@hnu.edu.cn)  \nXin He  \nXidian University Guangzhou, Guangdong, China [hexin@xidian.edu.cn](hexin@xidian.edu.cn)  \nAbstract  \nTemporal Graph Neural Networks (TGNNs) are widely used for learning from dynamic graphs in applications such as recommendation, social network analysis, and traffic forecasting. However, scaling TGNN training to large dynamic graphs remains challenging due to three intertwined bottlenecks: memory I/O, irregular computation, and temporal neighbor sampling. Existing systems often optimize these stages in isolation, leaving substantial performance headroom on the table. We present FAST, a holistic framework that accelerates end-to-end TGNN training by jointly optimizing sampling, memory I/O, and computation. FAST introduces SlimCache, which exploits within-batch compression and cross-batch caching to reduce host-device data movement under limited GPU memory budgets. It further designs thread-efficient graph operators tailored to sparse temporal subgraphs, improving GPU cache locality and reducing the latency of aggregation and edge softmax. In addition, FAST employs a topology-aware sampling strategy that improves CPU cache locality and accelerates temporal neighbor sampling. Extensive experiments on real-world large dynamic graphs show that FAST achieves an average of 2.1× (up to 4.7×) speedup over stateof-the-art systems without sacrificing model accuracy.The code of FAST is publicly available at [https://github.com/NoneBone/FAST](https://github.com/NoneBone/FAST).  \n1 Introduction  \nDynamic graphs naturally model timestamped interactions between real-world entities, with nodes representing entities and timestamped edges capturing their temporal relationships. Learning from such data is fundamental to applications including recommendation, social network analysis, traffic forecasting, and fraud detection [9, 29]. Temporal Graph Neural Networks (TGNNs) [18, 20, 25] have become the dominant approach by capturing evolving message dynamics through recursive temporal message passing, combining temporal neighbor sampling with temporally encoded neighborhood aggregation to learn expressive, time-aware node representations. TGNNs excel at modeling evolving relationships, such as user interest drift in streaming services [17], enabling accurate future behavior prediction.  \nTo scale TGNNs to large dynamic graphs, existing systems typically adopt a timestamp-ordered batch training paradigm [18, 25] . The input graph is partitioned into chronologically ordered batches,  \nAccepted to the 55th International Conference on Parallel Processing (ICPP 2026) .  \neach containing a sequence of interactions. For each batch, the system (1) performs temporal neighbor sampling to construct a computation subgraph,(2) reads node states and feature data from host memory and transfers them to the GPU, and (3) executes forward and backward propagation to update representations and optimize model parameters. While effective in small-scale settings, this pipeline suffers from the following three severe bottlenecks when scaled to large graphs.  \nMemory I/O Bottleneck. A fundamental bottleneck is the massive data access overhead between host and GPU memory. Each mini-batch requires copying dynamically sampled subgraphs and associated features, saturati","cbCaiu1DuRfp6qtH","https://ap.wps.com/l/cbCaiu1DuRfp6qtH","pdf",3198339,1,11,"English","en",105,"# Abstract\n# Introduction\n## Memory I/O Bottleneck\n## Computation Bottleneck\n## Sampling Bottleneck","[{\"question\":\"What problem does FAST address in Temporal GNN training?\",\"answer\":\"FAST targets the difficulty of scaling Temporal Graph Neural Network (TGNN) training on large dynamic graphs, where memory I/O, irregular computation, and temporal neighbor sampling form intertwined bottlenecks.\"},{\"question\":\"How does SlimCache improve training efficiency?\",\"answer\":\"SlimCache reduces host–device data movement by combining within-batch compression with cross-batch caching, using careful handling of node and edge reuse patterns under limited GPU memory budgets.\"},{\"question\":\"What improvements does FAST make beyond memory I/O?\",\"answer\":\"FAST also designs thread-efficient graph operators for sparse temporal subgraphs to improve GPU cache locality and lowers latency for aggregation and edge softmax, and it uses a topology-aware sampling strategy to accelerate temporal neighbor sampling while improving CPU cache locality.\"}]",1784198252,28,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":27},"fast-a-holistic-framework-for-optimizing-memory-io-computation-and-sampling-in-temporal-gnn-training","",{"@graph":35,"@context":85},[36,53,68],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/fast-a-holistic-framework-for-optimizing-memory-io-computation-and-sampling-in-temporal-gnn-training/84790/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What problem does FAST address in Temporal GNN training?","Question",{"text":75,"@type":76},"FAST targets the difficulty of scaling Temporal Graph Neural Network (TGNN) training on large dynamic graphs, where memory I/O, irregular computation, and temporal neighbor sampling form intertwined bottlenecks.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does SlimCache improve training efficiency?",{"text":80,"@type":76},"SlimCache reduces host–device data movement by combining within-batch compression with cross-batch caching, using careful handling of node and edge reuse patterns under limited GPU memory budgets.",{"name":82,"@type":73,"acceptedAnswer":83},"What improvements does FAST make beyond memory I/O?",{"text":84,"@type":76},"FAST also designs thread-efficient graph operators for sparse temporal subgraphs to improve GPU cache locality and lowers latency for aggregation and edge softmax, and it uses a topology-aware sampling strategy to accelerate temporal neighbor sampling while improving CPU cache locality.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"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":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":45,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":45,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":45,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":45,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":45,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]