CNS: CORE: Small: Scaling Graph Machine Learning Workloads on Modern Storage

CNS：核心：小型：在现代存储上扩展图机器学习工作负载

• 批准号: 2237193
• 负责人: Vasiliki Kalavri
• 金额: $ 46.98万
• 依托单位: Trustees of Boston University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-10-01 至 2026-09-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2237193&HistoricalAwards=false
• 关键词: CNS; CORE; Small; Scaling; Graph

This project designs and develops GNNSuite, a novel unified framework for graph machine learning that leverages emerging storage technology and enables users to deploy large graph neural network (GNN) models on a single commodity machine with high efficiency and low cost. The project's core novelties include methods and software tools for training and serving GNN models on larger-than-memory graphs without affecting the accuracy of downstream tasks. GNNSuite targets classification and prediction use cases, such as chemical synthesis, recommender systems, fraud detection, and large-scale distributed service management.The project involves three sets of tasks that address challenges in large-scale GNN training and inference. First, the investigator designs a training module that employs single-pass in-situ neighborhood sampling on disk-resident data and prefetching optimizations that maximize resource utilization. Second, she develops a three-layered data organization approach that spans memory and secondary storage to facilitate scalable GNN inference on graph streams. Third, the investigator designs and implements a continual training module that leverages experience replay and modern storage capabilities to provide efficient incremental model updates. Project results have the potential to radically increase the accessibility of GNNs, accelerate the integration of graph machine learning tasks in online business analytics pipelines, and inform future research on the next generation of computational storage.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

该项目设计并开发了GNNSuite，这是一种新颖的图形机器学习统一框架，利用新兴的存储技术，使用户能够在单个商品机器上以高效率和低成本部署大型图形神经网络（GNN）模型。该项目的核心创新包括在不影响下游任务准确性的情况下，在大于内存的图上训练和服务GNN模型的方法和软件工具。GNNSuite针对分类和预测用例，如化学合成、推荐系统、欺诈检测和大规模分布式服务管理。该项目涉及三组任务，以解决大规模GNN训练和推理中的挑战。首先，研究者设计了一个训练模块，该模块采用磁盘驻留数据的单次原位邻域采样和预取优化，最大限度地提高资源利用率。其次，她开发了一种跨越内存和二级存储的三层数据组织方法，以促进图流上可扩展的GNN推理。第三，研究者设计并实现了一个持续的培训模块，该模块利用经验回放和现代存储能力来提供有效的增量模型更新。项目结果有可能从根本上增加gnn的可访问性，加速在线业务分析管道中图形机器学习任务的集成，并为下一代计算存储的未来研究提供信息。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。