FoMR: Using Machine Learning to Design Next Generation Caches and Data Prefetchers

FoMR：使用机器学习设计下一代缓存和数据预取器

• 批准号: 1823546
• 负责人: Calvin Lin
• 金额: $ 22.5万
• 依托单位: University of Texas at Austin
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2018
• 资助国家: 美国
• 起止时间: 2018-10-01 至 2022-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1823546&HistoricalAwards=false
• 关键词: FoMR; Using; Machine; Learning; Design

While the general public typically focuses on processor clock speed as a measure of microprocessor performance, it is often the memory system that limits overall system performance, particularly for today's data-hungry uses of computing. This project explores a transformative approach to designing computer hardware, particularly the memory system: Rather than rely solely on human insight and intuition, this approach adapts and leverages machine learning techniques to explore larger and richer design spaces in a more thorough and systematic manner. This approach enables hardware designers to consider more complex design factors than are currently possible. This project consists of two phases. The first phase seeks to dramatically improve traditional memory system components, such as cache replacement policies and data prefetchers, which have been heavily studied but are now ripe for innovation with the use of machine learning. The second phase considers complicating factors such as criticality, the interaction among these components, and the use of new memory technologies.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.

虽然一般公众通常关注处理器时钟速度作为微处理器性能的衡量标准，但通常是存储器系统限制了整体系统性能，特别是对于当今数据饥渴的计算用途。 该项目探索了一种变革性的方法来设计计算机硬件，特别是内存系统：这种方法不是仅仅依赖于人类的洞察力和直觉，而是适应并利用机器学习技术，以更全面和系统的方式探索更大，更丰富的设计空间。 这种方法使硬件设计人员能够考虑比目前可能的更复杂的设计因素。 该项目包括两个阶段。 第一阶段旨在大幅改进传统的内存系统组件，例如缓存替换策略和数据预取器，这些组件已经得到了大量研究，但现在已经成熟，可以使用机器学习进行创新。该奖项反映了NSF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。

项目成果

Temporal Prefetching without the Off-Chip Metadata

没有片外元数据的时间预取

• 发表时间: 2019
• 期刊: International Symposium on Microrchitecture
• 作者: Hao Wu, Krishnendra Nathella

Efficient Metadata Management for Irregular Data Prefetching

• DOI: 10.1145/3307650.3322225
• 发表时间: 2019-06
• 期刊: 2019 ACM/IEEE 46th Annual International Symposium on Computer Architecture (ISCA)
• 作者: Hao Wu;Krishnendra Nathella;Dam Sunwoo;Akanksha Jain;Calvin Lin

A hierarchical neural model of data prefetching

• DOI: 10.1145/3445814.3446752
• 发表时间: 2021-04
• 期刊: Proceedings of the 26th ACM International Conference on Architectural Support for Programming Languages and Operating Systems
• 作者: Zhan Shi;Akanksha Jain;Kevin Swersky;Milad Hashemi;Parthasarathy Ranganathan;Calvin Lin

A Fast Work-Efficient SSSP Algorithm for GPUs

一种适用于 GPU 的快速、高效的 SSSP 算法

• DOI: 10.1145/3437801.3441605
• 发表时间: 2021
• 期刊: Annual Symposium on Principles and Practice of Parallel Programming
• 作者: Wang, Kai;Fussell, Don;Lin, Calvin
• 通讯作者: Lin, Calvin

Practical Temporal Prefetching With Compressed On-Chip Metadata

• DOI: 10.1109/tc.2021.3065909
• 发表时间: 2022-11
• 期刊: IEEE Transactions on Computers
• 影响因子: 3.7
• 作者: Hao Wu;Krishnendra Nathella;Matthew Pabst;Dam Sunwoo;Akanksha Jain;Calvin Lin