CAREER: Architecting a Hardware-Software Co-Designed Data Management System for Heterogeneous Memory Computers

职业：为异构内存计算机构建软硬件协同设计的数据管理系统

• 批准号: 2144883
• 负责人: Jason Lowe-Power
• 金额: $ 60.51万
• 依托单位: University of California-Davis
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-07-01 至 2027-06-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2144883&HistoricalAwards=false
• 关键词: CAREER; Architecting; Hardware; Software; Co

Data processing is integral to society, and the amount of data our computing systems are asked to process is growing rapidly. For example, some of today’s artificial intelligence (AI) and machine learning (ML) workloads have billions of parameters requiring terabytes of working memory for training and inference. To efficiently process this deluge of data, the data must first be moved from distant random-access memory to memory closer to the processor. In today’s systems, this data movement itself is one of the most expensive aspects of computation in terms of both performance and energy/power. To improve the efficiency of computing systems, this CAREER project aims to answer the question “where should data be placed and when should it move from one memory type to another?” The main insight driving this CAREER proposal is that bridging the semantic gap between the application which knows how the data is used and the hardware which directly interacts with the data will enable a new set of optimizations. Leveraging this insight, this CAREER project will develop a new interface for data movement which separates the data movement mechanism from the data movement policies. This separation of concerns will enable programmer hints and runtime-directed data movement which will significantly increase the performance and efficiency of heterogeneous memory systems.This CAREER project will improve the efficiency and performance of a wide variety of applications including AI/ML which increasingly depend on both high memory performance and large memory capacities. To broaden the impact, the techniques developed in this project will be realized in software, as a hardware-software codesign, and as hardened logic. Further, the software models, systems, and hardware prototypes will be released open source as they are developed. This CAREER project will also lay the foundation for a sustainable undergraduate research internship program at UC Davis. This program will give experience with open-source software development, career development opportunities, and hands-on hardware-software co-design experience.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.

数据处理是社会不可或缺的一部分，我们的计算系统需要处理的数据量正在快速增长。例如，当今的一些人工智能 (AI) 和机器学习 (ML) 工作负载具有数十亿个参数，需要 TB 的工作内存来进行训练和推理。为了有效地处理大量数据，必须首先将数据从遥远的随机存取存储器移动到更靠近处理器的存储器。在当今的系统中，就性能和能源/功率而言，这种数据移动本身是计算中最昂贵的方面之一。为了提高计算系统的效率，这个职业项目旨在回答“数据应该放在哪里以及何时应该从一种内存类型移动到另一种内存类型？”的问题。推动这一职业提案的主要见解是，弥合知道如何使用数据的应用程序与直接与数据交互的硬件之间的语义差距将实现一组新的优化。利用这种洞察力，该 CAREER 项目将开发一个新的数据移动界面，将数据移动机制与数据移动策略分开。这种关注点分离将使程序员提示和运行时引导的数据移动成为可能，从而显着提高异构内存系统的性能和效率。这个职业项目将提高各种应用程序的效率和性能，包括日益依赖于高内存性能和大内存容量的人工智能/机器学习。为了扩大影响，该项目开发的技术将以软件、硬件-软件协同设计和硬化逻辑的形式实现。此外，软件模型、系统和硬件原型将在开发时开源发布。该职业项目还将为加州大学戴维斯分校可持续的本科生研究实习项目奠定基础。该计划将提供开源软件开发经验、职业发展机会以及软硬件协同设计实践经验。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力优势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

Efficient Large Scale DLRM Implementation On Heterogeneous Memory Systems

异构内存系统上的高效大规模 DLRM 实现

• 发表时间: 2023
• 期刊: Proceedings
• 作者: Hildebrand, Mark;Lowe-Power, Jason;Akella, Venkatesh
• 通讯作者: Akella, Venkatesh

SoK: Limitations of Confidential Computing via TEEs for High-Performance Compute Systems

• DOI: 10.1109/seed55351.2022.00018
• 发表时间: 2022-09
• 期刊: 2022 IEEE International Symposium on Secure and Private Execution Environment Design (SEED)
• 作者: Ayaz Akram;V. Akella;S. Peisert;Jason Lowe-Power