SPX: Collaborative Research: Cross-stack Memory Optimizations for Boosting I/O Performance of Deep Learning HPC Applications

SPX：协作研究：用于提升深度学习 HPC 应用程序 I/O 性能的跨堆栈内存优化

• 批准号: 2318628
• 负责人: Yue Cheng
• 金额: $ 32.06万
• 依托单位: University of Virginia Main Campus
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-10-01 至 2024-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2318628&HistoricalAwards=false
• 关键词: SPX; Collaborative; Research; Cross; stack

New computing applications are emerging in smart networks, scientific explorations, business management, security, and healthcare. These applications depend on very large amounts of data. This data must be used in a fast and efficient manner. The use of large supercomputers to analyze such data is on the rise. The techniques they use are referred to as deep learning (DL) high-performance computing (HPC). Researchers are using DL HPC to make sense of this flood of data and obtain useful information. To do this they must redesign HPC systems. A key challenge is how to use resources such as data storage and computer memory at a huge scale. This project will build Metis, a high-performance data storage system that uses new, end-to-end, hardware-supported memory and storage design to meet the needs of DL HPC applications. The goal is to satisfy the challenge posed by increasing data management performance for next-generation supercomputers. The project will connect several different computing communities and increase interactions among them. The project includes educational and engagement activities which will greatly increase the community's understanding of HPC systems. These activities include broadening participation activities to attract and retain new students. Special emphasis will be given to students from underrepresented groups. The project will encourage student interest in design and research in large-scale computing systems design.This project brings together researchers in micro-architecture, distributed computing systems, namely cloud and HPC systems, storage systems, and power/energy modeling to boost DL HPC data processing performance. The research will yield a fundamentally new software-hardware co-designed memory compression technique that transparently compresses DL application memories with negligible runtime performance overhead. Metis will leverage the novel compression substrate to enable a distributed, intelligent, operating-system-level data cache that effectively exploits the physical memory freed via program-memory compression. The developed techniques will open doors for innovative HPC and scientific applications in a broad range of disciplines, which have not been previously possible. Metis' focus on addressing the challenges of increasing performance in the Exascale era, along with engaging researchers from multiple areas, aligns it very well with the goals and objectives of the SPX program. Additionally, the research will also create new knowledge on design principles of memory compression, and yield insights to provide seamless integration of DL applications into the next-generation DL-aware supercomputer infrastructure.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.

新的计算应用程序正在智能网络，科学探索，业务管理，安全和医疗保健中出现。这些应用程序取决于大量数据。这些数据必须以快速有效的方式使用。使用大型超级计算机来分析此类数据正在上升。它们使用的技术称为深度学习（DL）高性能计算（HPC）。研究人员正在使用DL HPC来理解这一数据泛滥并获得有用的信息。为此，它们必须重新设计HPC系统。一个关键的挑战是如何使用大规模使用数据存储和计算机存储器之类的资源。该项目将构建METIS，这是一种高性能数据存储系统，该系统使用新的，端到端的，硬件支持的内存和存储设计来满足DL HPC应用程序的需求。目的是通过提高下一代超级计算机的数据管理绩效来满足挑战。该项目将连接几个不同的计算社区并增加它们之间的互动。该项目包括教育和参与活动，这些活动将大大提高社区对HPC系统的理解。这些活动包括扩大参与活动以吸引和留住新学生。来自代表性不足的团体的学生将特别重视。该项目将鼓励学生对大规模计算系统设计的设计和研究的兴趣。本项目将微观架构，分布式计算系统的研究人员汇集在一起​​，即云和HPC系统，存储系统以及功率/能量建模，以提高DL HPC数据处理性能。这项研究将产生一种新的软件硬件共同设计的内存压缩技术，该技术可以透明地压缩DL应用程序记忆，并用可忽略不计的运行时性能开销。 METIS将利用新颖的压缩基板来启用分布式，智能，操作系统级的数据缓存，该数据缓存有效利用通过程序 - 记忆压缩释放的物理内存。开发的技术将在广泛的学科中为创新的HPC和科学应用打开大门，这是不可能的。梅蒂斯（Metis）专注于解决Exascale时代提高绩效的挑战，以及来自多个领域的研究人员，它与SPX计划的目标和目标非常一致。此外，这项研究还将创建有关记忆压缩设计原理的新知识，并产生见解，以将DL应用程序无缝整合到下一代DL-WAWARE SUPERCOUTATER基础架构中。该奖项反映了NSF的法定任务，并通过使用基金会的知识优点和广泛的影响来通过评估来进行评估，并认为有价值。

项目成果

SHADE: Enable Fundamental Cacheability for Distributed Deep Learning Training

• 发表时间: 2023
• 作者: Redwan Ibne Seraj Khan;Ahmad Hossein Yazdani;Yuqi Fu;Arnab K. Paul;Bo Ji;Xun Jian;Yue Cheng;A. R. Butt

InfiniStore: Elastic Serverless Cloud Storage

InfiniStore：弹性无服务器云存储

• DOI: 10.14778/3587136.3587139
• 发表时间: 2023
• 期刊: Proceedings of the VLDB Endowment
• 影响因子: 2.5
• 作者: Zhang, Jingyuan;Wang, Ao;Ma, Xiaolong;Carver, Benjamin;Newman, Nicholas John;Anwar, Ali;Rupprecht, Lukas;Tarasov, Vasily;Skourtis, Dimitrios;Yan, Feng
• 通讯作者: Yan, Feng