Collaborative Research: SHF: Medium: A hardware-software co-design approach for high-performance in-memory analytic data processing

协作研究：SHF：中：用于高性能内存分析数据处理的硬件软件协同设计方法

• 批准号: 2407690
• 负责人: Jignesh Patel
• 金额: $ 40万
• 依托单位: Carnegie-Mellon 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=2407690&HistoricalAwards=false
• 关键词: Collaborative; Research; SHF; Medium; hardware

Database analytics is crucial for decision-making across various industries and fields of inquiry. However, there is a challenge: analyzing large amounts of data using traditional methods takes more time and money as the volume of data grows. Resolving this issue is vital for enterprises to stay competitive through fast and accurate data-driven decision-making and to keep up with rapid growth in data volumes. In the past, hardware and software for analytics could be developed separately, benefiting from Moore's Law (doubling of transistor density with each transistor generation) and Dennard scaling (which allowed Moore's Law to proceed without increasing power density). This scaling allowed the industry to steadily improve the performance of general-purpose hardware. However, we have now reached the physical limits of these trends and need new hardware approaches to enhance analytics speed and efficiency. Furthermore, processing units are now much faster than memory, so applications with large volumes of data are increasingly bottlenecked by memory accesses. This research, therefore, focuses on memory devices, in particular designing ``intelligent'' memory capable of computing results near the stored data, and proposes a solution by redesigning both hardware and software components of an analytics pipeline to work synergistically, addressing the data analytics performance issue from the ground up. This innovative approach has the potential to significantly improve the efficiency of data analytics. The project is a collaboration between one database and software researcher at the University of Wisconsin-Madison (UW) and two computer architecture and systems researchers at Cornell University and the University of Virginia (UVA). The project is organized into four thrusts. Thrust 1 aims to develop mechanisms for in-place data analytics query processing on the dynamic random access memory (DRAM) side and explore the synergies between intelligent DRAM and other processing units. Thrust 2 focuses on processing in memory (PIM) designs for static random-access memory (SRAM)-based caches, exploring associative processing (AP) and its applicability to data analytics workloads. Thrust 3 takes a holistic approach to accelerate analytics queries across both SRAM and DRAM-based PIM designs. It proposes a Domain-Specific Language (DSL)-based approach using an operational algebra, decomposing queries into a dataflow graph and optimizing their execution across different PIMs and the Central Processing Unit (CPU). Finally, Thrust 4 addresses the need for evaluation frameworks in the database-hardware co-design approach by developing a simulation infrastructure and benchmarks that can be used by the broader architecture and database research communities. Besides training students involved in this project across the hardware-software boundaries, the project will also support outreach efforts in entrepreneurship education at UW. Additionally, there are outreach plans to Native American high school students through an effort at Cornell, and at UVA, the project will contribute to ongoing efforts to build long-term collaborations with Historically Black Colleges and Universities (HBCUs) and Minority-Serving Institutions (MSIs) in the Virginia area.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.

数据库分析对于各个行业和调查领域的决策至关重要。然而，存在一个挑战：随着数据量的增长，使用传统方法分析大量数据需要花费更多的时间和金钱。解决这个问题对于企业通过快速准确的数据驱动决策保持竞争力以及跟上数据量的快速增长至关重要。在过去，用于分析的硬件和软件可以单独开发，受益于摩尔定律（每一代晶体管的晶体管密度增加一倍）和Dennard缩放（允许摩尔定律在不增加功率密度的情况下进行）。这种扩展使业界能够稳步提高通用硬件的性能。然而，我们现在已经达到了这些趋势的物理极限，需要新的硬件方法来提高分析速度和效率。此外，处理单元现在比内存快得多，因此具有大量数据的应用程序越来越多地受到内存访问的影响。因此，这项研究的重点是存储设备，特别是设计能够在存储数据附近计算结果的“智能”存储器，并通过重新设计分析管道的硬件和软件组件来协同工作，从根本上解决数据分析性能问题，提出了一种解决方案。这种创新方法有可能显著提高数据分析的效率。该项目是威斯康星大学麦迪逊分校（UW）的一名数据库和软件研究人员与康奈尔大学和弗吉尼亚大学（UVA）的两名计算机体系结构和系统研究人员之间的合作。该项目分为四个重点。Thrust 1旨在开发动态随机存取存储器（DRAM）端的就地数据分析查询处理机制，并探索智能DRAM和其他处理单元之间的协同作用。Thrust 2专注于基于静态随机存取存储器（SRAM）的缓存的内存处理（PIM）设计，探索关联处理（AP）及其对数据分析工作负载的适用性。Thrust 3采用整体方法来加速基于SRAM和DRAM的PIM设计中的分析查询。它提出了一种基于领域特定语言（DSL）的方法，使用运算代数，将查询分解为一个简单的图，并优化它们在不同的PIM和中央处理器（CPU）上的执行。最后，Thrust 4通过开发可被更广泛的架构和数据库研究社区使用的仿真基础设施和基准来解决数据库硬件协同设计方法中对评估框架的需求。除了培训参与该项目的学生跨越硬件和软件的界限，该项目还将支持在华盛顿大学创业教育的推广工作。此外，通过康奈尔大学和弗吉尼亚大学的努力，该项目将有助于与历史上的黑人学院和大学（HBCU）以及少数民族服务机构（MSI）建立长期合作关系该奖项反映了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。