CRII: SHF: A Memory-Centric Hardware Accelerator for Large Scale Data Clustering

聚类是几乎所有科学和工程学科中分析数据的重要工具。美国国家科学院（NAS）最近公布了“统计数据分析的七大巨头”，其中数据聚类起着核心作用。该报告还强调，需要更多可扩展的解决方案，以实现未来大规模数据分析的时间和空间聚类。因此，硬件和软件的创新，可以显着提高能源效率和性能的数据聚类技术是必要的，使未来的大规模数据分析实用。为了实现这一目标，所提出的研究展示了未来解决数据聚类问题的完全不同的愿景，其中大规模聚类问题被映射到以内存为中心的非冯·诺依曼计算基底上，并在数据阵列内就地解决，性能和能效比当代计算机系统高出几个数量级。拟议的项目将利用电阻式随机存取技术的最新发展存储器（RRAM）和算法方法，用于在大规模并行框架（诸如位串行秩序滤波器）内重新制定聚类问题，以构建用于未来聚类应用的极低功率和快速存储器基板。在软件层面，将开发新的算法，以将科学和工程领域的不同类型的异构数据聚类问题（包括数值和非数值数据点）映射到拟议的忆阻加速器上。编程模型，软件模块和应用程序库的硬件-软件协同设计，动态资源管理和内存分配将开发给用户控制的数据聚类过程在运行时。在硬件层面上，我们将研究用于优化由新型电阻单元和互连网络构造的存储器模块的功率和性能的技术。架构和软件创新将通过发表的论文传播到更广泛的研究界，以及在非冯·诺依曼计算机系统中新兴的原位计算平台和软硬件接口的教程。这个项目的教育部分将涉及集成的细胞结构，互连网络，分层软件接口，并控制策略到一个先进的计算机体系结构课程的教学大纲。

Clustering is a crucial tool for analyzing data in virtually every scientific and engineering discipline. The U.S. National Academy of Sciences (NAS) has recently announced "the seven giants of statistical data analysis" in which data clustering plays a central role. This report also emphasizes that more scalable solutions are required to enable time and space clustering for the future large scale data analyses. As a result, hardware and software innovations that can significantly improve energy-efficiency and performance of the data clustering techniques are necessary to make the future large scale data analysis practical. To this goal, the proposed research demonstrates a radically different vision of solving data clustering problems in the future, where large-scale clustering problems are mapped onto a memory-centric, non-Von Neumann computation substrate and solved in situ within the data arrays, with orders of magnitude greater performance and energy efficiency than contemporary computer systems.The proposed project will leverage recent developments in resistive random access memory (RRAM) and algorithmic approaches for reformulating clustering problems within massively parallel frameworks, such as bit serial rank order filters, to build an extremely low power and fast memory substrate for future clustering applications. At the software level, novel algorithms will be developed to map different types of heterogeneous data clustering problems (including numerical and non-numerical data points) from scientific and engineering domains onto the proposed memristive accelerator. Programming models, software modules, and application libraries for hardware-software co-design, dynamic resource management, and memory allocation will be developed to give the user control of the data clustering process at runtime. At the hardware level, we will investigate techniques for optimizing power and performance of the memory modules constructed from novel resistive cells and interconnection networks. Architecture and software innovations will be disseminated to the broader research community through published papers, as well as tutorials on the emerging in situ computing platforms and software-hardware interfaces in non-Von Neumann computer systems. The educational component of this project will involve integrating the cell structure, the interconnection networks, the hierarchical software interface, and the control policies into the syllabus of an advanced computer architecture course.