SHF: Medium: Breaking the Physical Divide between Computation and NAND-Flash Storage

SHF：媒介：打破计算和 NAND 闪存存储之间的物理鸿沟

• 批准号: 1302557
• 负责人: Mahmut Kandemir
• 金额: $ 80万
• 依托单位: Pennsylvania State Univ University Park
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2013
• 资助国家: 美国
• 起止时间: 2013-07-01 至 2017-06-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1302557&HistoricalAwards=false
• 关键词: SHF; Medium; Breaking; Physical; Divide

Almost all application segments today experience data explosion, meaning that they need to store, access, manipulate and transform extremely large amounts of data stored in different mediums in a fashion that is simultaneously performance-aware and energy-aware. These data-hungry market segments include (i) consumer applications in the mobile and home electronics segment, (ii) desktop applications that are providing rich content and user experience, (iii) scientific applications that generate petabytes of data for analyzing experiments and real-world phenomena on temporal and spatial scales unheard of before, (iv) enterprise applications which tirelessly store all kinds of data/knowledge for auditability, analytics, and optimization, (v) datacenters and cloud platforms which use storage to hold large virtual machine images of the workloads for consolidation across different servers, (vi) Internet services and social networking platforms which need to store, track and manage user patterns, and (vii) cyber-physical applications which continuously sense and store physical world data for real-time analytics and control. Current computer infrastructures are poorly equipped to cope with this data demand. The primary reason for this is the inherent physical divide between computation and storage. While both computation and storage technologies have undergone tremendous improvements in the last decades, the interactions and interfaces between them have not, thereby limiting the performance of critical data-intensive applications. If not addressed in a timely fashion, this problem has the potential to slow down scientific discoveries and engineering breakthroughs. This project addresses the data management problem by breaking the physical divide between computation and NAND-flash storage. Doing so can potentially allow the communication bandwidth between computation and storage to scale together with the parallelism-driven scaling of both computation resources and storage resources. It can also allow each to become more aware of the intentions and operations of the other, opening a wide spectrum of possibilities in more efficiently managing storage. This will in turn allow better co-design, co-management, and co-evolution of the two for better scalability in the future, as applications start imposing even more stringent computing and storage demands. Specifically, this project investigates three main strategies for bridging the physical divide between compute and NAND-flash storage. The first strategy enables better cooperation between flash storage and host; the second strategy elevates NAND-flash storage to directly interface with the processors, similar to main memory DIMMs (dual inline memory modules) interfacing to the on-chip cores through memory controllers; and the last strategy explores different placement options for tighter integration of NAND-flash storage with computational resources. The broader impacts of this research include student training, participation of under-represented groups, recruiting workshops, incorporation of educational modules into existing and future courses, and public domain simulation tools. Further, through the Visit In Engineering Weekend (VIEW) program, the project fosters interest in computer science and engineering. The project provides hands-on-design activities to motivate the VIEW participants in new areas of computer science and engineering related to storage system and data management.

如今，几乎所有的应用程序都经历了数据爆炸，这意味着它们需要以同时具有性能感知和能量感知的方式存储、访问、操作和转换存储在不同介质中的大量数据。 这些数据饥渴的细分市场包括（i）移动的和家用电子产品细分市场中的消费者应用，（ii）提供丰富内容和用户体验的桌面应用，（iii）生成PB级数据的科学应用，用于在前所未有的时间和空间尺度上分析实验和真实世界现象，（iv）不知疲倦地存储各种数据/知识以实现可重复性的企业应用，分析和优化，（v）数据中心和云平台，使用存储来保存工作负载的大型虚拟机映像，以便跨不同服务器进行整合，（vi）互联网服务和社交网络平台，需要存储，跟踪和管理用户模式，以及（vii）网络物理应用程序，持续感知和存储物理世界数据，以进行实时分析和控制。目前的计算机基础设施装备不足，无法科普这种数据需求。其主要原因是计算和存储之间固有的物理划分。虽然计算和存储技术在过去几十年中经历了巨大的改进，但它们之间的交互和接口却没有，从而限制了关键数据密集型应用程序的性能。如果不及时解决，这个问题有可能减缓科学发现和工程突破。该项目通过打破计算和NAND闪存存储之间的物理鸿沟来解决数据管理问题。这样做可以潜在地允许计算和存储之间的通信带宽与计算资源和存储资源的并行驱动的缩放一起缩放。它还可以让每个人更加了解另一个人的意图和操作，从而为更有效地管理存储提供广泛的可能性。这反过来又将允许更好的协同设计，协同管理和协同进化，以便在未来实现更好的可扩展性，因为应用程序开始施加更严格的计算和存储需求。具体而言，该项目研究了三种主要策略，用于弥合计算和NAND闪存存储之间的物理鸿沟。第一种策略实现了闪存和主机之间更好的合作;第二种策略提升了NAND闪存存储，使其直接与处理器接口，类似于通过存储器控制器与片上内核接口的主存储器DIRECT（双列直插式存储器模块）;最后一种策略探索了不同的布局选项，以实现NAND闪存存储与计算资源的更紧密集成。这项研究的更广泛的影响包括学生培训，代表性不足的群体的参与，招聘研讨会，将教育模块纳入现有和未来的课程，以及公共领域的模拟工具。此外，通过参观工程周末（VIEW）计划，该项目培养了对计算机科学和工程的兴趣。该项目提供动手设计活动，以激励VIEW参与者在与存储系统和数据管理相关的计算机科学和工程新领域。