Experimental Partnerships: Multifacet: Exploiting Prediction and Speculation in Multiprocessor Memory Systems

实验合作伙伴：Multifacet：利用多处理器内存系统中的预测和推测

• 批准号: 9971256
• 负责人: Mark Hill
• 金额: $ 108.19万
• 依托单位: University of Wisconsin-Madison
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 1999
• 资助国家: 美国
• 起止时间: 1999-09-01 至 2003-01-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=9971256&HistoricalAwards=false
• 关键词: Experimental; Partnerships; Multifacet; Exploiting; Prediction

9971256Hill, Mark D.Wood, David A.University of Wisconsin - MadisonCISE Experimental Partnerships: Multifacet: Exploiting Prediction and Speculation in Multiprocessor Memory SystemsThe goal of this research is to improve the performance of multiprocessor servers that form the computational infrastructure for internet web servers, databases, and other demanding applications. It broadly examines exploiting prediction and speculation, heretofore limited to processor cores, in processors, memory, and the interactions between them. A driving focus of the work is multicast snooping, a new cache coherence approach that uses prediction to both increase effective bandwidth over a classical symmetric multiprocessor and reduce latency over a traditional directory-based system. Multicast snooping uses prediction to "guess" which processors have copies of a block and multicast address transactionsto those processors. A novel network organization combines simultaneous multicasts to disjoint multicast sets while still providing ordering guarantees. A simplified directory verifies the prediction and takescorrective action when incorrect. Refining multicast snooping provides research challenges in areas such as hardware and software mask prediction, network implementation, forward progress, high availability, and atheoretical foundation for protocol and network behaviors.The application of prediction and speculation to multiprocessor systems also exposes research opportunities applicable to multiprocessors more generally, including upgrade speculation, false sharing speculation, locksuperposition, and speculation implementation issues. Hardware only, as well as hardware/software solutions strategies are being explored. The research plan encompasses design (refinement of existing ideas and development of new ones), evaluation (pruning the design space and comparing with existing alternatives), validation (ensuring designs work, including a new application of Lamport's logical clocks), demonstration (transferring technology to industry), and infrastructure development (tool adaptation and construction).

9971256 Hill，Mark D.Wood，大卫A.威斯康星州大学麦迪逊分校CISE实验合作伙伴关系：多方面：利用预测和推测在多处理器内存系统这项研究的目标是提高多处理器服务器的性能，形成互联网Web服务器，数据库和其他要求苛刻的应用程序的计算基础设施。 它广泛地研究了利用预测和推测，迄今为止仅限于处理器内核，处理器，内存和它们之间的相互作用。 工作的一个驱动重点是组播窥探，一个新的高速缓存一致性的方法，使用预测，既增加了有效的带宽在一个经典的对称多处理器和减少延迟在传统的基于目录的系统。 多播监听使用预测来“猜测”哪些处理器拥有块的副本，并将多播地址事务发送到这些处理器。 一种新的网络组织结合同时多播不相交的多播集，同时仍然提供排序保证。 一个简化的目录验证预测，并在不正确时采取纠正措施。 改进组播监听技术在硬件和软件掩码预测、网络实现、前向进展、高可用性以及协议和网络行为的理论基础等方面提供了研究挑战。预测和推测在多处理器系统中的应用也揭示了适用于更普遍的多处理器系统的研究机会，包括升级推测、假共享推测、锁叠加、和投机实施问题。 目前正在探索仅硬件以及硬件/软件解决方案战略。 研究计划包括设计（现有想法的改进和新想法的开发），评估（修剪设计空间并与现有替代方案进行比较），验证（确保设计工作，包括Lamport逻辑时钟的新应用），演示（将技术转移到行业）和基础设施开发（工具适应和建设）。