多核处理器容错关键技术研究

The advancement of IC technology led to billions of transistors on die. Chip Multiprocessors provide an efficient and scalable solution to make use of those transistors. However, reliability becomes one of the key issues in the CMP architecture design with the increasing of integration. This project intends to tolerate faults from the key modules of CMP, such as Caches and Network on-Chip(NoC), based on analyses of the reliability and failure model of CMP. Specifically,it includes Cache fault tolerance in Cache line level, Cache fault tolerance in module level with network address mapping, NoC fault tolerance by last-level Cache (LLC) capacity borrowing, as well as fault tolerant reconfigurable NoC. Coordinatively using the on-chip memories, maintaining the integrity of on-chip cacheable address space, and tolerating faults of on-chip Caches and NoC are critical to this project. The network mapping technology of cacheable address space, LLC-based reconfigurable NoC and smoothly degradation are our main contributions. Totally, researches of this project have great theoretical and practical significance for improving the reliability and availability of CMP systems.

随着集成电路工艺技术不断进步，单芯片上集成数十亿晶体管成为可能。片上多核处理器为有效利用这些晶体管资源提供了一种高效、可扩展的方案。但随着集成度的不断提高，可靠性问题成为多核处理器研发和应用重要问题之一。本项目拟通过对处理器可靠性和故障分析，研究多核处理器各级Cache、片上网络（NoC）等关键模块故障时的容错技术。包括：基于组划分的Cache行级容错技术;基于网络地址映射的Cache模块级容错技术;基于最后级Cache（LLC）容量借用的NoC容错技术，以及支持容错的可重构NoC架构等关键技术。协同利用片上存储资源、维护片上可缓存地址空间完整、实现片上Cache、NoC容错是本项目研究的关键。缓存空间的网络映射技术、基于LLC的NoC重构与系统适度降级容错技术是项目研究的创新所在。因此，项目研究对于提高多核处理器的可靠性和多核应用系统的可用性具有重要的理论和实践意义。

随着集成电路工艺技术不断进步，单芯片上集成数十亿晶体管成为可能。片上多核处理器为有效利用这些晶体管资源提供了一种高效、可扩展的方案。但随着集成度的不断提高，可靠性问题成为多核处理器研发和应用的重要问题之一。本项目拟通过对处理器可靠性和故障进行研究分析，研究多核处理器各级Cache、片上网络（NoC）等关键模块故障时的容错技术。为解决多核处理器的可靠性问题，本项目从基于组划分的Cache行（Cache line）级容错技术、基于网络地址映射的Cache模块级故障容错技术、基于LLC容量借用的NoC故障容错技术等方向着手，并提出Cache组管理部件（Cluster Cache Monitor, CCM）、依据贝叶斯决策理论的最优数据源选择（Optimal Data-Provider Selection, ODPS）、基于数据访问类型的Cache替换策略（Data access Type Aware Replacement Policy, DTARP）、可扩展地址映射（Scalable Address Mapping, SAM）、片上网络消息缓冲区（Pull-Off Buffer, POB）等技术具体实现了上述目标。