NoC众核系统的核级冗余容错管理和替换方法研究

NoC-based many-core systems have much better performance than the multi-core systems, but face the much serious core-fault problems which are introduced by permanent failures. The core-level redundancy technology, widely used in the industry, has been the hot topic in the NoC-based many-core systems fault-tolerance, due to the complementary with NoC. The project focuses on the core management and replacement method in the NoC-based many-core fault-tolerant systems with core-level redundancy. Firstly, breaking the boundaries between the managing cores and the computing cores, increasing the replacement alternative space for the managing cores redundancy, the automatically reorganization core management structure and the mechanism aiming at any fault state will be introduced, and focusing on how to increase the robustness and the reliability of the core manage systems. Secondly, Combining the virtual topology technology with the NoC topology recovery technology, the NoC architecture and the mapping method for the virtual topology, which can support the core topology recovery, will be introduced to resolve the performance gap between the core-level replacement afore and after. Finally, according to the online fault recovery needs, low cost implementation mechanism of the core-level redundancy replacement will be proposed. The purpose of the project is to get a more completely fault-tolerant redundancy methods in the NoC-based many-core systems, and provide the theoretical method and experimental data for application of the core-level redundancy on NoC-based many core fault-tolerant systems, providing with important theoretical and practical value.

NoC众核系统具备比多核系统更好的性能，但对于永久性核心故障问题更加敏感。广泛应用于工业界的核级冗余技术，因与NoC具备互补性，成为NoC众核系统容错领域的研究热点。本课题针对基于核级冗余技术的NoC众核容错系统，研究其核心管理和替换方法中的难点和重点问题。具体来讲：通过打破管理核和计算核的角色界限，增加管理核冗余替换备选空间，提出能够在任意故障下自动重组的核心管理结构及其重构机制，研究提高核心管理系统鲁棒性、可靠性的方法；通过将虚拟拓扑技术和NoC拓扑还原技术相结合，提出能够支持核心物理拓扑恢复的NoC架构及其配套的虚拟拓扑映射方法，解决核级冗余替换中存在的性能差别问题；最后根据在线故障恢复需求，研究低开销核级冗余替换实现机制，以期得到一套较为完整的NoC众核系统核级冗余容错方法。为核级冗余技术在NoC众核容错系统应用中提供理论方法和实验数据支持，具有重要的理论意义和实际应用价值。

近年来，随着集成电路工艺和处理器设计技术的发展，学术界和工业界一致认为基于片上网络（Network-on-Chip，NoC）互连结构的众核系统具有比多核系统更好的性能，但同时也存在着更多的设计挑战。其中，如何克服由于芯片制造过程的缺陷以及老化等问题引起的硬错误而导致的单个或多个核心永久性故障问题，一直是研究人员关注的热点和难点问题。本项目以研究NoC众核系统芯片的容错方法为背景，以突破核级冗余技术面临的挑战为目的，提出了具备故障不敏感的高鲁棒性核心管理策略、低性能差异乃至无差异的冗余核替换方法以及相配套的低开销在线容错实现机制，与此同时，由于NoC众核系统与上层应用密切相关，为此，我们在原定的研究计划的基础上，对上层应用提前进行了相关研究。以当前学术界和工业界非常热门的汽车自动驾驶相关应用为背景，在NoC众核系统平台的基础上，对汽车自动驾驶的相关应用进行了细致的研究，并提出了道路去反光和去雾算法。.