众核片上资源动态划分与管理研究

Research on the dynamic division and management of resources on multiprocessor system-on-chip

为了提高芯片的可扩展性多采用基于No C的分簇管理方案,现有的基于应用的动态实时分簇管理方案已有较深入的研究,然而关于固定分簇方案的研究较为缺乏,包括在该方案下的核级容错策略。在此背景下设计了一种基于固定分簇方案的核级容错策略,提出了片上区域重划分算法,并完成了芯片的MATLAB建模及实现。进行了故障注入实验,将区域重划分算法与随机分簇算法就分簇后的片上平均曼哈顿距离进行比较,得到了比较好的结果,加入侧边冗余核之后,将区域重划分算法与工程常用的行列替换策略进行比较,结果也表明该算法优于行列替换策略。

To increase the scalability of cores, many methods are used, including Network on Chip（ No C） and cluster-based dis-tributed management scheme. The application-based re-clustering algorithm has been delved deeply, while fixed-sized cluster is less developed, including the core-level fault tolerant scheme under such method. Under such environment, a re-clustering scheme based on fixed-sized cluster was proposed in order to achieve fault tolerance, including dynamic re-clustering algorithm. This mod-eling of chip was finished on MATLAB, and the proposed dynamic re-clustering algorithm was compared with several other algorithms.Core error injections were did and the Average Manhattan Distance（ AMD） of the dynamic re-clustering algorithm was compared with random re-clustering algorithm. The results show that the dynamic re-clustering algorithm is far better than random re-clustering algorithm. Then backup cores to the side of the chip were added, and the dynamic re-clustering algorithm was compared with the same-row-replacing algorithm what was commonly used in industry. The dynamic re-clustering algorithm still shows advantages.