High Reliability Self-recovery Methods And Optimization Based On Many Core Architecture

Network-on-Chip (NoC) is one of the representatives in Many-coretechnology, which is the hotspot in nowadays research. As semiconductorshrinks down, Soft and hard errors become increasingly prominent.Fault-tolerant technology is an important way to enhance reliability ofmany-core systems. Currently, repair rate is not high and time delaybecomes very long after been repaired.NoC has two kinds of communication style, namely exclusivecommunication style and addressing exchange style. The studies in thispaper provide two methods aimed at improving reliability of these twostyles.For exclusive communication, this paper proposes a low-latencyself-recovery method through part-reconfiguration. Optimal repair designis calculated by software in advance for every error pattern, the results arestored in an additional memory; when an error happens in the system,control module only needs to find an appropriate solution in memory torepair the system. This method can repair faults happen in links, routersand cores and can be used in many kinds of NoC structures. This methodreduces an average of19%of the time delay after been repaired, the repairprocess can be completed in four clock cycles, improving system repairrate by about32%in average.For addressing exchange, this paper proposes a self-recovery structure.In this structure, four routers form a cluster. For each cluster, one sparerouter is added, for each cluster line, one spare cluster is added, providingdual-protection for the system. The simulation results show, based on 14*14Mesh structure, compared to other methods, this method canimprove system’s reliability by40%in average and provide121%more ofthe mean time to failure (MTTF).