A Local-Sparing Design Methodology for Fault-Tolerant Multiprocessors

We present a comprehensive design methodology for constructing low-cost multiprocessors that use local spares to tolerate the failure of either processor clusters or individual processors. We first formalize the concepts of global- and local-sparing in terms of graph automorphisms. We then present a method for partitioning a multiprocessor graph by its automorphisms and for incorporating local-sparing to tolerate faults. We emphasize local-sparing designs, since they offer higher reliability-to-cost ratios and can reconfigure faster and in a localized manner. When the spare clusters in each local subsystem are certain sizes, our designs are optimal in the number of spare intersubsystem links. They are all efficient (optimal in some cases) in terms of the number of spare intrasubsystem links. We present switch-based implementations that significantly reduces the spare link complexities of the designs. These implementations are equally efficient for any spare cluster size, so they yield efficient local-sparing designs that can tolerate individual processor faults (cluster size of one). Algorithms for fast, localized, and incremental reconfiguration of our FT designs are also developed. Finally, we demonstrate that our local-sparing designs have higher reliability-to-cost ratios than previous designs.