Design of a highly reliable cube-connected cycles architecture

The cube-connected cycles (CCC) architecture is an attractive supercomputer system, because it not only preserves all the desired features of hypercubes but also is suitable for VLSI implementation. However, the CCC tends to suffer from considerable performance degradation when a fault arises. This paper introduces a fault-tolerant CCC constructed from adding a spare PE to every cycle and making "cross connections" among cycles, called the XCCG. The reconfiguration procedure for the XCCG is simple and can be done in a distributed manner. The XCCG is shown to exhibit significantly enhanced reliability. The XGCG layout is discussed, and its area overhead is found to be moderate if the PE size is far larger than the link/switch size. As a result, this design approach is particularly useful for situations where the PE is relatively complex.