Efficient bi-level reconfiguration algorithms for fault tolerant arrays

Abstract

Considers the problem of reconfiguring processor arrays subject to computational loads that alternate between two modes. A strict mode is characterized by a heavy computational load and severe constraints on response time while a relaxed mode is characterized by a relatively light computational load and relaxed constraints on response time. In the strict mode, reconfiguration is performed by a distributed local algorithm in order to achieve fast recovery from faults. In the relaxed mode, a global reconfiguration algorithm is used to restore the system to a state that maximizes the probability that future faults occurring in subsequent strict modes will be repairable. Several new results are given for this problem. Efficient reconfiguration algorithms are described for a number of general classes of architectures. These general algorithms obviate the need for architecture-specific algorithms for architectures in these classes. It is unlikely that similar algorithms can be obtained for related classes of architectures since the reconfiguration problem for these classes is NP-complete. Finally, a general approximation algorithm is described that can be used for any architecture. Experimental results are given, suggesting that this algorithm is very effective.<>