Reconfiguration and experiments on a faulty hypercube

It is often desirable to reconfigure a faulty hypercube so as to retain as many fault-free nodes as possible. This inspires us to identify maximal incomplete subcubes in a faulty hypercube, as the subcube so reconfigured is often much larger than a complete subcube obtainable and is likely to retain performance batter. An efficient algorithm is first presented to find incomplete subcubes in a faulty hypercube. To study the actual performance difference on an incomplete system and a complete one, tree applications are then implemented on both systems. Among the three applications, Gaussian elimination takes exactly the same mapping scheme on an incomplete system as on a complete one, while FFT requires some efforts to make it adaptive to the incomplete topology. Our experimental results collected on the Intel iPSC/860 show that an incomplete system may outperform its complete counterpart significantly. The measured results of the three applications indicate that reconfiguring a faulty hypercube into an incomplete subcube is beneficial in practice