Scaling theory for fault stealing algorithms in large systolic arrays

The performance of fault-stealing algorithms for very large, multipipeline systolic arrays is considered. Extensions of an existing algorithm are proposed, and with these extensions the algorithm is shown to work for large array sizes. Using the modified algorithms as a testbed, a scaling theory that predicts, on the basis of performance for a single small array, the performance of the algorithm for arbitrary array size, defect rate, and number of spares is introduced. The theory differs from current approaches in that it has both analytical and empirical components, and in that it accurately predicts system performance, rather than providing bounds on it.<
>