A proposed approach to recovery from transient faults in wavefront processor arrays

The need for a transient fault recovery procedure in a wavefront processor array is examined. The idea of data rollback to recover from transient faults is considered. Data rollback is achieved by creating checkpoints at different instants of time; when an error occurs, backtracking is done to a consistent state and computation resumes from that state. An algorithm for data rollback in wavefront processor arrays is developed. After the fault is detected, the error information is propagated throughout the array. A distributed recovery algorithm is employed, so that there is no single point of failure in the fault recovery mechanism. An important criterion in the data-rollback algorithm is that the processing elements are designed to be totally self checking; this ensures that there is no rollback propagation