Implementing an Approximate Probabilistic Algorithm for Error Recovery in Concurrent Processing Systems

We have developed a probabilistic algorithm for improved error recovery in a system of concurrent processes. Simulations for various lengths of checkpoint intervals have shown that in most cases the probabilistic method is more cost effective than the iterative rollback method. However, implementation of the probabilistic algorithm requires knowledge of the distribution function of the latency times between error occurrence and error detection. In this paper, we present a method for obtaining an approximate empirical distribution function for the latency times using the iterative rollback method. The cost effectiveness of the probabilistic method, when based on the approximate distribution function, is investigated for various parameters (number of data points collected, length of error interval). We show that using the probabilistic algorithm in conjunction with the approximate distribution function still leads to significant cost reduction over the iterative method, while not requiring knowledge of the theoretical distribution function, making this implementation universally applicable.