Optimizing and evaluating algorithms for replicated data concurrency control

Abstract

Many algorithms for replicated data concurrency control are based on voting methods. Techniques are developed for optimizing the assignment of votes in an environment where intersite communications costs are nonuniform and individual site reliabilities vary. These techniques apply to all algorithms that are based on voting. Availability is considered as a realistic measure of reliability, and so is incorporated in an optimization model. The optimization model is based on minimizing communications costs subject to a given availability constraint. A semi-exhaustive algorithm is described for solving this model. The algorithm utilizes a signature-based method for identifying equivalent vote combinations and an efficient technique for computing availability. It is compared to an equal vote assignment to estimate the extent of possible savings in communications costs.<>