Stochastic modeling for performance evaluation of database replication protocols

Abstract

Performance is often the most important non-functional property for database systems and associated replication solutions. This is true at least in industrial contexts. Evaluating performance using real systems, however, is computationally demanding and costly. In many cases, choosing between several competing replication protocols poses a difficulty in ranking these protocols meaningfully: the ranking is determined not so much by the quality of the competing protocols but, instead, by the quality of the available implementations. Addressing this difficulty requires a level of abstraction in which the impact on the comparison of the implementations is reduced, or entirely eliminated. We propose a stochastic model for performance evaluation of database replication protocols, paying particular attention to: (i) empirical validation of a number of assumptions used in the stochastic model, and (ii) empirical validation of model accuracy for a chosen replication protocol. For the empirical validations we used the TPC-C benchmark. Our implementation of the model is based on Stochastic Activity Networks (SAN), extended by bespoke code. The model may reduce the cost of performance evaluation in comparison with empirical measurements, while keeping the accuracy of the assessment to an acceptable level.