An adaptive approach to accelerated evaluation of highly available services

Abstract

We motivate and describe improved fast simulation techniques for the accelerated performance evaluation of highly available services. In systems that provide such services, service unavailability events are rare due to a low component failure rate or high resource capacity. Using traditional Monte Carlo simulation to evaluate such services requires a large amount of runtime. Importance sampling (IS) has been applied to certain instances of such systems, focusing on single-class and/or homogeneous resource demands. In this article, we formulate highly available services as multiresource losstype systems, and we present two IS methods for fast simulation, extending to multiple classes and nonhomogeneous resource demands. First, for the cases in which component failure rates are small, we prove that static IS using the Standard Clock (S-ISSC) method exhibits the bounded relative error (BRE) property. Second, for estimating failure probabilities due to large capacity or fast service in systems that have nonrare component failure rates, we propose adaptive ISSC (A-ISSC), which estimates the relative probability of reaching each possible state of system failure in every step of the simulation. Using A-ISSC, IS methods which are proven to be efficient can be extended to multidimensional cases, while still retaining a very favorable performance, as supported by our validation experiments.