Efficient optimization of software performance models via parameter-space pruning

When performance characteristics are taken into account in a software design, models can be used to identify optimal configurations of the system's parameters. Unfortunately, for realistic scenarios, the cost of the optimization is typically high, leading to computational difficulties in the exploration of large parameter spaces. This paper proposes an approach to provably exact parameter-space pruning for a class of models of large-scale software systems analyzed with fluid techniques, efficient and scalable deterministic approximations of massively parallel stochastic models. We present a result of monotonicity of fluid solutions with respect to the model parameters, and employ it in the context of optimization programs with evolutionary algorithms by discarding candidate configurations a priori, i.e., without ever solving them, whenever they are proven to give lower fitness than other configurations. An extensive numerical validation shows that this approach yields an average twofold runtime speed-up compared to a baseline optimization algorithm that does not exploit monotonicity. Furthermore, we find that the optimal configuration is within a few percent from the true one obtained by stochastic simulation, whose solution is however orders of magnitude more expensive.