Efficient Load Balancing in Partitioned Queries Under Random Perturbations

摘要

This work investigates a particular instance of the problem of designing efficient adaptive systems, under the condition that each adaptation decision incurs some nonnegligible cost when enacted. More specifically, we deal with the problem of dynamic, intraquery load balancing in parallel database queries across heterogeneous nodes in a way that takes into account the inherent cost of adaptations and thus avoids both overreacting and deciding when to adapt in a completely heuristic manner. The latter may lead to serious performance degradation in several cases, such as periodic and random imbalances. We follow a control theoretical approach to this problem; more specifically, we propose a multiple-input multiple-output feedback linear quadratic regulation (LQR) controller, which captures the tradeoff between reaching a balanced state and the cost inherent in such adaptations. Our approach, apart from benefitting from and being characterized by a solid theoretical foundation, exhibits better performance than state-of-the-art heuristics in realistic situations, as verified by thorough evaluation.