Online Load Balancing in Parallel Database Queries with Model Predictive Control

This work deals with a commonly encountered instantiation of the more generic problem of runtime load balancing. More specifically, we study load balancing in the context of query plans that include partitioned database operators running on remote nodes with time-varying connection speeds and loads. In such query plans, data is processed by similar operator instances placed on different machines in parallel, e.g., using a cloud infrastructure. The goal is to adaptively distribute the data to respond to changes in the environment, in such a way that all machines finish at the same time. Dynamic modifications in the data distribution policy are easily enforced, but they result in non-negligible adaptation overheads due to associated data movements. The emphasis of our approach is on taking into account the inherent cost of adaptations with a view to avoiding over-reacting, which may lead to serious performance degradation. We follow a rigorously founded control theoretical approach to this problem and we present a Model Predictive Control (MPC) controller, which tolerates monitoring inaccuracies and exhibits better performance under periodic and random machine perturbations than other state-of-the-art approaches, as shown by our experiments.