Search-Based Parameter Tuning on Application-Level Load Balancing for Distributed Embedded Systems

Load balancing techniques for distributed embedded systems tend to be very parameter-rich, and finding adequate parameters for a given scenario is not trivial. Furthermore, the parameter values that are adequate for one scenario are rarely applicable to another that has, for instance, a different application profile or network topology. In this paper, we present a search-based parameter tuning approach that aims to automate the process of configuring a parameter-rich load balancing technique. It considers the service availability and energy dissipation figures obtained by each configuration of the load balancing technique, and uses those values to explore the parameter space towards optimised solutions. To accelerate the search, we also present a number of improvements to the simulator used to evaluate each configuration. The proposed parameter tuning approach is then evaluated by analysing the best configurations it can find for several scenarios, and we use Principle Component Analysis to identify which of the parameters have the most critical effect on the quality of the solutions.