Improved Prediction-based Dynamic Load Balancing Systems for HLA-Based Distributed Simulations

Abstract

Due to the dependency of High-Level Architecture (HLA)-Based simulations on the resources of distributed environments, simulations can face load imbalances and can suffer from low performance in terms of execution time. HLA is a framework that simplifies the implementation of distributed simulations; it also has been built with dedicated resources in mind. As technology is nowadays shifting towards shared environments, the following two weaknesses have become apparent in HLA: managing federates and reacting towards load imbalances on shared resources. Moreover, a number of dynamic load management systems have been designed in order to provide a solution to enable a balanced simulation environment on shared resources. These solutions use specific techniques depending on simulation characteristics or load aspects to perform the balancing task. Load prediction is one of such techniques that improves load redistribution heuristics by preventing load imbalances. In this thesis, a number of enhancements for a prediction technique are presented, and their efficiency are compared. The proposed enhancements solve the observed problems with Holt’s implementations on dynamic load balancing systems for HLA-Based distributed simulations and provide better forecasting. As a result, these enhancements provide better predictions for the load oscillations of the shared resources. Furthermore, a number of federate migration decision-making approaches are introduced to add more intelligence into the process of migrating federates. The approaches aim to solve a dependency problem in the prediction-based load balancing system on the prediction model, thus making similar systems adapt to any future system improvements.