Automated learning of load-balancing strategies for a distributed computer system
Pankaj Mehra · 1993
Workstations interconnected by a local-area network are the most common examples of distributed systems. The performance of such systems can be improved via load balancing, which migrates tasks from the heavily loaded sites to the lightly loaded ones. Load-balancing strategies have two components: load indices and migration policies. This thesis presents SMALL (Systematic Method for Automated Learning of Load-balancing strategies), a system that learns new load indices and tunes the parameters of given migration policies. The key component of SMALL is DWG, a dynamic workload generator that allows off-line measurement of task-completion times under a wide variety of precisely controlled loading conditions. The data collected using DWG are used for training comparator neural networks, a novel architecture for learning to compare functions of time series. After training, the outputs of these networks can be used as load indices. Finally, the load-index traces generated by the comparator networks are used for tuning the parameters of given load-balancing policies. In this final phase, SMALL interfaces with the TEACHER system of Wah, et al. in order to search the space of possible parameters using a combination of point-based and population-based approaches. Together, the components of SMALL constitute an automated strategy-learning system for performance-driven improvement of existing load-balancing software.