The Learning Behaviour of a Scheduler using a Stochastic Learning Automaton

This paper discusses a load balancing heuristic in a general-purpose distributed computer system. We implemented a task scheduler based on the concept of a Stochastic Learning Automaton on a network of Unix workstations. The used heuristic and our implementation are shortly described. Creating an executable artiicial workload, a number of experiments examined diierent learning schemes. Using a linear reward{ penalty scheme resulted in the best performance of the scheduler. Another series of experiments looked at diierent ways to evaluate the goodness of a scheduling decision, another aspect of the learning behaviour. Instead of using a simple binary (qualitative) measure, a quantitative evaluation allowed for a more stable and therefore better learning behaviour.