Q-learning scheduler and load balancer for heterogeneous systems

Abstract

Distributed computing is a viable and cost-effective alternative to the traditional model of computing. Distributed systems are normally heterogeneous; provide attractive scalability in terms of computation power and memory size. Generally, in such systems no processor should remain idle while others are overloaded. Large degrees of heterogeneity add additional complexity to the scheduling problem. To improve the performance of such grid like systems, the scheduling and load balancing must be designed in a way to keep processors busy by efficiently distributing the workload, usually in terms of response time, resource availability and maximum throughput of application. Dynamic load balancing is NP complete. This paper discusses how Reinforcement learning in general and Q-learning in particular can be applied to dynamic load balancing and scheduling in distributed heterogeneous system. We consider a grid like environment consisting of multi-nodes. Experimental results suggest that Q-learning improves the quality of load balancing in large scale heterogeneous systems. © 2007 Asian Network for Scientific Information.