Performance Evaluation of Dynamic Load Balancing System for Clusters
Dan Tang · Chinese Journal of Computers · 2004
A cluster system consists of a collection of computing nodes that share resources. Cluster systems utilize load balancing technology to share processing power and other resources, and thus to improve system performance, by transparently transferring load between nodes. In this paper, the Stochastic Petri Nets are employed to model the dynamic load balancing systems for clusters. The model we represent can avoid the disadvantages such as the lack of flexibility and the obvious difference from practical systems of previous work. We detailed the local process part of the model in order to analyze the performance of five dynamic load balancing schemes. We also discussed the impact of some workload characters of clusters on the performance of dynamic load balancing systems. It is found that:(1) Dynamic load balancing has the better system performance over that obtained with static load balancing;(2) Compared to the traditional load balancing algorithms that consider only the running queue, the load balancing algorithms which consider the I/O request queue can get better performance;(3) Even with extreme workload characters, the dynamic load balancing system can also get ideal performance, therefore implementing even a very simply dynamic load balancing system is also necessary.