Distributed Computer System, Dynamic Load Balancing Studies

Dynamic load balancing is a hot reaearch point in Distributed Systems. By using on-line tracing technique, this paper can predicate the behavior of a job, such as its resource requirements and its approximate execution time. Therefore, we can recognize those short-lived jobs which are not worth transferring. And according a job's resource requirements, we can also find a node to execute it which satisfies its need best. In additional, we point out that in dynamic load balancing systems, using resource utilizations as load index is better than using CPU queue lengths. In this paper, we give an algorithm for predicating the execution time of various jobs under various load environments. Based on this predication, the overhead and benefit of transferring a job can be calculated. A new selection policy based on the overhead and benefit of transferring a job is designed and implemented. We consider that the best node for a job is the one on which the job will run the shortest time. Experiment measurement results show that it is able to reduce mean response time and improve resource utilization of systems.