Scheduling of near-future workload in distributed computing systems

A major issue in distributed computing systems is the choice of an adequate load balancing policy. 'Traditional' load balancing policies assign independent tasks to servers, while scheduling policies assume complete knowledge of task dependencies and treat the assignment problem as a large search problem. However, both approaches are based on assumptions that do not hold in distributed computing systems: the job context of tasks exists and thus dependencies between tasks, but they cannot be predicted a long time in advance. In this paper, a dynamic decentralized load balancing policy is introduced that utilizes neat-future workload predictions based on knowledge about the job context. A performance evaluation by simulation and a comparison to widely used load balancing policies is presented.<>