Modeling the Cost of Redistribution in Scheduling

Distributed heterogeneous resources coordinated by fast networks provide a new platform for resource-intensive parallel applications. However, careful scheduling of such applications is required to achieve their performance potential. Since resource load varies dynamically, the scheduling mechanism may consider redistribution of application tasks to improve performance. In this paper, we focus on the development of dynamically parametrizable models to determine the cost (in terms of execution delay) of redistribution. When such models are combined with predictive models for application execution, a scheduler can determine whether redistribution is pro table at a given point in time.