In: IEEE Computer Soc. Press, Proc. of 6th International Workshop on Petri Nets and Performance Models - PNPM'95, Durham, N. Carolina, USA, pages 73-82. 1995.
Abstract: A multiprocessor system can be subdivided into partitions of processors, each of which can be dedicated to the execution of a parallel program. The partitioning of the system can be done statically at system configuration time, adaptively prior to the execution time, or dynamically during execution time. Since, in a dynamically partitioned multiprocessor system, partitioning can occur anytime during the execution of a program, designing an analytical model for such a system is a difficult task. In this paper a Petri net model of a dynamically partitioned multiprocessor system is presented. The workload consists of parallel programs which are characterized by their execution signatures. Repartitioning overhead is an important parameter and is modeled explicitly. The model is used to perform a series of sensitivity analysis experiments which give insight into the behavior of such systems. Several dynamic processor allocation policies have been implemented. Equal Slope and Shortest Job First Preemptive Resume achieve the best performance of the policies considered.