A Self Monitoring Approach to Adaptive Processor Allocation

Given a multiprocessor system and a workload composed of several parallel programs , the processor allocation problem is to determine the number of processors that should be allocated to each parallel program. One simple, but common strategy is to give the entire system to an arriving job, thus executing the workload in a unipro-gramming, rst-come-rst-serve fashion. However, it is often possible to improve performance by allocating only a portion of the available processors to each incoming job. Several strategies exist for determining this allocation. An adaptive strategy is one in which the number of processors allocated to a waiting program may vary depending on values of the current wait queue length, the number of other jobs in execution, the number of free processors, or other components of the current system state description. Determining the \optimal" adaptive allocation strategy is a diicult problem. The typical approach to adaptive allocation requires a priori knowledge of the workload. In this paper an approach to adaptive allocation is presented which monitors the performance of allocation decisions that are made and makes future decisions based on the perfromance of prior decisions. No a priori knowledge of the workload is assumed. The method used for self monitoring is a stochastic learning automaton. The method is validated through simulation. The results of this study show that a stochastic learning automaton can be used to nd an eeective adaptive processor allocation strategy.