Effects of job loading policies for multiprogramming systems in processing a job stream

Abstract

The scheduling of jobs for multiprogramming systems includes the selection of jobs to be loaded into memory (job loading policy or memory schedule) and the scheduling for CPU processing (CPU schedule). There has been a successful empirical claim for the optimal CPU schedule; its optimality has been proved in a Markovian model of job-stream processing that uses the first-come-first-loaded (FCFL) job loading policy. We extend this model to gain insight into the effects of job loading policies. The model studied consists of an input stream of jobs of two classes and a multiple-resource system (the model of a multiprogramming system) with a stack for waiting jobs. The system consists of a finite amount of memory and a cyclic queue of a single (CPU) server station and a multiple (I/O) server station. The values of parameters describing each class of jobs are distinct except the mean I/O service time and the amount of memory required. The estimate of the maximum processing capacity (throughput bound) of the system is obtained and is shown to be achieved by the combination of the empirically claimed optimal CPU schedule and a job loading policy whereby the set of jobs in memory is kept to be (nearly) balanced with respect to the job stream. Furthermore, we show that the job loading policies independent of the system status have no improvement over the FCFL policy. Our investigation, supported by numerical calculations, suggests that much more care may be needed in implementing the job loading policy that aims at the optimal processing capacity than in implementing the optimal CPU schedule. This agrees with what has been conjectured on the basis of empirical studies.