Multiprogrammed Memory Management for Random-Sized Programs

We consider a probabilistic model of a computer system with multipro-gramming and paging. The applied work-load is derived from measurements in scientific computer applications and is characterized by a great variance of compute time. Throughput of a cyclic model is computed approximately presuming program sizes with negative exponential distribution. After a review of previous results for a memory allocation policy with prescribed number, n, of working sets at least to be loaded, an adaptive memory allocation policy is introduced which dynamically changes the number, n. Thereby, it is possible to reach the goal of having always enough memory available to load the parachor of each program. Simulation results establish our approximations as being very good. CPU scheduling is chosen to be through-put optimal. Our results are useful to demonstrate the benefits of allocation policies with adaptive controlled degree of multiprogramming. Previous contributions to this problem are to that date only by means of simulation [5].