Optimal Multiprogramming Control for Parallel Computations

Traditionally, jobs on parallel computers are run one at a time, and control of parallelism so far was mainly guided by the desire to determine the optimal number of processors for the algorithm under consideration. Here, we want to depart from this course and consider the goal of optimizing the performance of the overall parallel system, assuming more than one job is available for execution. Thus the central issue of this paper is the question how the available processors of a parallel machine should be distributed among a number of jobs. In order to obtain guidelines for such multiprogramming control, we use the speedup-behaviour and the accumulated processor time of a job as its characterization.