An adaptive microscheduler for a multiprogrammed computer system

The objective of this paper is the description of the development and verification of a new, adaptive method of internal scheduling of resources, with the goal of the optimization of computer system performance. A general system effectiveness measure is defined which parametrically encompasses the prototypical system effectiveness measures to be considered. The adaptive internal scheduler then selects such tasks for resource allocation request fulfillment that a local system effectiveness measure, derived from the general measure, is optimized, leading to semi-optimization of the general measure. The adaptive scheduler functions, as a second-order exponential estimator. A predicator-corrector algorithm functions as the adaptive controller by varying the estimator's parameters and the time of application of the estimator in response to the nature of the sequence of deviations between the predicted and actual values of resource utilization. In order to validate the new scheduler, a workload description in the form of task profile distributions was gathered by a software monitor on the Georgia Tech B5700 running a live job stream. A simulator was developed to allow the comparison of the new scheduler with other nonadaptive schedulers shown to be good by various researchers, under various general system effectiveness measure prototypes. The simulators was validated by running it with the B5700 TSSMCP scheduler against the B5700 workload job profiles. Values resulting from the simulation checked against those of the measured B5700 system quite well. The results of other simulation runs show that the new adaptive scheduler is clearly statistically superior to other schedulers under most measures considered and is inferior to no other scheduler under any measure considered, at least in that environment. Only the new internal scheduler is described here.