An adaptive policy driven scheduler

The theory of policy driven schedulers (Ref. [1]) is extended to cover cases in which the scheduling parameters are allowed to adapt dynamically as the system's job load varies. The system under consideration offers batch, time sharing and limited real time services. Data from simulated and live loads are presented to evaluate both the static and the adaptive schedulers. A policy driven scheduler makes its decisions with respect to a set of policy functions, f i (t). Each of the policy functions corresponds to a different type of user and specifies the amount of computing resources that the system will try to give a user in that group within a given total amount of elapsed time. It is found that the policy functions must be set conservatively in order to avoid response problems during periods of heavy load, but that during more lightly loaded periods the conservative settings result in widely disparate rates of service to similar jobs. One solution is to vary the policy functions as the job load changes. A dynamic algorithm is presented that maintains responsiveness during heavy loads and provides fairly uniform service rates at other times.