Design of loading policies for a high-performance computer system

We use a systematic simulation study of computer-system loading policies to design a loading policy for a high-performance system. Our results illustrate the performance advantages of ‘round-robin’ disciplines, the interference caused by finite memory, and how this interference can be mitigated by using carefully selected loading policies that exploit information about job resource requirements. Our choice of performance measures allows us to assess response-time predictability and fairness as well as overall efficiency. Using the insights obtained from our study, we introduce a new type of policy, the NEST policy, which is structured to employ the principles of good scheduling, while avoiding shortcomings of commonly used approaches.