This paper uses trace-driven simulations to evaluate two novel Web server scheduling policies called KSRPT and T-SRPT. K-SRPT is a multi-threaded version of SRPT (Shortest Remaining Processing Time) that can have up to K jobs in service at a time. TSRPT is a hybrid policy that dynamically switches from PS (Processor Sharing) to SRPT when the number of jobs in the system exceeds T, and back to PS when the load diminishes. Slowdown profile plots are used to analyze the performance of these policies relative to PS, SRPT, and FSP (Fair Sojourn Protocol). Simulation results show that these new parameterized policies offer smooth performance tradeoffs between SRPT and PS. Stability issues for T-SRPT are also discussed.
Mingwei Gong, Carey L. Williamson