Cycle stealing and truncation for reducing delay in heavy-tailed queues.

Abstract

Significant research effort has recently targeted reducing delay in “heavy-tailed” (highly variable) systems via truncating service times. However, these schemes induce idle times for some servers. This thesis proposes “cycle stealing” as a remedy to reduce idle times and further reduce delays. Servers are arranged into two stations in series. Arrivers, who find either station busy, are served at the first station for a maximum time T. Among this demand stream, those with service requirements exceeding T, continue service at the second station. To this truncation scheme, we add cycle stealing, where an arriver, who finds the second station idle, joins it. Simulation-optimization results demonstrate superior performance of this system with respect to the parallel multi server system and a similar system having no cycle stealing. We also develop an approach for determining an appropriate value of T analytically, with no simulation. This approach, based on analyzing a simpler proxy system, also yields superior performance.