Speed scaling in fork-join queues: a comparative study

Frequency scaling plays an important power-saving role in computer systems. In fork-join systems, dynamic adaptation of the server speeds can significantly reduce system power consumption while maintaining high throughput. In previous work, we studied a rate adaptation policy that dynamically chooses server speeds based on the difference in join-queue lengths, with each server knowing only its own join-queue length and that of one other server. In this work, we increase the information available to each server, and choose speeds based on the knowledge of the join-queue lengths of two other servers. We show that, under a specific canonical configuration of the service rates, the new system has exactly the same throughput and subtask dispersion as before, but with reduced power consumption. We use time-reversal analysis to derive the exact stationary performance of this new model under saturation conditions, and use simulation to study more general cases.