Power-aware temporal isolation with variable-bandwidth servers

Abstract

Variable-bandwidth servers (VBS) control process execution speed by allocating variable CPU bandwidth to processes. VBS enables temporal isolation of EDF-scheduled processes in the sense that the variance in CPU throughput and latency of each process is bounded independently of any other concurrently running processes. In this paper we aim at reducing CPU power consumption with VBS by CPU voltage and frequency scaling while maintaining temporal isolation. Scaling to lower frequencies is possible whenever there is CPU slack in the system. We first show that, in the presence of CPU slack, frequency scaling of EDF-scheduled, possibly non-periodic tasks (as they arise with VBS) is safe up to full CPU utilization and propose a frequency-scaling VBS algorithm that exploits CPU slack to minimize operating frequencies with maximal CPU utilization while maintaining temporal isolation. Additional power may be saved by redistributing computation time of individual processes while still maintaining temporal isolation if the system has knowledge of future events. We introduce an offline algorithm as an optimal baseline and an online algorithm that approximates the baseline. While the offline algorithm works for various, possibly complex power consumption models, the online algorithm may reduce power consumption only for a simplified power consumption model by reducing the CPU utilization jitter in the system.