Understanding Microquanta Process Scheduling for Cloud Applications

Abstract

Process schedulers are responsible for arbitrating CPU resources among services. Unfortunately, traditional sched-ulers, working at millisecond scale and characterized by strict priority schemes, are no longer suitable to meet increasingly stringent and diverse requirements imposed by many workloads. Recognizing this aspect, the research community has recently proposed new schedulers operating at microsecond granularity. This work studies microsecond-scale schedulers and policies for data center applications from the configuration versus perfor-mance standpoint. We demonstrate that for the best performance, workload-dependent parameter tuning is fundamental. Specifically, even a slight misconfiguration can also lead to 110% higher tail latency with respect to its best-case scenario. Our results call for a new set of process scheduling schemes that are workload-aware.