Pelican: Power Scheduling for QoS in Large-scale Data Centers with Heterogeneous Workloads
Bing Luo, Wei Chen, Xingxing Liu, Xiaozhong Li, Lifei Zhang, Weisong Shi · 2019
Given the high cost of building and operating large-scale data centers in terms of the power budget, it is crucial to maximize available power resources for overall quality of service(QoS). In this paper, we present a power scheduling system, Pelican, for large-scale data centers to increase overall QoS by over-provisioning the number of servers for heterogeneous workloads. We adopt a two-level design that separate power scheduling mechanism and power controlling policy without affecting the already over-complicated job scheduler. Our power controller leverage both priority and performance as key indicators to control the power of a server, which improves QoS while significantly reducing the risk of power violations. We implemented and deployed our Pelican in a real overprovisioned production data center at Baidu Inc. Our evaluations show that we are able to utilize 12.7% more computing resources without any violation and achieve 11.5% offline service throughput gain while not affecting the latency of online services so as to improve overall QoS.