TS-BatPro: Improving Energy Efficiency in Data Centers by Leveraging Temporal–Spatial Batching

Abstract

With the rapid scaling of data centers, understanding their power characteristics and optimizing data center energy consumption is a critical task. Typically, data centers are provisioned for peak load even when they are mostly operating at low utilization levels. This results in wasteful energy consumption requiring smart energy saving strategies. At the same time, latency critical application workloads have stringent quality of service (QoS) constraints that need to be satisfied. Optimizing data center energy with QoS constraints is challenging since different workloads can have variabilities in job sizes and distinct system utilization levels. Also, server configuration (e.g., the number of cores per server) can be different across data centers. Therefore, a single configuration for energy management that works well across these various scenarios, is not practical. In this paper, we propose TS-BatPro, a novel framework that judiciously integrates spatial and temporal job batching to save energy for multicore data center servers while meeting the QoS constraints for application workloads. TS-BatPro performs effective global job batching and scheduling by modeling job performance and power characteristics of multicore servers without any hardware modifications. TS-BatPro works on commodity server platforms and comprises two components: 1) a temporal batching engine that batches incoming jobs such that the servers can continue to reside in low-power states and 2) a spatial batching engine that prioritizes scheduling job batches to a small subset of servers. We develop a prototype of TS-BatPro on physical testbed with a cluster of servers and evaluate TS-BatPro on a variety of workloads. Our results show that TS-BatPro is able to achieve significant amount of energy savings under various job response time constraints and traffic patterns.