Exploring the Potential of CMP Core Count Management on Data Center Energy Savings

A data center’s power and energy consumption is a crucial design issue, and is often a fundamental determiner of its performance potential. Conventional power/energy management approaches data centers have focused on uniprocessor rather than multiprocessor servers and on techniques such as Dynamic Voltage and Frequency Scaling (DVFS). Given the increasing use of chip multiprocessors (CMPs) and the decreasing leverage of DVFS, it is time to reexamine energy and power management possibilities with an eye towards future data center implementations. In particular, it is interesting to consider whether per-core power management techniques on a multi-core chip can offer sufficient leverage, and whether inter-core resource contention on CMPs may influence what sorts of core scheduling techniques make sense. In this paper, we focus on cross-server core count management techniques and evaluate a range of design questions for CMP-based data center power managers. Our results are based on a highly-parameterized simulation environment. We evaluate several possible approaches for selecting the number of cores to use, and selecting how the desired core count should be mapped onto the available CMPs. We evaluate these techniques on a real-world workload representing 24 hours of Facebook requests. For the scenarios considered, total core energy savings of 35% are possible from appropriate core count selection and mapping. Overall, this work represents a high-level view of CMPs as data center servers, and can be a guide for more detailed future studies on power management and server design questions.