Per-Core Power Modeling and Control-Theoretic Power Management in Multicore Systems

  • Muhammad Yasin

Student thesis: Master's Thesis


Power and energy have emerged as first order constraints for the design of multicore processors. They can impact the packaging costs, chip reliability and in case of embedded applications battery life. Many of the existing power management schemes make use of ad hoc techniques that lack a systematic design approach and do not offer any guarantee on satisfying the given power budget. The goal of this research is to address this problem by proposing a formal control theoretic power management framework for modern multicore processors. Initially, we develop per-core power models for Xeon E5-2360 Intel Sandy Bridge processors. We make use of the Intel RAPL interface to get power and performance counter data. We analyze the performance counters crucial for modeling and build the models using linear regression. The average modeling errors for SPEC CPU2006 benchmarks is below 4% using only a small number of performance monitor counters. We implement two different types of controllers for power management on Dell PowerEdge T620 servers. The first is a traditional single-input, single-output proportional controller. The second is an optimal controller is based on advanced multiple-input, multiple-output model predictive control theory. It has the distinct advantage of meeting the power budget specification and all system and optimization constraints.
Date of AwardAug 2013
Original languageAmerican English
SupervisorIbrahim Elfadel (Supervisor)


  • Power Management; Multi-Core; Power Modeling.

Cite this