Compact Thermal Modeling and Control for Multi-Core Processors

Abstract

High end multi-core CPUs with high core frequency and power are reaching the point where CPU performance is limited mainly by the amount of heat that can be dissipated by the cooling equipment. High operating temperatures increase leakage power, complicate package design, and degrade reliability and performance. It may also lead to hardware failures. While one option is to spend money on a more expensive thermal package, a better solution is to apply a thermal management technique that will trade off processor’s temperature with performance loss so as to avoid thermal emergencies. To control temperature efficiently in microprocessors, thermal behavior has to be accurately predicted. Precise modeling allows the use of formal methods that offer tight control over temperature. While the most accurate way is to synthesize as many accurate thermal sensors as possible, this will need substantial area and power on top of the cost incurred. A possible way is to use performance counters that are already embedded in microprocessors to predict core temperatures. Specifically, it proposes a compact, accurate thermal model for estimating core temperatures using the built-in performance counters in the Intel® Xeon® machines. The model is developed and tested using workloads from the SPEC CPU2006 benchmark suite. Results show that the model can predict temperatures with an error of less than 1%. The thesis then presents dynamic thermal management techniques using formal feedback control theory which has the benefits of runtime adaptivity and tight control over temperature. The thermal model is used to design a proportional controller with its command setpoint dependent on the per-core CPU utilization. This controller is evaluated and compared with the default operating system governors and an ad-hoc controller. The proportional feedback controller is shown to outperform the other default or ad-hoc control policies.

Date of Award	Aug 2013
Original language	American English
Supervisor	Ibrahim Elfadel (Supervisor)