Workload-Driven Power and Thermal Management in Multicore Processors

Abstract

Power budget and heat dissipated from processor are the main constraints for CPU performance in multicore processor. The use of performance counters (PCs) to develop percore power and thermal proxies for multicore processors are now well established. These proxies are typically obtained using traditional linear regression techniques. These techniques have the disadvantage of requiring the full PC set regardless of the workload run by the multicore processor which is computationally burden when scaling the number of cores in the processor. Typically a computationally expensive principal component analysis is conducted to find the PCs most correlated with each workload. In this work, we use a more recent algorithm of least-angle regression to efficiently develop power and thermal proxies that include only PCs most relevant to the workload. Such PCs can be considered workload signatures in the PC space and used to categorize the workload and to trigger specific power and thermal management action. Our new power and thermal proxies are trained and tested on workloads from the PARSEC and SPEC CPU 2006 benchmarks with an average error of less than 3%. Feedback is now a common control policy for power and thermal management in multicore processors. Such feedback policies can be implemented at the hardware, firmware or operating systems level. Each of these implementations have its own overhead that consumes processor clock cycles and competes with workloads for processor resources. Also, each of these implementations uses proportional, integral and derivative control algorithms, having distinct closed-loop characteristics in terms of response time, stability, and steady-state error. Furthermore, high-level feedback policy implementations using programming languages such as C/C++, Python, Perl and Shell scripting suffer from their inability to account for processor idles and sleep state. In this research, a digital PID controller is designed and implemented for the power and thermal management in a multi-core processor. We compare this VLSI PID controller to both state-of-the-art software feedback loops and analog controllers. Result shows that digital controller achieves about an order magnitude saving in power and more than 2 orders of magnitude saving in response time with respect to an analog controller.

Date of Award	Aug 2015
Original language	American English
Supervisor	Ibrahim Elfadel (Supervisor)