Runtime energy consumption estimation for server workloads based on chaotic time-series approximation

  • Authors:

  • Adam Wade Lewis;Nian-Feng Tzeng;Soumik Ghosh

  • Affiliations:

  • University of Louisiana at Lafayette, Lafayette, LA;University of Louisiana at Lafayette, Lafayette, LA;University of Louisiana at Lafayette

  • Venue:
  • ACM Transactions on Architecture and Code Optimization (TACO)
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

This article proposes a runtime model that relates server energy consumption to its overall thermal envelope, using hardware performance counters and experimental measurements. While previous studies have attempted system-wide modeling of server power consumption through subsystem models, our approach is different in that it links system energy input to subsystem energy consumption based on a small set of tightly correlated parameters. The proposed model takes into account processor power, bus activities, and system ambient temperature for real-time prediction on the power consumption of long running jobs. Using the HyperTransport and QuickPath Link structures as case studies and through electrical measurements on example server subsystems, we develop a chaotic time-series approximation for runtime power consumption, arriving at the Chaotic Attractor Predictor (CAP). With polynomial time complexity, CAP exhibits high prediction accuracy, having the prediction errors within 1.6% (or 3.3%) for servers based on the HyperTransport bus (or the QuickPath Links), as verified by a set of common processor benchmarks. Our CAP is a superior predictive mechanism over existing linear auto-regressive methods, which require expensive and complex corrective steps to address the nonlinear and chaotic aspects of the underlying physical system.