Model Discovery for Energy-Aware Computing Systems: An Experimental Evaluation

Zhichao Li, Radu Grosu, Koundinya Muppalla, Scott A. Smolka, Scott D. Stoller, and Erez Zadok

We present a model-discovery methodology for energy-aware computing systems that achieves high prediction accuracy. Model discovery, or system identification, is a critical first step in designing advanced controllers that can dynamically manage the energy-performance trade-off in an optimal manner. Our methodology favors Multiple-Inputs-Multiple-Outputs (MIMO) models over a collection of Single-Input-Single-Output (SISO) models, when the inputs and outputs of the system are coupled in a nontrivial way. In such cases, MIMO is generally more accurate than SISO over a wider range of inputs in predicting system behavior. Our experimental evaluation, carried out on a representative server workload, validates our approach. We obtained an average prediction accuracy of 77% and 76% for MIMO power and performance, respectively. We also show that MIMO models are consistently more accurate than SISO ones.