A Statistical Approach to Power Estimation for x86 Processors

With the growing significance of green computing and difficulty in obtaining accurate real time power measurements, there is an increasing need for accurate and reliable power estimation techniques for energy-aware performance optimization. In this paper we present a statistical approach for building accurate power models using Performance Monitoring Counters (PMC) as effective proxies for x86 systems. The selection of PMC events is based on statistical methods described in literature for ARM systems. The models are trained and validated through a synthetic workload generator as well as standardized benchmarks, using k-fold cross validation technique. We demonstrate the accuracy of the resultant models across different voltage and frequency states using sophisticated reference measurements. Furthermore, we analyze the significance of the chosen PMC events, which form the framework for the regression based power models.

I. Introduction

Modern HPC systems, ranging from the current petascale to future exascale supercomputers, are constrained by power and energy consumption. As such, to balance performance and power consumption, there is a growing need for accurate real-time power information for efficient power management. Power usage information can be obtained either by using power meters or through model-based power estimation techniques implemented in hardware or software. While physical power measurements are becoming more widely available, they always present a trade-off between resolution, accuracy, scalability and cost [1]. In particular, measurement sensors need to be placed somewhere in the power distribution chain, effectively limiting the possibility to observe components with a common voltage source (e.g. multiple cores). Power estimation models can complement measurements in terms of general availability, component resolution and temporal granularity.