No Hardware Required: Building and Validating Composable Highly Accurate OS-based Power Models

In this paper, we present an automatic framework for modeling nodeand cluster-level power consumption, using only portable OS-level performance counters. We evaluate these models using an emerging class of MapReduce-style workloads, executed on current server-class systems as well as energy-efficient low-power desktops, high-end laptops, and embedded systems. We also validate generic, cross-platform (with respect to model features) cluster power models for our four workloads running on six types of clusters. Our models yield highly accurate predictions without the intrusiveness and/or the correctness and portability problems of hardware performance counters or board-level measurements. We define a new metric called Dynamic Range Error (DRE) to describe how well the model characterizes the dynamic system behavior (a tighter bound than MSE or median error) and facilitate interand intra-cluster model accuracy comparisons. Using this metric, we quantify the tradeoffs between model complexity and accuracy for different workloads. The generic feature model removes the feature selection process and only degrades prediction accuracy by at most 1% DRE when compared to the best cluster power model for the workloads and clusters we studied. To the best of our knowledge, this is the most complete study of system power modeling covering such a wide variety of platforms, workloads, and models.