CHAOS: Composable Highly Accurate OS-based power models

Models of computers' power consumption enable a variety of energy-efficiency optimizations and reduce data center instrumentation costs. In this paper, we present Composable, Highly Accurate, OS-based (CHAOS) full-system power models for machines and clusters. CHAOS models, which use high-level OS performance counters, yield highly accurate predictions without the intrusiveness and portability problems of hardware counters or board-level instrumentation. Furthermore, they are automatically generated by a low-overhead software framework (less than 1% CPU utilization on a mobile-class processor). We evaluate CHAOS models using MapReduce-style workloads, executed on server-class systems as well as energy-efficient low-power desktops, laptops, and embedded systems. We also generate and validate a generic, cross-platform feature set for cluster power models. To facilitate comparisons across different models and platforms, we define a metric called Dynamic Range Error (DRE) to describe how well the model characterizes the dynamic system behavior. Using this metric, we quantify the tradeoffs between model complexity and accuracy for different workloads. Our results show that the generic cross-platform feature set degrades prediction accuracy by at most 1% DRE compared to power models using the best cluster-specific feature set. 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.