Accurate and Low-Overhead Process-Level Energy Estimation for Modern Hard Disk Drives

Abstract

Measuring the energy consumed by modern secondary storage devices and attributing it to the causative application's processes can guide optimization of application energy efficiency. Prior work has focused on hard disk drive or file system energy management using custom-built measurement systems and models, but suffers from limitations that prohibit the attribution of the hardware energy consumption to the causative processes. This paper presents a novel process-level disk drive energy estimation system that monitors the operating system kernel to predict the storage power consumption created by individual processes of a software application. Evaluation reveals that the system is accurate with prediction errors around 7% even for workloads involving simultaneous file system operations from multiple processes. The system creates virtually no power consumption overhead in the CPU and about 4% overhead in the RAM module. Results using I/O benchmarks demonstrate that a large portion of the energy usage (up to 100% in some cases) in modern disk drives can occur after the causative processes have already completed execution. Hence, results obtained using power meters or direct energy measurement systems based solely on a benchmark's execution lifetime can drastically underestimate the benchmark application's storage energy demands.

Energy efficient task scheduling on a multi-core platform using real-time energy measurements

“…Direct energy measurements [11], platform/device energy models [12] or a combination of both [13] have all been successfully applied. In this paper, EAS utilizes energy data using a direct energy measurement platform called DEEP [11] that adds current sensing instrumentation to the computing platform.…”