Low-cost estimation of sub-system power

Real-time, fine-grained power consumption information enables energy optimization and adaptation for operating systems and applications. Due to the high cost associated with dedicated power sensors, however, most computers do not have the ability to measure disaggregated power consumption at a component or subsystem level. We present DiPART (Disaggregated Power Analysis in Real Time), a tool to estimate subsystem power consumption based on performance (event) counters and a single, system-wide power sensor. With only one power sensor for overall system power consumption, DiPART is able to self-adapt to variations in subsystem power consumption present across nominally identical hardware. We validate the approach using a cluster of Intel Atom-based nodes that has been instrumented for subsystem (CPU, RAM and disk) power measurements. DiPART was tested across nodes in the cluster using varied benchmarks, resulting in a 40% reduction in estimation error when compared to a static model.