Accurate on-line prediction of processor and memory energy usage

Access & Terms of Use
metadata only access
Minimising energy use is an important factor in the operation of many classes of embedded systems - in particular, battery-powered devices. Dynamic voltage and frequency scaling (DVFS) provides some control over a processor's performance and energy consumption. In order to employ DVFS for managing a system's energy use, it is necessary to predict the effect this scaling has on the system's total energy consumption. Simple (yet widely-used) energy models lead to dramatically incorrect results for important classes of application programs. Predicting the energy used under scaling requires (i) a prediction of the dependency of program performance (and hence duration of execution) on the frequencies and (ii) a prediction of the power drawn by the execution as a function of the frequencies and voltages. As both of these characteristics are workload-specific our approach builds a model that, given a workload execution at one frequency setpoint, will predict the run-time and power at any other frequency setpoint. We assume temporal locality (which is valid for the vast majority of applications) so predicting the characteristics of one time slice, frame, or other instance of a task, will imply the characteristics of subsequent time slices, frames or instances (e.g. MPEG video decoding). We present a systematic approach to building these models for a hardware platform, determining the best performance counters and weights. This characterisation, done once for a particular platform, produces platform-specific but workload-independent performance and power models. We implemented the model on a real system and evaluated it under a comprehensive benchmark suite against measurements of the actual energy consumption. The results show that the model can accurately predict the energy use of a wide class of applications and is highly responsive to changes in the application behaviour.
Persistent link to this record
Link to Open Access Version
Additional Link
Snowdon, David
Petters, Stefan
Heiser, Gernot
Conference Proceedings Editor(s)
Other Contributor(s)
Corporate/Industry Contributor(s)
Publication Year
Resource Type
Conference Paper
Degree Type
UNSW Faculty