Can we trust our energy measurements? A study on the Odroid-XU4

TL;DR

This study investigates how sampling frequency affects energy measurement accuracy on embedded devices, revealing that lower frequencies can lead to significant errors, and recommends a minimum of 350Hz for reliable results.

Contribution

The paper systematically analyzes the impact of sampling frequency on energy measurement accuracy for the Odroid-XU4, providing guidelines for reliable energy measurement practices.

Findings

Sampling at 350Hz yields accurate energy measurements comparable to high-frequency sampling.

Sampling at 1Hz can result in errors up to 80%.

Standard 4kHz sampling is more accurate but may be unnecessary for some applications.

Abstract

IoT devices, edge devices and embedded devices, in general, are ubiquitous. The energy consumption of such devices is important both due to the total number of devices deployed and because such devices are often battery-powered. Hence, improving the energy efficiency of such high-performance embedded systems is crucial. The first step to decreasing energy consumption is to accurately measure it, as we base our conclusions and decisions on the measurements. Given the importance of the measurements, it surprised us that most publications dedicate little space and effort to the description of their experimental setup. One variable of importance of the measurement system is the sampling frequency, e.g. how often the continuous signal's voltage and current are measured per second. In this paper, we systematically explore the impact of the sampling frequency on the accuracy of the measurement system. We measure the energy consumption of a Hardkernel Odroid-XU4 board executing nine Rodinia benchmarks with a wide range of runtimes and options at 4kHz, which is the standard sampling frequency of our measurement system. We show that one needs to measure at least at 350Hz to achieve equivalent results in comparison to the original power traces. Sampling at 1Hz (e.g. Hardkernel SmartPower2) results in a maximum error of 80%.