An Approach Using Performance Models for Supporting Energy Analysis of Software Systems

Measurement-based experiments are a common solution for assessing the energy consumption of complex software systems. Since energy consumption is a metric that is sensitive to several factors, data collection must be repeated to reduce variability. Moreover, additional rounds of measurements are required to evaluate the energy consumption of the system under different experimental conditions. Hence, accurate measurements are often unaffordable because they are time-consuming. In this study, we propose a model-based approach to simplify the energy profiling process and reduce the time spent performing it. The approach uses Layered Queuing Networks (LQN) to model the scenario under test and examine the system behavior when subject to different workloads. The model produces performance estimates that are used to derive energy consumption values in other scenarios. We have considered two systems while serving workloads of different sizes. We provided 2K, 4K, and 8K images to a Digital Camera system, and we supplied bursts of 75 to 500 customers for a Train Ticket Booking System. We parameterized the LQN with the data obtained from short experiment and estimated the performance and energy in the cases of heavier workloads. Thereafter, we compared the estimates with the measured data. We achieved, in both cases, good accuracy and saved measurement time. In case of the Train Ticket Booking System, we reduced measurement time from 5 h to 35 min by exploiting our model, this reflected in a Mean Absolute Percentage Error of 9.24% in the estimates of CPU utilization and 8.72% in energy consumption predictions.