A system-call based model of software energy consumption without hardware instrumentation

Abstract

The first challenge to develop an energy efficient application is to measure the application's energy consumption, which requires sophisticated hardware infrastructure and sig nificant amounts of developers' time. Models and tools that estimate software energy consumption can save developers time, as application profiling is much easier and more widely available than hardware instrumentation for measuring software energy consumption. Our work focuses on modelling software energy consumption by using system calls and machine learning tech niques. This system call based model is validated against actual energy measurements from five different Android applications.These results demonstrate that system call counts can successfully model software energy consumption if the idle energy consump tion of an application is estimated or known. In the absence of any knowledge of an application's idle energy consumption, our system call based approach is still useful to compare the energy consumption among different versions of the same application.

Green Software Engineering: The Curse of Methodology

“…This rule-of-thumb specifically avoids mis-classifying many of the 90% of changes which do not affect the energy profile of an application. This model was extended and generalized as a regression problem by Shaiful et al [19] who estimate not only change, but actual energy usage. The system call based models are general and relative to the products themselves, they model applications that face the user quite well.…”

Search-Based Energy Optimization of Some Ubiquitous Algorithms

“…Often, CPU time is used as a proxy for energy use [15], but it has been found that this can be inaccurate, particularly because this omits CPU idle states [24]. As one alternative, both [16] and [17] used a regression model built on the number of a system calls made by a running program. Static or dynamic analysis of program paths with respect to known power consumption of hardware components can also be used [25].…”

“…In this paper, we describe OPACITOR, a tool implementing a new approach to measuring the energy consumption of applications running on a JVM. Previously, run time [15] or system calls [16], [17] have been used as proxies for energy use, but it has been shown [18] that, independent of the total number of CPU cycles, there can be large differences in the energy consumed by different opcodes. An alternative, high-level hardware measurements [19], [20], are non-trivial to implement and subject to noise that hinders a search process.…”

Did I make a mistake? Finding the impact of code change on energy regression

“…Power (P) is defined as the rate of work completion and measured in watts whereas energy (E) is the total amount of work done for a given time and expressed in joules [5,10,11]. The difference between power and energy, E = P · T , is one of the first requirement to understand in order to develop energy efficient system.…”

“…For example, a test suit of size 3 means there are 3 test cases in the test suit. We started with test suit size 2 and then repeated the procedure with test suites of sizes 4,7,10,13,16,27,40,64, and 73. The selected sizes of test suites are similar to Inozemtseva et al, but not identical because we used different projects in our paper.…”