Search-Based Energy Testing of Android

Automated construction of energy test oracles for Android

“…Prior research has shown that energy defects manifest themselves under specific contextual settings [28]. Specifically, some energy defects, e.g., wakelocks and resource leaks, happen under specific sequences of lifecycle callbacks, while others manifest themselves under peculiar hardware states, e.g., poor network signal, no network connection, or low battery.…”

“…Proper feature selection, i.e., feature engineering, is fundamental to the application of DL techniques, as the quality and quantity of features greatly impact the utility of a model. We chose our features to reflect the changes in the state of app lifecycle and hardware elements during the execution of a test, as these factors have shown to play an important role in manifestation of energy defects [28]. Figure 2 demonstrates the representation of SV at the highest level in the first row and at a finer granularity for all the entries.…”

“…We referred to Android documentation [1,2,6] to determine the former. For the latter, we followed a systematic approach similar to that presented in the prior work [28] to obtain all the Android APIs and constant values in the libraries that allow developers to monitor or utilize hardware components. Specifically, we performed a keyword-based search on the Android API documentation to collect hardware-relevant APIs and fields, identified all the hardware states that can be changed or monitored at the software level, and constructed State Vector as demonstrated in Figure 2.…”

“…While ACETON is able to effectively detect the outcome of tests in mutants, we also wanted to see how it performs on Android apps that have real but similar energy defects. To that end, we referred to a prior work [28], which provides a dataset of 14 Android apps with real energy defects. Each app is accompanied by a test generated using their test generation tool, which is manually confirmed to reproduce the energy defect.…”

“…In fact, advancements on mobile app testing have in large part focused on functional correctness, rather than non-functional properties, such as energy efficiency [29]. To alleviate this shortcoming, recent studies have tried to generate effective energy tests [19,28]. While the proposed techniques have shown to be effective for generating energy-aware test inputs, they either use manually constructed oracles [19,28] or rely on observation of power traces, i.e., series of energy consumption measurements throughout the test execution, to determine the outcome of energy testing [9,10,29].…”