Please note: This PhD defence will be given online.
Rafael
Olaechea, PhD
candidate
David
R.
Cheriton
School
of
Computer
Science
Supervisors: Professors Joanne Atlee, Krzysztof Czarnecki
A software product-line (SPL) is a family of related software systems that are jointly developed and reuse a set of shared assets. Each individual software system in an SPL is called a software product and includes a set of mandatory and optional features, which are independent units of functionality. Software-analysis techniques, such as model checking, analyze a model of a software system to determine whether the software system satisfies its requirements. Because many software-analysis techniques are computationally intensive, and the number of software products in an SPL grows exponentially with the number of features in an SPL, it tends to be very time consuming to individually analyze each product of an SPL. Family-based analyses have adapted standard software-analysis techniques (e.g., model checking, type checking) to simultaneously analyze all of the software products in an SPL, reusing partial analysis results between different software products to speed up the analysis. However, these family-based analyses verify only the functional requirements of an SPL, and we are interested in analyzing the quality of service that different software products in an SPL would exhibit. Quantitative analyses of a software system model (e.g., of a weighted transition system) can estimate how long a system will take to reach its goal, how much energy a system will consume, and so on. Quantitative analyses are known to be computationally intensive. In this thesis, we investigate whether executing a family-based quantitative analysis on a model of an SPL is faster than individually analyzing every software product of the SPL.
First, we present a family-based trace-checking analysis that facilitates the reconfiguration of a dynamic software product line (DSPL), which is a type of SPL in which features can be activated or deactivated at runtime. We assessed whether executing the family-based trace-checking analysis is faster than executing the trace-checking analysis on every software product in three case studies. Our results indicated that the family-based trace checking analysis, when combined with simple data-abstraction over an SPL model’s quality-attribute values to facilitate sharing of partial-analysis results, is between 1.4 and 7.7 times faster than individually analyzing each software product. This suggests that abstraction over the quality-attribute values is key to make family-based trace-checking analysis efficient.
Second, we consider an SPL’s maximum long-term average value of a quality attribute (e.g., because it represents the long-term rate of energy consumption of the system). Specifically, the maximum limit-average cost of a weighted transition represents an upper bound on the long-term average value of a quality attribute over an infinite execution of the system. Because computing the maximum limit-average cost of a software system is computationally intensive, we developed a family-based analysis that simultaneously computes the maximum limit-average cost for each software product in an SPL. We assessed its performance compared to individually analyzing each software product in two case studies. Our results suggest that our family-based analysis will perform best in SPLs in which many products share the same set of strongly connected components.
Finally, because both of our family-based analyses require as input a timed (weighted) behaviour model of a Software Product Line, we present a method to learn such a timed (weighted) behaviour model. Specifically, the objective is to learn, for each transition t, a regression function that maps a software product to a real-valued weight that represents the duration of transition t’s execution in that software product. We apply supervised learning techniques, linear regression and regularized linear regression, to learn such functions. We assessed the accuracy of the learnt models against ground truth in two different SPL and also compared the accuracy of our method against two different state-of-the-art methods: Perfume and a Performance-Influence model. Our results indicate that the accuracy of our learnt models ranged from a mean error of 3.8% to a mean error of 193.0%. Our learnt models were most accurate for those transitions whose execution times had low variance across repeated executions of the transition in the same software product, and in which there is a linear relationship between the transition’s execution time and the presence of features in a software product.
To join this PhD defence on WebEx, please go to https://uwaterloo.webex.com/uwaterloo/j.php?MTID=m5f029cb3cc82c61171630374c8d8e012.