With large scale and complex configurable systems, it is hard for users to choose the right combination of options (i.e., configurations) in order to obtain the wanted trade-off between functionality and performance goals such as speed or size. Machine learning can help in relating these goals to the configurable system options, and thus, predict the effect of options on the outcome, typically after a costly training step. However, many configurable systems evolve at such a rapid pace that it is impractical to retrain a new model from scratch for each new version. In this paper, we propose a new method to enable transfer learning of binary size predictions among versions of the same configurable system. Taking the extreme case of the Linux kernel with its$\approx 14,500$configuration options, we first investigate how binary size predictions of kernel size degrade over successive versions. We show that the direct reuse of an accurate prediction model from 2017 quickly becomes inaccurate when Linux evolves, up to a 32% mean error by August 2020. We thus propose a new approach for transfer evolution-aware model shifting (tEAMS). It leverages the structure of a configurable system to transfer an initial predictive model towards its future versions with a minimal amount of extra processing for each version. We show thattEAMSvastly outperforms state of the art approaches over the 3 years history of Linux kernels, from 4.13 to 5.8.

Numerous software systems are highly configurable and provide a myriad of configuration options that users can tune to fit their functional and performance requirements (e.g., execution time). Measuring all configurations of a system is the most obvious way to understand the effect of options and their interactions, but is too costly or infeasible in practice. Numerous works thus propose to measure only a few configurations (a sample) to learn and predict the performance of any combination of options' values. A challenging issue is to sample a small and representative set of configurations that leads to a good accuracy of performance prediction models. A recent study devised a new algorithm, called distance-based sampling, that obtains state-of-the-art accurate performance predictions on different subject systems. In this paper, we replicate this study through an in-depth analysis of x264, a popular and configurable video encoder. We systematically measure all 1,152 configurations of x264 with 17 input videos and two quantitative properties (encoding time and encoding size). Our goal is to understand whether there is a dominant sampling strategy over the very same subject system (x264), i.e., whatever the workload and targeted performance properties. The findings from this study show that random sampling leads to more accurate performance models. However, without considering random, there is no single "dominant" sampling, instead different strategies perform best on different inputs and non-functional properties, further challenging practitioners and researchers.