Learning Very Large Configuration Spaces: What Matters for Linux Kernel Sizes

Linux kernels are used in a wide variety of appliances, many of them having strong requirements on the kernel size due to constraints such as limited memory or instant boot. With more than ten thousands of configuration options to choose from, obtaining a suitable trade off between kernel size and functionality is an extremely hard problem. Developers, contributors, and users actually spend significant effort to document, understand, and eventually tune (combinations of) options for meeting a kernel size. In this paper, we investigate how machine learning can help explain what matters for predicting a given Linux kernel size. Unveiling what matters in such very large configuration space is challenging for two reasons: (1) whatever the time we spend on it, we can only build and measure a tiny fraction of possible kernel configurations; (2) the prediction model should be both accurate and interpretable. We compare different machine learning algorithms and demonstrate the benefits of specific feature encoding and selection methods to learn an accurate model that is fast to compute and simple to interpret. Our results are validated over 95,854 kernel configurations and show that we can achieve low prediction errors over a reduced set of options. We also show that we can extract interpretable information for refining documentation and experts' knowledge of Linux, or even assigning more sensible default values to options.