The field of software engineering is rapidly evolving, exposing practitioners to a growing collection of tools, languages, frameworks, and paradigms. Thus, designing a new software system consists in selecting components from this collection, which is akin to creating a configuration. Currently, the criterion to configure such systems often revolves around ease of development, which leads to oversized, power-hungry bloatware. This paradigm is not aligned with frugal or environmental concerns. Thus, this dissertation looks into the ability to leverage the configuration of software systems to optimize their performance. In particular, a specific focus is made on the energy consumption and the size of software systems. A prerequisite to optimizing a system is to understand its current performance. To gain a comprehensive understanding, this dissertation begins with a thorough analysis of a configurable software system, considering multiple performance indicators. This analysis reveals that different configurations indeed yield varying performances, and these variations can be attributed to specific options. Leveraging insights gained from performance analysis enables the creation of high-performance configurations for this system. Furthermore, certain performance indicators proved correlated across configurations. Consequently, the optimization objective can be simplified by ignoring redundant performance indicators. The creation of optimized configurations of this system was performed manually, which is only possible for small configuration spaces. To address larger configuration spaces, we propose an algorithm that outlines how to evaluate the performance of each option and subsequently improves a given configuration based on this performance data. However, optimizing a configuration by selecting high-performance options brought out limitations, as options can interact with each other: in some situations, pairing high-performance options may result in subpar performances. Similarly, low-performance options can prove unexpectedly efficient when paired together. We thus designed the algorithm to leverage such interactions. Applying this algorithm to a large set of configurations showed that most of them can reach near-optimal performances, with only a limited set of modifications. However, performance constraints are not limited to a single performance indicator. Depending on the context, the energy consumption of a system may not be the single most impactful indicator to optimize for environmental concerns. Thus, the optimization algorithm must be generalized to accommodate multiple performance indicators. This improved algorithm is validated on a pair of performance indicators: the execution time and the software size. The key finding from this validation is that half of the assessed configurations can reach a local optimum by adjusting just one setting. To automate configuration optimization, this algorithm is implemented as an open-source tool called the ICO tool suite.