Power-line communication (PLC) is a technology that has become very popular, as it offers easy and high-throughput connectivity in local networks. Although it is widely adopted in hybrid PLC/WiFi networks and commercially successful, PLC has received little attention from the research community. In this thesis, we build the foundations for evaluating, exploiting and boosting PLC performance. When deploying hybrid networks, there are open questions that arise: Does PLC perform better than WiFi? How can we accurately estimate PLC capacity for deciding to which medium data should be forwarded? To answer these questions, we conduct an experimental study with PLC and WiFi stations and delve into the spatio-temporal variations of capacity. We uncover crucial differences between the two mediums and prove that PLC largely extends coverage and augments network reliability. We discover that PLC links are highly asymmetric and that temporal variation occurs on three time-scales. There is a high correlation between link quality and its variability, which has a direct impact on probing overhead and on accurate link-metric estimations. We propose guidelines for link-metric estimation. The subsequent open questions are related to efficiency when stations contend for the medium. We investigate the PLC CSMA/CA protocol that bears a resemblance to that of WiFi, in the sense that it uses a binary exponential backoff. WiFi stations double the contention window only after experiencing a collision. In contrast, PLC enables the stations to also double their contention window before a collision. PLC introduces a new variable, called deferral counter, that regulates the frequency of this proactive reaction based on congestion in the network. We introduce a model for evaluating performance. The model relies on the decoupling assumption that asserts that the backoff processes of the stations are independent and has been widely used for modeling WiFi. Our model boils down to a single fixed-point equation of the collision probability. We prove the uniqueness of the solution and exploit the model to devise configurations that significantly boost performance for best-effort applications. We corroborate our model and performance gains via extensive simulation and measurements on WiFi and PLC hardware. After delving into the average performance of PLC CSMA/CA, we explore the short-term dynamics. We find analytically, experimentally and in simulation that, contrary to WiFi, PLC is short-term unfair. This yields high delay-variance (jitter) and affects delay-sensitive applications. The deferral counter introduces unfairness and determines a tradeoff between throughput and fairness, which we extensively study. We reveal that this unfairness leads to strong dependence between stations, which can penalize the accuracy of models relying on the decoupling assumption. To improve the modeling accuracy of PLC CSMA/CA, we propose another model that does not resort to the decoupling assumption and takes into account the short-term dynamics. The resulting model is more complex compared to the first one, but it performs better for small number of stations. Here too, we prove that the model admits a unique solution. This is the first model of the PLC CSMA/CA that reaches this level of accuracy. By using this model and our research on fairness, we propose an algorithm that yields configurations with low jitter, given throughput constraints and a set of possible configurations.