Energy-efficient computing is becoming increasingly important. Among the reasons, one can mention the massive consumption of large data centers that consume as much as 180,000 homes. This trend, combined with environmental concerns, makes energy efficiency a prime technological and societal challenge. Currently, widely used power distribution units (PDUs) are often shared amongst nodes to deliver aggregated power consumption reports, in the range of hours and minutes. However, in order to improve the energy efficiency of software systems, we need to support process-level power estimation in real-time, which goes beyond the capacity of a PDUs. In particular, the CPU is considered by the research community as the major power consumer within a node and draws attention while trying to model the system power consumption. Over the last few years, a dozen of ad hoc power models have been proposed to cope with the wide diversity and the growing complexity of modern CPU architectures. In this thesis, we rather propose PowerAPI for learning power models and building software-defined power meters that provide accurate power estimation on modern architectures. With the emergence of cloud computing, we propose BitWatts and WattsKit for leveraging software power estimation in VMs and clusters. A finer level of estimation may be required to further evaluate the effectiveness of the software optimizations and we therefore propose codEnergy for helping developers to understand how the energy is really consumed by a software. We deeply assessed all above approaches, thus demonstrating the usefulness of PowerAPI to better understand the software power consumption on modern architectures.