Despite their major impact on our modern societies, large scale distributed systems remain extremely challenging to study and assess as they aggregate million of heterogeneous elements in complex hierarchies. These artifacts, among the most complex ever built, pose a novel methodological challenge: the reductionist approach, that explains the system from a theoretical perspective through the interactions of its components, turn out to be inefficient to tackle the complexity of these systems. Experimentation reveals mandatory, but no unique experimentation methodology is sufficient to study the correctness and performance of large scale distributed systems. This document presents my research toward the practical resolution of these methodological issues. My results were implemented in the SimGrid framework, a scientific instrument to simulate realistically large scale distributed systems. The first part presents our contribution to the simulation performance, through an innovative architecture for Discrete-Event Simulators. The second part details how we integrated a full-featured model checker to the environment to formally assess properties on simulated systems. Finally, the third part extends the scope of our tool toward the study of real applications.
