Software Performance Engineering (SPE) has recently considered as an important issue in the software development process. It consists on evaluate the performance of a system during the design phase. The authors have recently proposed a new methodology to generate a Stochastic Automata Network (SAN) model from a UML model, to consequently obtain performance predications from UML specifications. In this paper, they expand their idea to cover more complex UML models taking in advantage the modularity of SAN in modeling large systems. A formal description of the generation process is presented. The new extension gives rise to a serious approach in SPE that they call UML2SAN.