System performance becomes more and more important in modern embedded systems. Traditionally, system performance is measured after the system has been implemented. When it fails to meet the requirement due to the design of the application at this stage, the cost of fixing them would be high. This paper presents a framework of formal method-based performance analysis (FMPA), whose goal is to detect potential performance problems at the early stage of the model-based procedure of system development. FMPA is aimed to analyse multiple performance criteria. It provides unified input models (UML-MARTE) and various formal models as analysis models. The feasibility of FMPA is illustrated by throughput and response time analysis with real-time model checking, by system reliability prediction using probabilistic model checking, and is further confirmed by the implementation of its support tool FMPAer.