The requirement of software performance as an important part of the software quality requirements is very concerning. The traditional software development methods that focus on the software performance issues later in the development process will bring high risks and high costs. If the performance of software architecture can be predicted at the early phases of the development cycle, the performance bottlenecks can be found in advance, and the possible optimization also can be worked out. In this paper, a model-based UML software architectures performance prediction method is introduced. This method selects and uses case diagrams, activity diagrams and component diagrams, and extends them to UML SPT (schedulability, performance and time) model by introducing the stereotypes and tagged values. It then transforms these UML SPT models into queueing network model through an algorithm which can handle the activity diagram with both branch nodes and confluence nodes. At last, uses the analysis theory of frequency domain to solve queuing network model to derive the performance parameters and performance bottlenecks. At the same time, the design of an automatic performance analysis tool for UML software architecture is introduced, and an instance of performance prediction is given.