The architecture of a software system is a critical artifact in the software lifecycle and should be evaluated as early as possible. Recent efforts to software architecture evaluation are concentrated on scenario-based methods which are qualitative, subjective and need not any special architecture description languages. This paper investigates an approach to metrics based quantitative evaluation of UML software architecture. UML is a visual modeling language with well-formed hierarchical syntax and semantics, and is uniformly applied in various development stages. With supplementation UML has been adapted to describing software architecture. By utilization of these features three types of metrics for UML diagrams are proposed. They measure the amount of information, visual effect and connectivity degree in different UML diagrams. The application of these metrics in quantitative evaluating qualities at the architecture-level such as system scale, complexity and structural characteristics is discussed.