Model-driven design and development is based on the principle that all relevant attributes of a design can be associated with a design model. For performance-related attributes this is a difficult challenge. In component-oriented software development, recent approaches are based on associating performance annotations with the UML design descriptions. These are then transformed to an appropriate performance model. Problems with annotation approaches include that there are numerous design descriptions, none of which were developed for performance modelling purposes, and that the deployment aspect of UML is relatively underdeveloped. In this paper we offer a different approach, based on the Structure and Performance modelling paradigm (SP). This is a variable granularity static modelling framework designed specifically to capture those properties of software and its deployment which affect system performance. A coarsegrained structural model may be defined early in the design and refined as development proceeds, removing the need for a separate deployment model. By appropriate decomposition, it enables each designer or developer to provide quantitative estimates within their own domain. These can then be combined automatically. This helps to overcome the semantic gap between designers/developers and performance analysts. We consider two alternative avenues for exploiting the benefits of SP: either by enhancing the UML deployment diagram with OCL constraints, or by simply replacing it. The latter approach offers an evolutionary design/deployment model with the built-in advantage of quantifiability.

Traditionally performance modellers have tended to ignore the difficulty of obtaining parameter vaules which represent the resource demands of multi-layered software. In practice the use of performance engineering in large-scale systems development is limited by the cost of acquiring appropriate performance information about the various software components. However, if this information cart be reused when components are combined in different ways, then the cost of measurement can be more easily justified. Such reuse can be achieved by means of a composite work model.