PERFORMANCE PROTOTYPING - GENERATING AND SIMULATING A DISTRIBUTED IT-SYSTEM FROM UML MODELS

In this paper, we present the concept of "performance prototyping" - the automatic generation and deployment of small components emulating the intended behaviour of real components under design into real IT-infrastructure and environments. Allowing far more effective and consistent production of prototypes than manual prototyping, performance prototyping enables the designer of systems and their infrastructure to assess the impact of various load scenarios, design choices and configuration alternatives very early in the project, and thus allows to synchronize infrastructure planning and system development closely. Rather than the "build first - tune, change & upgrade later" approach, performance prototyping enables to design, plan and build hard-, soft and middleware in closer coordination and to meet performance targets in fewer cycles. The basic concepts of the UML-based notation of performance aspects is presented which was designed to be compatible with current UML-tools and fit into their normal usage in development practises. We then discuss the interaction and differences of performance prototyping ("in-vivo performance simulation"), performance prediction in dedicated methods and tools ("in vitro" performance simulation) and load-testing as well as the differences to manual prototyping and benchmarking. A method of integrating performance prototyping into commercial UML-tools is presented, particularly with view on the challenges of generating multi-target and multi-protocol prototypes that interact across targets, platforms and protocols in the way prescribed by the model. We then describe the model, prototype, experiments and findings based on a JSP example before the conclusion of the paper.