Design-time performance analysis of component-based real-time systems

In current real-time systems, performance metrics are one of the most challenging properties to specify, predict and measure. Performance properties depend on various factors, like environmental context, load profile, middleware, operating system, hardware platform and sharing of internal resources. Performance failures and not satisfying related requirements cause delays, cost overruns, and even abandonment of projects. In order to avoid these performancerelated project failures, the performance properties should be obtained and analyzed already at the early design phase of a project. In this thesis we employ principles of component-based software engineering (CBSE), which enable building software systems from individual components. The advantage of CBSE is that individual components can be modeled, reused and traded. The main objective of this thesis is to develop a method that enables to predict the performance properties of a system, based on the performance properties of the involved individual components. The prediction method serves rapid prototyping and performance analysis of the architecture or related alternatives, without performing the usual testing and implementation stages. The involved research questions are as follows. How should the behaviour and performance properties of individual components be specified in order to enable automated composition of these properties into an analyzable model of a complete system? How to synthesize the models of individual components into a model of a complete system in an automated way, such that the resulting system model can be analyzed against the performance properties? The thesis presents a new framework called DeepCompass, which realizes the concept of predictable assembly throughout all phases of the system design. The cornerstones of the framework are the composable models of individual software components and hardware blocks. The models are specified at the component development time and shipped in a component package. At the component composition phase, the models of the constituent components are synthesized into an executable system model. Since the thesis focuses on performance properties, we introduce performance-related types of component models, such as behaviour, performance and resource models. The dynamics of the system execution are captured in scenario models. The essential advantage of the introduced models is that, through the behaviour of individual components and scenario models, the behaviour of the complete system is synthesized in the executable system model. Further simulation-based analysis of the obtained executable system model provides application-specific and system-specific performance property values. To support the performance analysis, we have developed a CARAT software toolkit that provides and automates the algorithms for model synthesis and simulation. Besides this, the toolkit provides graphical tools for designing alternative architectures and visualization of obtained performance properties. We have conducted an empirical case study on the use of scenarios in the industry to analyze the system performance at the early design phase. It was found that industrial architects make extensive use of scenarios for performance evaluation. Based on the inputs of the architects, we have provided a set of guidelines for identification and use of performance-critical scenarios. At the end of this thesis, we have validated the DeepCompass framework by performing three case studies on performance prediction of real-time systems: an MPEG-4 video decoder, a Car Radio Navigation system and a JPEG application. For each case study, we have constructed models of the individual components, defined the SW/HW architecture, and used the CARAT toolkit to synthesize and simulate the executable system model. The simulation provided the predicted performance properties, which we later compared with the actual performance properties of the realized systems. With respect to resource usage properties and average task latencies, the variation of the prediction error showed to be within 30% of the actual performance. Concerning the pick loads on the processor nodes, the actual values were sometimes three times larger than the predicted values. As a conclusion, the framework has proven to be effective in rapid architecture prototyping and performance analysis of a complete system. This is valid, as in the case studies we have spent not more than 4-5 days on the average for the complete iteration cycle, including the design of several architecture alternatives. The framework can handle different architectural styles, which makes it widely applicable. A conceptual limitation of the framework is that it assumes that the models of individual components are already available at the design phase.