| Abstract: |
A fundamental aspect of the design of an embedded system is the prediction of its performance in terms of timing, memory, or energy early in the design process. The objective of this task,
typically referred to as system-level performance evaluation, is twofold. On one hand, it is instrumental for pre-validating a system design before any resources are invested for the actual
implementation and, on the other hand, the performance evaluation is a central driver for the exploration of the design space. For systems with strict performance requirements such as hard
real-time systems the performance evaluation needs to be provably correct, that is, it has to cover the worst-case performance scenarios. Furthermore, the evaluation should be fast such that it
can be employed for the exploration of large design spaces. Recent research efforts have led to analytical andmodularmethods for worst-case performance evaluation at the system level. These
methods ensure the correctness of the performance evaluation and are fast even for large-scale systems. However, they suffer from limited modelling scope and analysis accuracy. As a consequence,
when applying these methods to complex systems, one often experiences considerable abstraction losses, which lead to overly pessimistic performance results. This thesis introduces several formal
models and methods that refine the modelling capabilities of analytical performance evaluation and prevent abstraction losses. The results build on the existing framework for Modular Performance
Analysis (MPA), but apply also to other analytical formalisms. The main contributions of this thesis can be summarized as follows: • The modelling scope of analytical performance evaluation
is extended to systems with cyclic dependencies. • New models and methods are introduced for handling structured event or data streams in analytical performance evaluation. • A novel
hybrid analysis methodology is presented that combines analytical and state-based system evaluation. • New design methods for energy-efficient real-time systems are introduced. |
