TypicalCPA - Typical Worst Case Compositional Performance Analysis

• Applicant: Professor Dr.-Ing. Rolf Ernst
• Subject Area: Computer Architecture, Embedded and Massively Parallel Systems
• Term: from 2015 to 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 270604602

Numerous applications in control engineering and communication technology are robust towards a limited number of deadline misses. This robustness can be exploited for the efficient dimensioning of embedded systems. However, a formal verification of the timing behavior is required, which checks the maximum frequency of deadline misses. Realizing this verification task is subject of the current project: For complex embedded systems with heterogeneous components so called (m,k)-guarantees shall be formally derived. The (m,k)-guarantees indicate that no more than m out of k consecutive executions are affected by a deadline miss. In the current project, we strive to gain an in-depth understanding of the behavior of transiently overloaded systems.In the first project phase the available single component (m,k)-verification method was significantly improved before it was integrated in the existing Compositional Performance Analysis (CPA) framework. For this purpose, the coupling mechanisms between the individual component analyses were gradually extended to a „Typical Worst Case Compositional Performance Analysis (TypicalCPA)“.What is still missing towards a practically usable system analysis is the efficient propagation of event models with overload and the derivation of end-to-end metrics. The results shall, then, be used to develop an (m,k) system design. In the first step of the new phase, the computation of the coupling mechanisms shall be optimized, in order to increase the accuracy of the (m,k)-guarantees. The coupling of component-related analyses results from the flow of activation events through the system. We want to consider the fact that the nature of an activation event can change while being propagated through the system, such that - depending on time and place – it is cause of overload or not. The obtained results will allow investigating the spatial and temporal propagation of overload in the system. Subsequently, the definition and computation of end-to-end system metrics under consideration of various activation semantics shall be advanced. This includes computing the change of the relative data age and the dispersion of the relative data age in the system due to (m,k)-losses. Moreover, the system design shall be optimized, such that demanding (m,k)-guarantees can be realized.The concluding detailed evaluation of the obtained results shall be based on industrial case studies, with a continued focus on Ethernet-backbone networks, as well as on synthetic test cases, where new results shall be provided with respect to the automatic generation of test systems with transient overload behavior.

• DFG Programme: Research Grants
• International Connection: France
• Cooperation Partner: Sophie Quinton, Ph.D.