Scenario Based Design Space Exploration

Nowadays, embedded systems need to support multiple applications simultaneously. It is not enough anymore to simply use a mobile phone to call somebody, but users also want to be able to simultaneously do other things, like listen to mp3’s or watch videos. This calls for a change in embedded system design which already takes the interaction between the multiple applications into account in the early design phases. Scenario based design is one of the design methodologies which is able to support embedded systems with multiple target applications. In this thesis, we introduce and study two aspects of multi-application scenario based design methodology. The first aspect is the identification of scenarios. The scenario identification technique extracts application scenarios based on a training workload and efficiently stores them into a scenario database. This scenario database can reduce the storage requirements of the event traces of all the target applications in the embedded system to only 1.4 percent of the original size. A second aspect is the scenario based design space exploration. The scenario based design space exploration exploits the scenarios in the embedded system by performing a design space exploration without exhaustively evaluating all the scenarios in the multi-application workload. This improves the performance of the scenario based design space exploration, especially when the number of applications that can execute concurrently increases. To select a representative subset of scenarios, two approaches for a scenario based design space exploration are proposed: a combined and a separate genotype approach. A small case study is done to verify our scenario based design space exploration. In this case study, several design space explorations of a multi-application workload with a MJPEG-application and a stochastically generated application are performed. Our results shows that a scenario based design space exploration outperforms a design space exploration with multi-application support where a random sample of the scenarios in the training workload is chosen. Moreover, the properties of the stochastic applications show the strengths and weaknesses of the individual approaches.