Identifying Data-Dependent System Scenarios in a Dynamic Embedded System E. Hammari 1 , F. Catthoor 2 , P. G. Kjeldsberg 1 , J. Huisken 3 , K. Tsakalis 4 and L. Iasemidis 4 1 Dept.of Electronics and Telecom., Norwegian Univ. of Science and Technology, Trondheim, Norway 2 imec and Katholieke Univ. Leuven, Leuven, Belgium; 3 imec/Holst Centre, Eindhoven, The Netherlands 4 Dept.of Electrical Engineering, Arizona State Univ., Tempe, AZ, USA Abstract— System scenario-based design methodologies are applied to reduce the costs of dynamic embedded systems. At design-time, the system is optimized for a set of system scenarios with different costs, e.g., alternative scheduling of tasks. At run-time, certain parameters are monitored, scenario changes are detected, and mapping and scheduling are reconfigured accordingly. In this process, optimized identification of parameters and system scenarios is es- sential. Previously, the parameters have been limited to control variables, or variables with a limited number of distinct values. This paper presents a scenario identification approach based on polyhedral partitioning of the parameter space for systems where the parameters may have a wide range of data-dependent values. We evaluate our approach on a biomedical application. The results indicate that with 20 system scenarios the average execution time cost can be reduced with a factor 3 and brought within 15% of the theoretically best solution for the workload-adaptive designs. Keywords: Application-specific embedded systems, run-time re- configuration, system scenario-based design 1. Introduction Increasingly, modern applications are becoming dynamic resulting in input data dependent variations in system costs, e.g., execution time and energy consumption. An over di- mensioned solution based on a few extreme workloads can be very costly, or even impossible to implement, and a workload-adaptive reconfigurable design will be necessary [14]. System scenario based design methodologies [5] provide a systematic way of constructing workload-adaptive embedded systems and have been successfully applied to multiple designs in multimedia and wireless domains [3], [11], [12], [13], [15], [17], [18]. Through structural analysis and pro- filing of the application code at design-time, a set of system scenarios with different costs is identified along with the parameters that determine the cost variations. The system is then separately optimized for each system scenario and augmented with a scenario predictor and switching mecha- nism. At run-time, the active system scenario is predicted up front from the actual parameter values and the system is switched to the most cost-optimal configuration for this system scenario. Note, that system scenarios are conceptually different from the more common use-case scenarios. While both of them aim at reducing the total costs, use-case scenarios are extracted from the obvious system parameters, modes or usage pattern which can be detected without detailed knowl- edge of the algorithmic implementation. System scenarios are identified from the observed costs and then characterized in terms of implementation parameters. System scenarios do not depend on obvious parameters, modes or usage patterns and can hence be efficiently applied even if the application do not contain any of them. This paper targets the scenario identification technique in system scenario based design methodologies, in partic- ular for systems having parameters with widely varying data-dependent values. Existing techniques assume that the parameters are control variables and/or that they have a limited number of possible parameter values. They make use of enumeration and apply a bottom-up approach to cluster these values into system scenarios [6], [4].However when the parameters are data-dependent, they may have thousands or even millions of possible data values making bottom-up clustering and enumeration-based prediction impractical (see Section 3). Our method should then instead be used because it performs a scalable top-down polyhedral partitioning of the parameter space. This is our first main contribution. Secondly, we apply our scenario identification technique to a real application and demonstrate the feasibility of our approach for different number of system scenarios. The paper is organized as follows. Section 2 gives a motivating example for our work. In Section 3 the existing techniques for scenario identification are reviewed and the necessary terminology is introduced. Our proposed approach for scenario identification is detailed in Section 4. Experi- mental results are presented in Section 5, followed by our conclusions and plans for future work. 2. Motivational example Recent biomedical applications for outpatient care have a dynamic nature and are at the same time subject to strict cost constraints. They continuously monitor patient’s signals for