Browsing by Author "Wunderlich, Hans-Joachim (Prof. Dr. rer. nat. habil)"

Now showing 1 - 1 of 1

Open Access
Multi-level analysis of non-functional properties
(2014) Hatami Mazinani, Nadereh; Wunderlich, Hans-Joachim (Prof. Dr. rer. nat. habil)
System properties are usually classified into functional (behavioral) and non-functional properties (NFPs). While functional properties refer to system behavior, NFPs are attributes, or constraints of a system. Power dissipation, temperature distribution on the chip, vulnerability to soft and intermittent errors, reliability and robustness are all examples of NFPs. The exponential increase of system complexity and the transistor's smaller feature sizes pose new challenges to functional as well as non-functional properties of the system. Therefore, it becomes more important to understand and model their impact at early design phases. This work targets dynamic, quantifiable NFPs and aims at providing a basis for the accurate analysis of this class of NFPs at early design phases. It proposes an accurate and efficient NFP characterization and analysis method for complex embedded systems. An efficient NFP prediction method helps designers understand how the devices behave over time, identify NFP bottlenecks within circuits and make design trade-offs between performance and different NFPs in the product design stage. It assists manufacturers build their circuits such that no performance degradation due to specific NFPs dominate over the life of an operating device. The developed methodology is based on an efficient multi-level system-wide simulation that considers the target system application. High NFP evaluation speed is achieved using a novel piecewise evaluation technique which splits the simulation time into evaluation windows and efficiently evaluates NFP models once per window by partial linearization. The piecewise evaluation method is a fast, yet accurate replacement for a cycle-accurate NFP evaluation. To consider the mutual impact of different NFPs on each other, all NFP models are integrated into a common evaluation framework. The effect of some positive or negative feedback between different NFPs is dynamically considered during simulation. Evaluations are based on target applications instead of corner case analysis to provide a realistic prediction. The contributions of this work can be summarized as follows: (1) Generality: This work proposes a holistic, scalable NFP prediction methodology for multiple, interdependent NFPs. The NFP simulation and evaluation method is independent of a specific NFP, a particular model or a specific system or core. In addition, the method allows for multiple designs under analysis and multiple NFPs. As soon as the system is available at transaction level, it can be used for NFP estimation. (2) Speed up: The NFP-aware simulation is performed on a multi-level platform while low level simulation is accelerated using parallelism. The complete system simulation is always kept at transaction level. All the NFPs under analysis can be estimated with a single simulation run. In addition, the evaluation speed can be increased by increasing the size of the evaluation window. (3) Accuracy: The accuracy is a function of the accuracy of the selected model and the evaluation methodology. This work provides a method for integrating arbitrary low-level models into the system analysis. The right choice of low-level models may depend on the requirements for accuracy and efficiency. To preserve the low level evaluation accuracy, the required observables for piecewise evaluation are obtained at low level. The evaluation accuracy for the piecewise approach can be adjusted by calibrating the window size. Besides, rather than using statistical or worst-case analysis techniques (which may be too pessimistic in case of embedded systems with well defined applications), the complete system is simulated with the target applications and actual workloads to obtain higher accuracy for specific applications.