Techniques for Scenario Prediction and Switching in System Scenario Based Designs

In recent years, computing systems have become more efficient in delivering more performance. We have now reached the limits of economically viable cooling systems and packaging technologies for high performance systems. Furthermore, mobile and handheld systems have increased in popularity, resulting in a market demanding more advanced functionality in battery powered devices. However, battery capacity is not increasing fast enough. Heterogeneous platforms are getting more popular, as different parts of an application can be computed by efficient accelerators and system resources. These resources can be switched on and off when needed, thus saving power when the application does not require them. The challenge is to efficiently map a complex system to an embedded heterogeneous platform. The increased complexity requires an automated approach to handle the complexities in stringent real-time constraints, reactive, parallel and dynamic systems, reliable performance, and maximal cost and energy efficiency. 

The Task Concurrency Management (TCM) system scenario based design methodology is being developed to meet this challenge by using a 2-phase design- and run-time approach to focus on energy efficiency rather than increasing performance. For heterogeneous multi-processor systems, optimized mapping and scheduling tasks on processing units is essential to achieve energy efficiency. At design-time a set of system scenarios are identified and a scenario prediction and switching mechanism is developed. Furthermore, accelerators can be selected for inclusion in the heterogeneous platform. At run-time, upcoming scenarios are predicted and the platform is adapted accordingly, e.g., through powering accelerators and computing resources on and off when needed. The goal of this research project is to find optimal techniques for system scenario prediction and switching in the presence of input data dynamic behavior and heterogeneous platforms with accelerators.  

An M5 simulator platform with an ARM processor will be used as a research platform to explore and develop different techniques for system scenario prediction and switching. In the initial step, a simple processor will be used to map the TCM based algorithms and explore their efficiency.

The project will be done within the scope of the IEM Strategic research area on Energy Efficient Computing Systems at NTNU and in co-operation with KUL/IMEC Leuven and/or Eindhoven.