Author:

Keywords:

embedded systems, multiprocessors, resource management, MPSoC, system scenarios, task concurrency management, application dynamism, energy efficient

Abstract:

Developing software for contemporary embedded systems, featuring het-erogeneous multiprocessors, multiple power modes, complex data memory hierarchies, and advanced interconnects, is a daunting task. Mapping of emerging, dynamic software applications on complex MPSoC (Multi Processor System On-Chip) will be achieved through the use of Middle-ware components which will be able to mediate between embedded application software and the hardware platforms. State-of-the-art tools that help to map software tasks to hardware resources are limited because they do not take into account the inter-dependencies among processing, memory, and communication constraints. They require the right granularity of the models at different abstractions to leverage between the complexity of middle-ware components and optimization of system cost (ex: energy consumption) while satisfying the real-time constraints of applications. Moreover, the applications are becoming very dynamic due to their inputs and environment, which needs to be handled by the efficient resource management of MPSoC resources.We have focused on all these aspects to reduce the final system cost (ex: energy consumption) while optimizing the middle-ware components which includes both the design-time and run-time phases. In the design-time phase, the dynamism in the application is reflected in the appropriate number of scenarios using the System Scenario based Methodology. We have also extended the Task concurrency Management (TCM) methodology design-time phase. There, for each identified system scenario, we have provided the methodology to efficiently explore the search space of all possible mappings and extract only the few Pareto-optimal mapping solutions. We propose middle-ware components for the run-time decisions; some are specific to applications (like the scenario detection logic to identify in which scenario the application is in) and others are generic middle-ware components (like run-time resource managers for MPSoC resource processing elements, memories, and interconnect). We have proposed various methodologies for the design-time phase activities to meet the time-to market demand and also run-time phase middle-ware components to leverage between the system cost savings and overheads in terms of performance and energy consumption. We have used the 3D-WSS based Scalable Graphics game engine (graphics) as the driver application to understand the issues in the heterogeneous multiprocessor resource management and application dynamism. We have demonstrated our work on the MP3 (multimedia) decoder, H264 (multimedia) decoder, Cavity Detector (medical imaging) applications, and artificial TGFF (Task Graphs For Free) test benches. We show our experimental gains on our high-level virtual platform MPSoC functional simulator, which is developed in SystemC. Our middle-ware components experimental results have shown that we can either obtain up to 400% gains on performance or up to 70% energyreduction on the trade-off axes for the demonstrated application, when compared to state-of-the art approaches.