Design Automation of Multiprocessor Application-Specific Systems

Application-specific systems encompass a wide range of computing systems from small process controllers, as found in everything from washing machines to automotive engine management systems (often referred to as embedded systems); through to purpose-built supercomputers for modelling colliding super-galaxies. Whilst a range of tools exists to support designers once a system has been specified there are few if any tools available to assist in system-level specification. Our aim is to provide a tool that is not only capable of selecting the most appropriate processor cores required to meet performance deadlines, but also cost and energy criteria as well. Given an existing software application, our tool will select appropriate (possibly heterogeneous) processor cores from a design library; allocate the various software functions to each processor; and provide a static task execution schedule. Point-to-point communication links are used. Both their cost and energy requirements are also accounted for. Total system cost and energy requirements are hence provided. In cases where a solution's performance exceeds the performance deadline, the tool will use the available slack space to stretch the execution and reduce the operating voltage. Thereby resulting in a reduction in energy. Our tool uses a probabilistic search method derived from the field of Evolutionary Algorithms called Differential Evolution. We utilise the domain knowledge present in a scheduling algorithm to form the fitness function, thereby, creating a problem-space evolutionary search. Two schedulers of differing complexity are investigated: a priority-list scheduler and a critical-path scheduler. A number of benchmarks are used for comparison to previous work. In addition, an extensive study is made of two of the benchmarks to understand the effects a change of application has: Firstly, on the effectiveness of the evolutionary approach; Secondly, on the tuning parameters necessary to obtain rapid, consistent location of the global optimum solution or Pareto solutions