Multi approach for real-time systems specification: case study of GPU parallel systems

Abstract

Specifying and modelling real-time systems (RTS) is not an easy task to do because there are many aspects to treat such as time constraints, scheduling and multi-tasking processing. Current RTS parallel applications deal with large amounts of data which require using devices such as GPUs. GPU is one of the main architectures that guaranties parallel execution. The specification of scheduling and timing on GPUs had always been a research problematic. UML/MARTE is an efficient semi-formal specification tool thanks to the UML's diagrams and MARTE's profiles. But it still none valid specification because it is not proved. That is why we propose to couple MARTE with the formal method Event B to have a valid specification and to validate the task scheduling on the GPU. A set of refinements had been proposed to cover the GPU specification and to generate an executable pre-code CUDA and OpenCL.