Efficient Parallelization of Complex Automotive Systems

Abstract

As the automotive industry seeks to include more and more features in its vehicles while simultaneously attempting to reduce the number of "Electronic Control Units" (ECUs) that execute the corresponding embedded software, the necessary policy shift towards multi-core technology is in full swing. In order to eventually exploit the extra processing power, there is much additional effort needed for coping with the tremendously increased complexity of such systems. This is largely due to the elaborate parallelization process (partitioning, mapping and scheduling software parts as tasks on different cores) that results in a combinatorial explosion and thus spans a vast search space. Mastering this challenge requires innovative methods and appropriate tools that are specifically designed for the creation of embedded multi-core applications or the migration of legacy software [16]. On the basis of the concept presented in [25], we use the results of its data dependency analysis performed on an "AUTOSAR" model (AUTOSAR system descriptions) to determine advantageous partitions as well as initial task-to-core mappings. Afterwards, the extracted information serves as input for the simulation within an embedded multi-core timing tool suite. Here, the initial solution is evaluated with respect to the fulfillment of basic timing requirements and metrics like cross-core communication rates, average latencies or core workloads. A subsequent optimization process improves the initial solution and enables a comparative assessment. In order to demonstrate the benefit of this approach, we apply it to two models -- a fictional mid-sized and a real-life complex one -- and show the advantage compared to a parallelization process without preceding dependency analysis and initial partition/mapping suggestions.