Performance Evaluation and Modeling of a Multicore AUTOSAR System

Multicore processors have become common in personal computers and workstations for the past few years, and they are making their way to embedded devices. Meanwhile, Electronic Controller Unit (ECU) suppliers have also introduce multicore solutions in the automotive Electrics and Electronics(E/E) domain. The automotive E/E architectures are expected to adapt themselves to this change. This leads the AUTomotive Open System ARchitecture (AUTOSAR) standard to introduced multicore support in release 4.0. Because of the close ties and dependencies between the software and hardware, this adaptation is a complex task. The dependencies between hardware and software need to be handled carefully for any well performing multicore software. Based on the AUTOSAR solution, we believe that the cross-core communication could be a potential bottleneck and hence, this study measures SoftWare Components(SWC) communication time in inter-core and intra-core. In order to achieve this, a mocking of an AUTOSAR software was designed, implemented and tested on a dualcore MPC551x processor. Furthermore, a theoretical model for speedup gain prediction on heterogeneous dualcore systems is proposed. The model considers a scenario in which a task is fragmented into so-called slave tasks among cores in order to achieve speedup. By using this model, once can predict the possible speedup gain when migrating a software from a single-core to a multicore platform. The model is driven by extending Amdahl’s law and addressing the cross-core communication overhead in AUTOSAR and the heterogeneous nature of the MPC551x processor. The results show that cross-core communication has an overhead of 54%. The speedup curve shows that in tasks with large execution times, the speedup is 1.74 and that speedup is unity for tasks with an execution time about 28μs. The proposed model is evaluated by carrying out several test scenarios and comparing the results with the model which shows the model is more than 90% accurate.