Exploring Thread Mapping Strategies for RTEMS SMP Based Real-Time Systems

Kurzfassung

In recent years, multi-core processor configurations have become common in general-purposecomputing. Multi-core systems offer exciting opportunities such as power-efficient processingand parallel execution. But they are not the first choice when it comes to real-time computingsystems. Difficulties in developing a predictable system is one of the serious challenges thatprevent them from being used for real-time applications. In spite of their complexity, multi-core processors are inevitable in future real-time systems, especially the ones used in spaceapplications. This is due to the rising demand for high processing power and power-efficienthardware in space systems. Space research organisations around the world are taking efforts toovercome the challenges posed by multi-core real-time systems. DLR is currently interested inaddressing the real-time scheduling problem in multi-core systems. Because of the scope forRTEMS as a real-time operating system for multi-core systems, DLR is particularly interestedin analysing the RTEMS schedulers.Real-time task scheduling is a well-studied problem on uni-core processors for which thereare mathematically proven algorithms that can guarantee schedulability of task sets. Thesealgorithms are not applicable to multi-core systems due to the added spatial dimension(numberof cores) over the already existing uni-core scheduling problem. Partitioned scheduling is oneof the multi-core scheduling approaches where the the threads are statically assigned to theavailable cores and then each core is treated as a uni-core scheduling problem. The strategyemployed in assigning the threads to the cores is an active area of research. Recent studieshave highlighted the huge potential for partitioned scheduling. Most of the analysis in thisfield are simulation studies based on an ideal mathematical model of the scheduler algorithm.The outcome of these studies are applicable only for the ideal scenario. But, due to theimplementation overhead, a practical real-time scheduler does not behave similar to the idealmodel. Moreover, the hardware dependability plays a huge part in this aberration. It is difficultto model these behaviours in simulation based studies. Therefore, an experimental analysis ofthe RTEMS scheduler is presented in this thesis.