Measurement-based WCET estimation in multicore real-time systems

Sammanfattning: Increasing complexity and advancements of hardware continue to make WCET analysis a non-trivial problem especially on modern COTS processors. Finding a tightand reliable WCET upper bound is vital for schedulability analysis and increasing theresource utilization of a real-time system. On a system with multiple coresunpredictable interference to a programs execution time is expected. This behaviouris firstly difficult to model in details and secondly creates too high computationalcomplexity to analyze statically. Furthermore, results from such analysis are oftenvery pessimistic to be applicable in practice.A measurement-based approach is presented in this work where a program undertest is run with other programs on a multiprocessor to create possible worst casescenarios. The focus of the work is to explore the interference created by theco-running tasks and their influence on the WCET of a program. A closer look istaken at the properties of the hardware and software which cause these possiblevariations. The experiments are conducted using both a real target hardware, amulticore platform and also a simulator mimicking the hardware. The experimental results show that the WCET highly depends on the memory accesscharacteristics of the program in question and also the programs co-running at thesame time on other cores. It is shown that the combination of these characteristics,of the running and co-running programs, leads to large increases of execution timevariation. This behaviour is observed in both experiments with the real hardware aswell as the simulator with regards to the overall reaction of co-running tasks. This demonstrates that a simulator is a viable option for the estimation of WCET. The conclusion is that the WCET of a parallel program on a multicore platform can be bounded if the bad memory access patterns of all co-running tasks can be avoided, which may result in more optimized results and higher resource utilization.