Using measurements to derive the worst-case execution time

Execution time analysis is used in the development of real-time and embedded systems to derive the timing estimates required for schedulability analysis. The execution time of the analyzed program is typically obtained by combining results from program flow analysis (such as number of iterations in loops) with low-level timing information. This paper proposes a method for low-level timing analysis based on measurements of execution times of programs executing on the actual target architecture. The essence of the method is to derive a system of linear equations from a limited number of timing measurements of an instrumented version of the considered program. The solution to these equations give execution times for program fragments, from which execution time measures for the entire program can be derived. The main advantage with this approach is that architectural modeling is not needed, hence the risk of a discrepancy between model and real system is avoided. Also, compared to the non-exhaustive measurements performed in industry today, our approach is more structured and gives complete coverage in terms of the program paths considered. We present our method in the context of a simple, but reasonably realistic, processor model and show how it can be extended to architectures with pipelines.