Statistics and Computational Intelligence Tools for WCET Analysis in Multicore Architectures.

Real-time systems consist of a set of tasks (pieces of code) that are recurrently released to execute and must meet deadlines. The correct design of such a system requires information on worst-case execution time (WCET) for each of its tasks. However, estimating WCET is becoming increasingly difficult due to the high complexity of hardware and software present in nowadays modern platforms. This has motivated the use of techniques to derive probabilistic worst-case execution time (pWCET). Most existing approaches are based on measuring the execution time of each system task running in the target platform. As the measurements are carried out during design time, collected samples may lead to unreliable (due to possible measurement bias) or non-representative estimates (due to difficulties in reproducing operational conditions). The need to make the samples to conform with statistical modeling assumptions is an additional source of difficulty. In this context, our work aims at developing pWCET estimating approaches capable of circumventing the mentioned problems. We modelled execution time in function of either occurring hardware-level events, using Linear Regression Analysis; or the number of executed instructions, using Extreme Value Statistics. In both cases, we analyzed runs of 15 benchmark programs in different environments. Preliminary results indicate that both research directions offer richer information on execution time behavior when compared to existing techniques.