Métodos estatísticos e de inteligência computacional para análise temporal em sistemas de tempo real.

Abstract

Real-time systems (RTS) are composed of a set of tasks (code segments) that are recurrently launched to be executed and must meet deadlines. Designing such a system in a provably correct manner requires information about the worst-case execution time (WCET) for each of its tasks. However, estimating the WCET is becoming increasingly challenging due to the high complexity of hardware and software in modern platforms. This has motivated the use of techniques to derive the probabilistic worst-case execution time (pWCET). Most existing approaches rely on measuring the execution time of system tasks on the target platform. As measurements are taken during the design time, collected samples may lead to unreliable estimates (due to possible measurement bias) or non-representative ones (due to difficulties in reproducing operational conditions). The need to make samples compatible with the assumptions of statistical modeling is an additional source of difficulty. Given the complexities presented, two studies with distinct objectives were developed. In the first study, a representation of execution time is performed based on hardware events, considering different computational intelligence tools. Specifically, for a program under analysis, it is shown that the execution time T(n) per number n of executed instructions can be correlated with occurrences of hardwarerelated events. In the second study, a new approach for Measurement-Based Probabilistic Timing Analysis (MBPTA) is presented. Unlike the usual MBPTA, which considers only T(n) as the variable of interest, this new approach incorporates a variable of interest that considers both T(n) and n. Using tuples (n, T(n)) for different values of n allows for exploring multiple execution paths. Additionally, this new approach allows the set of measurements to be evaluated and improved. For this purpose, deep neural networks (DNN) were employed. Since the measurements are considered representative, it is possible to estimate probabilistic bounds on the execution time. Experimental results indicate a difference of up to 30% between the estimates obtained using samples refined by the proposed approach and those obtained using non-refined samples. The approaches are evaluated considering different hardware and program models, and the results obtained demonstrate effectiveness in the proposed studies.