Hardware architectures based on multiple cores, cache memory and branch prediction usually preclude the application of classical methods for determining execution time bounds for real-time tasks. As such bounds are fundamental in the designing of real-time systems, Measurement-Based Probabilistic Timing Analysis has been employed. A common choice is the derivation of probabilistic Worst-Case Execution Time via the use of Extreme Value Theory, a branch of statistics used to estimate the probability of rare events that are more extreme than observations. However, pWCET estimations are usually reported in a controlled or simulated environment. In this paper we apply MBPTA in a real multi-core platform, namely Raspberry Pi 3B, taking into consideration possible interference due to operating system and concurrent activities. The results indicate that although EVT is effective, it does not always produce adequate models and coherent pWCET estimations. As MBPTA is primarily called for when classical methods are not applicable, as it is the case for the studied platform, the results reported in this paper highlight risks and vulnerabilities when applying MBPTA-EVT for pWCET inference.

基于多核、高速缓冲存储器和分支预测的硬件架构通常妨碍应用经典方法来确定实时任务的执行时间界限。由于此类界限是实时系统设计的基础，因此采用了基于测量的概率时序分析。常见的选择是通过使用极值理论推导概率最坏情况执行时间，极值理论是统计学的一个分支，用于估计比观察结果更极端的罕见事件的概率。然而，pWCET 估计通常是在受控或模拟环境中报告的。在本文中，我们将MBPTA应用于真实的多核平台，即Raspberry Pi 3B，考虑到操作系统和并发活动可能产生的干扰。结果表明，虽然 EVT 是有效的，但它并不总是能产生足够的模型和一致的 pWCET 估计。由于 MBPTA 主要在经典方法不适用时才需要，正如所研究的平台的情况一样，本文报告的结果强调了应用 MBPTA-EVT 进行 pWCET 推理时的风险和漏洞。