Abstract

Estimating safe upper bounds on execution times of programs is required in the design of predictable real-time systems. When multi-core, instruction pipeline, branch prediction, or cache memory are in place, due to the considerable complexity traditional static timing analysis faces, measurement-based timing analysis (MBTA) is a more tractable option. MBTA estimates upper bounds on execution times using data measured under the execution of representative execution scenarios. In this context, understanding how hardware-related events affect the executing program under analysis brings about useful information for MBTA. This paper contributes to this need by modeling the execution behavior of programs in function of hardware-related events. More specifically, for a program under analysis, we show that the number of cycles per executed instruction can be correlated to hardware-related event occurrences. We apply our modeling methodology to two architectures, ARMv7 Cortex-M4 and Cortex-A53. While all hardware events can be monitored at once in the former, the latter allows simultaneous monitoring of up to 6 out of 59 events. We then describe a method to select the most relevant hardware events that affect the execution of a program under analysis. These events are then used to model the program behavior via machine learning techniques under different execution scenarios. The effectiveness of this method is evaluated by extensive experiments. Obtained results revealed prediction errors below 20%, showing that the chosen events can largely explain the execution behavior of programs.

中文翻译：

---

论硬件相关事件对实时程序执行的影响

摘要

在设计可预测的实时系统时需要估计程序执行时间的安全上限。当多核、指令流水线、分支预测或高速缓存到位时，由于传统静态时序分析面临相当大的复杂性，基于测量的时序分析（MBTA）是一个更容易处理的选择。MBTA 使用在代表性执行场景的执行下测量的数据来估计执行时间的上限。在这种情况下，了解硬件相关事件如何影响正在分析的执行程序可为 MBTA 提供有用的信息。本文通过对硬件相关事件函数中的程序执行行为进行建模来满足这一需求。更具体地说，对于正在分析的程序，我们表明每个执行指令的周期数可以与硬件相关事件的发生相关。我们将建模方法应用于两种架构：ARMv7 Cortex-M4 和 Cortex-A53。虽然前者可以同时监视所有硬件事件，但后者允许同时监视 59 个事件中的最多 6 个。然后，我们描述了一种选择影响所分析程序执行的最相关硬件事件的方法。然后，这些事件用于通过不同执行场景下的机器学习技术对程序行为进行建模。通过大量的实验评估了该方法的有效性。获得的结果显示预测误差低于 20%，表明所选择的事件可以在很大程度上解释程序的执行行为。