Virtual Timing Isolation for Mixed-Criticality Systems
Virtual Timing Isolation for Mixed-Criticality Systems
复制标题
混合关键系统的虚拟时序隔离
DOI：
10.4230/lipics.ecrts.2018.13
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
T. Ungerer
中科院分区：
文献类型：
--
作者：
Johannes Freitag;S. Uhrig;T. Ungerer
货架多层处理器的商业构成了核心之间的时机干扰，这使将它们应用于艰苦的实时系统（例如航空应用程序）中。本文提出了一个虚拟的时序隔离，将一个从所有其他内核上运行在一个核心上的一个主要应用程序。所提出的技术基于多层处理器外部的硬件，并且完全透明了主应用程序，即不需要对包括操作系统在内的软件进行修改。基本思想是应用基于单核执行的最坏情况执行时间分析，并接受多核执行期间预定义的放缓。如果放缓超过可接受的界限，则通过控制低临界核心的行为以保持主应用程序在给定界限内的进度来减少干扰。除了隔离关键应用程序的主要目标外，子目标还有效地使用了其他内核。为此，比较了控制非关键核心的三种不同机制，以有效使用完整处理器。
通过跟踪应用程序的指纹进行测量主应用程序的进度。与给定的基线（单核执行）相比，该技术可以在线量化任何执行速度的放缓。在本文中提出和评估了几种来补偿不可接受的放缓的对策，以及对指纹的准确评估。我们使用TacleBench基准套件进行的评估表明，在脉冲宽度调制控制的情况下，我们可以达到给定的可接受时间限制为4％的速度，而实际速度仅为3.27％，而在频率缩放的情况下为4.44％ 。
Commercial of the shelf multicore processors suffer from timing interferences between cores which complicates applying them in hard real-time systems like avionic applications. This paper proposes a virtual timing isolation of one main application running on one core from all other cores. The proposed technique is based on hardware external to the multicore processor and completely transparent to the main application i.e., no modifications of the software including the operating system are necessary. The basic idea is to apply a single-core execution based Worst Case Execution Time analysis and to accept a predefined slowdown during multicore execution. If the slowdown exceeds the acceptable bounds, interferences will be reduced by controlling the behavior of low-critical cores to keep the main application's progress inside the given bounds. Apart from the main goal of isolating the timing of the critical application a subgoal is also to efficiently use the other cores. For that purpose, three different mechanisms for controlling the non-critical cores are compared regarding efficient usage of the complete processor.
Measuring the progress of the main application is performed by tracking the application's Fingerprint. This technology quantifies online any slowdown of execution compared to a given baseline (single-core execution). Several countermeasures to compensate unacceptable slowdowns are proposed and evaluated in this paper, together with an accuracy evaluation of the Fingerprinting. Our evaluations using the TACLeBench benchmark suite show that we can meet a given acceptable timing bound of 4 percent slowdown with a resulting real slowdown of only 3.27 percent in case of a pulse width modulated control and of 4.44 percent in the case of a frequency scaling control.