Parametric measurement-based WCET estimation for multiprocessor platforms

Abstract

Real-time systems are bound to timing constraints. These constraints are meant to ensure that the application exhibits predictable behavior by having bounded response times. The worst case execution time (WCET) is an important property of programs, which must be bounded to allow for a response time analysis of tasks. While estimation of the WCET is a difficult problem, modern commercial off-the-shelf processors with multiple processor cores make the WCET estimation problem even harder to solve, because of the existence of shared hardware resource contention among co-running tasks. This work proposes a parametric WCET estimation tool, with which configurable and reproducible experiments can be created to investigate the co-runners’ problem for specific task sets. The created tool runs on two hardware platforms, both featuring ARM Cortex quad core processors. Three different benchmarks suites are implemented in the tool, which can be configured to run in arbitrary combinations on the processor cores. These are a set of synthetic benchmarks meant to stress the memory system, a subset of the Mälardalen WCET benchmark suite and a subset of the San Diego vision benchmark suite. The value of the tool is demonstrated through three sets of experiments. With these experiments, the effects of shared hardware resources are investigated in detail. We show that the experienced slowdown is highly dependent on multiple factors. First, a major factor is the sensitivity to the co-runners’ effects of the task itself. Two extremes are investigated, from a task being insensitive to a task being highly sensitive. Another factor is the size of the input data, which is shown to be a major contributor to the experienced slowdown. Finally, we evaluate the delayed execution of co-running tasks, which in some cases has a significant effect on the experienced slowdown. When we consider a task’s slowdown as a function of the delayed execution of its co-runners, the knowledge on this function’s behavior for specific tasks provides an optimization strategy that could be used to mitigate the problem of shared hardware resource contention.