Monitoring the performance of multicore embedded systems is crucial to properly ensure their timing requirements. Collecting performance data is also very relevant for optimization and validation efforts. However, the strategies used to monitor and capture data in such systems are complex to design and implement since they must not interfere with the running system beyond the point at which the system’s timing and performance characteristics start to get affected by the monitoring strategies. In this paper, we extend a monitoring framework developed in previous work to encompass three monitoring strategies, namely Active and Passive Periodic monitoring and Job-based monitoring. Periodic monitoring follows a given sampling rate. Active Periodic relies on periodic timer interrupts to guarantee deterministic sampling, while Passive Periodic trades determinism for a less invasive strategy, sampling data only when ordinary system events are handled. Job-based follows an event-driven monitoring that samples data whenever a job leaves the CPU, thus building isolated traces for each job. We evaluate them according to overhead, latency, and jitter, where none of them presented an average impact on the system execution time higher than \(0.3\%\). Moreover, a qualitative analysis is conducted in terms of data quality. On one hand, while Periodic monitoring allows for configurable sampling rates, it does not account for the rescheduling of jobs and may capture mixed traces. On the other hand, Job-based monitoring provides data samples tied to the execution of each job while disregarding sampling rate configuration and may lose track of instant measures.

监控多核嵌入式系统的性能对于正确确保其时序要求至关重要。收集性能数据对于优化和验证工作也非常相关。然而，用于监视和捕获此类系统中的数据的策略设计和实施起来很复杂，因为除了系统的时序和性能特征开始受到监视策略的影响之外，它们不得干扰正在运行的系统。在本文中，我们扩展了之前工作中开发的监控框架，以涵盖三种监控策略，即主动和被动定期监控以及基于工作的监控。定期监控遵循给定的采样率。主动周期性依靠周期性定时器中断来保证确定性采样，而被动周期性则用确定性换取侵入性较小的策略，仅在处理普通系统事件时才采样数据。基于作业遵循事件驱动的监控，每当作业离开 CPU 时都会对数据进行采样，从而为每个作业构建独立的跟踪。我们根据开销、延迟和抖动来评估它们，其中没有一个对系统执行时间的平均影响高于\(0.3\%\)。此外，还对数据质量进行了定性分析。一方面，虽然定期监控允许配置采样率，但它不考虑作业的重新安排，并且可能捕获混合跟踪。另一方面，基于作业的监控提供与每个作业的执行相关的数据样本，同时忽略采样率配置，并且可能会丢失即时测量的跟踪。