Interference Among AI-Based Components and Real-Time Tasks in Embedded Autonomous Systems

Autonomous systems such as robots combine AI-based components, like computer- vision networks, with real-time control tasks. To reduce cost, size, and power con- sumption, both increasingly run on the same embedded system-on-chip, with several CPU cores and a GPU. The two kinds of software then share caches, main memory, the power budget of the chip, and the operating system, so the activity of one can delay the other. For a real-time task, this interference lengthens the wake-up latency, the time between the instant the task should start and the instant it runs, and leaves less time before its deadline. On these platforms, interference has mostly been studied with synthetic workloads that load the memory as heavily as possible. They show how much a task can suffer, but not whether real AI components cause interference of the same size, nor through which channels it reaches the task. Studies of real vision applications mostly measure the application’s own performance. This thesis takes a first step by measuring how an industrial computer-vision appli- cation affects the mean wake-up latency of a periodic real-time task on an NVIDIA Jetson Orin Nano. The application, tkHPick, performs instance segmentation and pose estimation for robotic manipulation. The reference is MEMINTERF, a synthetic memory stressor from HeSoC-mark. The real-time task is cyclictest, in two versions. The first only wakes up and goes back to sleep, so it shows the delays caused by the operating system and the platform. The second also reads, at every activation, as much new data as a high-speed vision system processes per frame. With the first task, interference arrives mainly through the scheduler, the peripheral interrupts, and the over-current throttling of the board. Memory contention weighs little, because the accesses of the task stay in the private caches of its core. With the second, memory becomes a main channel: on an isolated core the mean latency grows about twice as much, by 39% with tkHPick and 52% with MEMINTERF. A real AI application can therefore disturb a real-time task almost as much as a synthetic stressor, even when the task has a CPU to itself.