Footnote:
Diese Datenquelle enthält auch Bestandsnachweise, die nicht zu einem Volltext führen.
Description:
With a rapidly growing functionality of embedded real-time applications, it becomes inevitable to integrate tasks of different safety integrity levels on one many-core processor leading to a large-scale mixed-criticality system. In this process, it is not sufficient to only isolate shared architectural resources, as different tasks executing on different cores also possibly interfere via the many-core processor’s thermal management. This can possibly lead to best-effort tasks causing deadline violations for safety-critical tasks. In order to prevent such a scenario, we propose a monitoring-based hardware extension that communicates imminent thermal violations between cores via a lightweight interconnect. Building on this infrastructure, we propose a thermal strategy such that best-effort tasks can be throttled in favor of safety-critical tasks. Furthermore, assigning static voltage/frequency (V/f) levels to each safety-critical task based on their worst-case execution time may result in unnecessary high V/f levels when the actual execution finishes faster. To free the otherwise wasted thermal resources, our solution monitors the progress of safety-critical tasks to detect slack and safely reduce their V/f levels. This increases the thermal headroom for best-effort tasks, boosting their performance. In our evaluation, we demonstrate our approach on an 80-core processor to show that it satisfies the thermal and deadline requirements, and simultaneously reduces the run-time of best-effort tasks by up to 45% compared to the state of the art.