Data-mining synthesised schedulers for hard real-time systems

The analysis of hard real-time systems, traditionally performed using RMA/PCP or simulation, is nowadays also studied as a scheduler synthesis problem, where one automatically constructs a scheduler which can guarantee avoidance of deadlock and deadline-miss system states. Even though this approach has the potential for a finer control of a hard real-time system, using fewer resources and easily adapting to further quality aspects (memory/energy consumption, jitter minimisation, etc.), synthesised schedulers are usually extremely large and difficult to understand. Their big size is a consequence of their inherent precision, since they attempt to describe exactly the frontier among the safe and unsafe system states. It nevertheless hinders their application in practise, since it is extremely difficult to validate them or to use them for better understanding the behaviour of the system. In this paper, we show how one can adapt data-mining techniques to decrease the size of a synthesised scheduler and force its inherent structure to appear, thus giving the system designer a wealth of additional information for understanding and optimising the scheduler and the underlying system. We present, in particular, how it can be used for obtaining hints for a good task distribution to different processing units, for optimising the scheduler itself (sometimes even removing it altogether in a safe manner) and obtaining both per-task and per-system views of the schedulability of the system