Aperiodic multiprocessor scheduling for real-time stream processing applications

This thesis is concerned with the computation of buffer capacities that guarantee satisfaction of timing and resource constraints for task graphs with aperiodic task execution rates that are executed on run-time scheduled resources. Stream processing applications such as digital radio baseband processing and audio or video decoders are often firm real-time embedded systems. For a real-time embedded system, guarantees on the satisfaction of timing constraints are based on a model. This model forms a load hypothesis. In contrast to hard real-time embedded systems, firm real-time embedded systems have no safety requirements. However, firm real-time embedded systems have to be designed to tolerate the situation that the load hypothesis is inadequate. For stream processing applications, a deadline miss can lead to a drastic reduction in the perceived quality, for instance the loss of synchronisation with the radio stream in case of a digital radio can result in a loss of audio for seconds. For power and performance reasons, stream processing applications typically require a multiprocessor system, on which these applications are implemented as task graphs, with tasks communicating data values over buffers. The established time-triggered and event-triggered design paradigms for real-time embedded systems are not applicable. This is because timetriggered systems are not tolerant to an inadequate load hypothesis, for example non-conservative worst-case execution times, and event-triggered systems have no temporal isolation from their environment. Therefore, we introduce our data-driven approach. In our data-driven approach, the interfaces with the environment are time-triggered, while the tasks that implement the functionality are data-driven. This results in a system where guarantees on the satisfaction of the timing constraints can be provided given that the load hypothesis is adequate. If the load hypothesis is inadequate, then for instance non-conservative worst-case execution times do not immediately result in corrupted data values and undefined functional