Real-Time Stream Processing in Embedded Systems

Modern real-time embedded systems often involve computational-intensive data processing algorithms to meet their application requirements. As a result, there has been an increase in the use of multiprocessor platforms. The stream processing programming model aims to facilitate the construction of concurrent data processing programs to exploit the parallelism available on these architectures. However, most current stream processing frameworks or languages are not designed for use in real-time systems, let alone systems that might also have hard real-time control algorithms. This thesis contends that a generic architecture of a real-time stream processing infrastructure can be created to support predictable processing of both batched and live streaming data sources, and integrated with hard real-time control algorithms. The thesis first reviews relevant stream processing techniques, and identifies the open issues. Then a real-time stream processing task model, and an architecture for supporting that model is proposed. An approach to the integration of stream processing tasks into a real-time environment that also has hard real-time components is presented. Data is processed in parallel using execution-time servers allocated to each core. An algorithm is presented for selecting the parameters of the servers that maximises their capacities (within an overall deadline) and ensures that hard real-time components remain schedulable. Response-time analysis is derived to guarantee that the real-time requirements (deadlines for batched data processing, and latency for each data item for live data) for the stream processing activity are met. A framework, called SPRY, is implemented to support the proposed real-time stream processing architecture. The framework supports fully-partitioned applications that are scheduled using fixed priority-based scheduling techniques. A case study based on a modified Generic Avionics Platform is given to demonstrate the overall approach. Finally, the evaluation shows that the presented approach provides a better schedulability than alternative approaches.