Managing the Latency of Data-Dependent Tasks in Embedded Streaming Applications

Streaming languages are well suited to the development of embedded many-core applications that process large volumes of data with high throughput. Because they enable periodic scheduling, cyclo-static models of computation and their variants are well fitted to modern real-time applications. Nevertheless, while schedules may be functionally correct-by-construction, they can still violate timing constraints and safety requirements of critical real time embedded systems (e.g. avionics). Such violations are usually caused by delays that are not accounted for, due to resource sharing (e.g. the communication medium). In this paper, we investigate the applicability of the theory of hard-real-time scheduling for periodic tasks in the context of streaming applications. More specifically, we consider Cyclo-Static Dataflow (CSDF) graphs with variable interprocessor communication (IPC) times, and real-time constraints imposed by hardware devices or control engineers. As a result, we establish a latency constraint on the initiation times of predecessor actors on which a given actor is dependent. Based on this constraint, we show how to guarantee real-time services and reduce inconsistencies in a CSDF application by introducing a fault-tolerant procedure. We evaluate the performance of our scheduling policy for a set of 12 real-life streaming applications. We find that in most of the cases, our approach yields the maximum achievable throughput as obtained under Self-Timed scheduling (STS).