This thesis focuses on addressing four research problems in designing embedded streaming systems. Embedded streaming systems are those systems thatprocess a stream of input data coming from the... Show moreThis thesis focuses on addressing four research problems in designing embedded streaming systems. Embedded streaming systems are those systems thatprocess a stream of input data coming from the environment and generate a stream of output data going into the environment. For many embeddedstreaming systems, the timing is a critical design requirement, in which the correct behavior depends on both the correctness of output data and on the time at which the data is produced. An embedded streaming system subjected to such a timing requirement is called a real-time system. Some examples of real-time embedded streaming systems can be found in various autonomous mobile systems, such as planes, self-driving cars, and drones. To handle the tight timing requirements of such real-time embedded streaming systems, modern embedded systems have been equipped with hardware platforms, the so-called Multi-Processor Systems-on-Chip (MPSoC), that contain multiple processors, memories, interconnections, and other hardware peripherals on a single chip, to benefit from parallel execution. To efficiently exploit the computational capacity of an MPSoC platform, a streaming application which is going to be executed on the MPSoC platform must be expressed primarily in a parallel fashion, i.e., the application is represented as a set of parallel executing and communicating tasks. Then, the main challenge is how to schedule the tasks spatially, i.e., task mapping, and temporally, i.e., task scheduling, on the MPSoC platform such that all timing requirements are satisfied while making efficient utilization of available resources (e.g, processors, memory, energy, etc.) on the platform. Another challenge is how to implement and run the mapped and scheduled application tasks on the MPSoC platform. This thesis proposes several techniques to address the aforementioned two challenges. Show less
This thesis presents a highly automated design framework, called DaedalusRT, and several novel techniques. As the foundation of the DaedalusRT design framework, two types of dataflow Models-of... Show moreThis thesis presents a highly automated design framework, called DaedalusRT, and several novel techniques. As the foundation of the DaedalusRT design framework, two types of dataflow Models-of-Computation (MoC) are used, one as timing analysis model and another one as the implementation model. The timing analysis model is used to formally reason about timing behavior of an application. In the context of DaedalusRT, the Mode-Aware Data Flow (MADF) MoC has been developed as the timing analysis model for adaptive streaming applications using different static modes. A novel mode transition protocol is devised to allow efficient reasoning of timing behavior during mode transitions. Based on the transition protocol, a hard real-time scheduling approach is proposed. On the other hand, the implementation model is used for efficient code generation of parallel computation, communication, and synchronization. In this thesis, the Parameterized Polyhedral Process Network (P3N) MoC has been developed to model adaptive streaming applications with parameter reconfiguration. An approach to verify the functional property of the P3N MoC has been devised. Finally, implementation of the P3N MoC on a MPSoC platform has shown that run-time performance penalty due to parameter reconfiguration is negligible. Show less
