Exploring resource/performance trade-offs for streaming applications on embedded multiprocessors

Embedded system design is challenged by the gap between the ever-increasing customer demands and the limited resource budgets. The tough competition demands ever-shortening time-to-market and product lifecycles. To solve or, at least to alleviate, the aforementioned issues, designers and manufacturers need model-based quantitative analysis techniques for early design-space exploration to study trade-offs of different implementation candidates. Moreover, modern embedded applications, especially the streaming applications addressed in this thesis, face more and more dynamic input contents, and the platforms that they are running on are more flexible and allow runtime configuration. Quantitative analysis techniques for embedded system design have to be able to handle such dynamic adaptable systems. This thesis has the following contributions: - A resource-aware extension to the Synchronous Dataflow (SDF) model of computation. - Trade-off analysis techniques, both in the time-domain and in the iterationdomain (i.e., on an SDF iteration basis), with support for resource sharing. - Bottleneck-driven design-space exploration techniques for resource-aware SDF. - A game-theoretic approach to controller synthesis, guaranteeing performance under dynamic input. As a first contribution, we propose a new model, as an extension of static synchronous dataflow graphs (SDF) that allows the explicit modeling of resources with consistency checking. The model is called resource-aware SDF (RASDF). The extension enables us to investigate resource sharing and to explore different scheduling options (ways to allocate the resources to the different tasks) using state-space exploration techniques. Consistent SDF and RASDF graphs have the property that an execution occurs in so-called iterations. An iteration typically corresponds to the processing of a meaningful piece of data, and it returns the graph to its initial state. On multiprocessor platforms, iterations may be executed in a pipelined fashion, which makes performance analysis challenging. As the second contribution, this thesis develops trade-off analysis techniques for RASDF, both in the time-domain and in the iteration-domain (i.e., on an SDF iteration basis), to dimension resources on platforms. The time-domain analysis allows interleaving of different iterations, but the size of the explored state space grows quickly. The iteration-based technique trades the potential of interleaving of iterations for a compact size of the iteration state space. An efficient bottleneck-driven designspace exploration technique for streaming applications, the third main contribution in this thesis, is derived from analysis of the critical cycle of the state space, to reveal bottleneck resources that are limiting the throughput. All techniques are based on state-based exploration. They enable system designers to tailor their platform to the required applications, based on their own specific performance requirements. Pruning techniques for efficient exploration of the state space have been developed. Pareto dominance in terms of performance and resource usage is used for exact pruning, and approximation techniques are used for heuristic pruning. Finally, the thesis investigates dynamic scheduling techniques to respond to dynamic changes in input streams. The fourth contribution in this thesis is a game-theoretic approach to tackle controller synthesis to select the appropriate schedules in response to dynamic inputs from the environment. The approach transforms the explored iteration state space of a scenario- and resource-aware SDF (SARA SDF) graph to a bipartite game graph, and maps the controller synthesis problem to the problem of finding a winning positional strategy in a classical mean payoff game. A winning strategy of the game can be used to synthesize the controller of schedules for the system that is guaranteed to satisfy the throughput requirement given by the designer