Predictable embedded multiprocessor architecture for streaming applications

The focus of this thesis is on embedded media systems that execute applications from the application domain car infotainment. These applications, which we refer to as jobs, typically fall in the class of streaming, i.e. they process on a stream of data. The jobs are executed on heterogeneous multiprocessor platforms, for performance and power efficiency reasons. Most of these jobs have firm real-time requirements, like throughput and end-to-end latency. Car-infotainment systems become increasingly more complex, due to an increase in the supported number of jobs and an increase of resource sharing. Therefore, it is hard to verify, for each job, that the realtime requirements are satisfied. To reduce the verification effort, we elaborate on an architecture for a predictable system from which we can verify, at design time, that the job’s throughput and end-to-end latency requirements are satisfied. This thesis introduces a network-based multiprocessor system that is predictable. This is achieved by starting with an architecture where processors have private local memories and execute tasks in a static order, so that the uncertainty in the temporal behaviour is minimised. As an interconnect, we use a network that supports guaranteed communication services so that it is guaranteed that data is delivered in time. The architecture is extended with shared local memories, run-time scheduling of tasks, and a memory hierarchy. Dataflow modelling and analysis techniques are used for verification, because they allow cyclic data dependencies that influence the job’s performance. Shown is how to construct a dataflow model from a job that is mapped onto our predictable multiprocessor platforms. This dataflow model takes into account: computation of tasks, communication between tasks, buffer capacities, and scheduling of shared resources. The job’s throughput and end-to-end latency bounds are derived from a self-timed execution of the dataflow graph, by making use of existing dataflow-analysis techniques. It is shown that the derived bounds are tight, e.g. for our channel equaliser job, the accuracy of the derived throughput bound is within 10.1%. Furthermore, it is shown that the dataflow modelling and analysis techniques can be used despite the use of shared memories, run-time scheduling of tasks, and caches