| Resumo: | MultiProcessor Systems-on-chip (MPSoCs) are versatile and powerful platforms designed to t the needs of modern embedded applications, such as radios and audio/video decoders. However, in a MPSoC running several applications simultaneously, resources must be shared while the timing constraints of each application must be met. Embedded streaming applications mapped on a MPSoC, are often modeled using data ow graphs. Data ow graphs have the expressivity and analytical properties to naturally describe concurrent digital signal processing applications. Many scheduling strategies have been analyzed using data ow models of applications, such as Time Division Multiplexing (TDM) and Non-preemptive Non-blocking Round Robin (NPNBRR). However, few approaches have focused on a preemptive Fixed Priority (FP) scheduling scheme. Fixed Priority scheduling is a simple and often used scheduling scheme. It is easy to implement in any platform and it is quite predictable under overload conditions. This dissertation studies the temporal analysis of a set of data ow modeled applications mapped on the same resources and scheduled with a xed priority scheme. Our objective is to improve the existing analysis for Single Rate Data ow (SRDF) graphs and develop the necessary concepts and techniques to extend it for applications modeled with state-of-art data ow avor, Mode Controlled Data ow (MCDF). MCDF is a more suitable model than SRDF, since it can capture the natural dynamic behavior of modern streaming applications, and therefore, reduce the gap between model and real application. To reach our main objective we present two contributions: improvement of the existing xed priority scheduling analysis for SRDF and a complete temporal analysis technique for MCDF graphs. We propose a novel method, for a general case of a set of n graphs, to determine the worst-case response time of a low priority task by characterizing the worst-case load that higher priority tasks generated on the processor. We also demonstrate, that in the case of graphs that exhibit a single dominant periodic/sporadic source, it is possible optimize the analysis for tighter results. We validate and compare our analysis with the current state-of-art technique for periodic streaming applications, and conclude that, for all the experimented graphs, our analysis always provides tighter results. Furthermore, we propose an implementation of a complete and optimal temporal analysis technique for MCDF graphs that is based on a nite state machine description of the graphs dynamic behavior. We propose solutions to include in the analysis speci c properties of MCDF graph, such as pipelining execution and intermodal dependencies. Despite being limited to time bounded and strongly connected graphs, results obtained using this analysis are as good or better than any currently used temporal analysis technique.
|
|---|