Modèle et algorithme d'ordonnancement pour architectures reconfigurables dynamiquement
Imène Benkermi · 2007
With increasing multimedia application complexity, designers try to propose hardware architectures able to support this complexity. Architectures, called SoCs (Systems on Chip), based on heterogeneous processing units (e. G. General purpose processors, FPGAs, DSP) integrated on a single chip, are more and more adopted, especially in embedded systems. These heterogeneous units can present different computing capacities in addition of different energy cost for the same portion of code. Furthermore, the presence of dynamically reconfigurable units allow to adapt the architecture to the variety of the application processing (intensive data processing and control) on data of different nature and width. These specific highly heterogeneous architecture consideration requires the use of specific methods and tools. In this thesis, software solutions to the specificity of the architectures considered is discussed. First, a general model of a system-on-chip based platform which includes dynamically reconfigurable modules is proposed. This model is essential prior to the implementation phase of an application and aims at providing a real simulation framework. The specific part of the operating system that ensures task scheduling, i. E. Dispatching tasks over processing units on the chip is then dealt with. The contribution of the proposed method in this thesis is the ability to take into account the heterogeneity of the computing unit characteristics in addition of the heterogeneity of the application task constraints in an on-line manner. To do so, we consider the extension of neural network use to on-line task scheduling on heterogeneous architectures. This method based on neural networks is well suited for on-line scheduling since the networks convergence is extremely rapid when implemented directly in silicium. An on-line scheduling algorithm is than constructed and simulations show the applicability and the efficiency of this method for heterogeneous systems.