Abstract

The substantial amount of research pertaining to the usage of optical networks for communication between cores in a multicore processor underlines the need for effective communication schemes. This necessitates the exploration of the efficiency of an optical network with a suitable benchmark which compares the different features essential for having an effective communication between the cores. As far as communication in a network is considered, the parameters that are most crucial are the delays and energy consumption. This thesis focuses on an industrial-sized application from the image processing field, Canny Edge Detector, to compare the performance in terms of network parameters which are the contention delay, latency and the energy consumption with the different settings on the network on chip simulator. The Canny Edge Detector application is implemented with various software parallelization schemes for better performance as compared to the normal serialized application. Also, to analyze the effectiveness of multicore processors, a comparison among sequential and parallelized coding techniques is performed in this thesis.
Software parallelization schemes applied to the algorithm executed on optical network architectures improve the latency and delay of the network up to 60% in the best case, while the total energy consumption values have a worst case overhead of around 50%. For almost all the configuration parameters, the parallelized schemes provide much better results for the outputs than the sequential implementation. The design parameters help determine the optimal amount of resources required for efficient execution of an image processing algorithm using a moderate to heavy workload on an NoC based on the minimal delay and energy consumption values.