Please use this identifier to cite or link to this item:
|Title:||Analysis on the Efficacy of Parallel Cuckoo Search on CUDA||Authors:||Monzer Khamash||Supervisor:||Dr. Ayed Salman||Keywords:||CUDA : Parallel Cuckoo||Issue Date:||2018||Publisher:||Kuwait university - college of graduate studies||Abstract:||The field of metaheuristic computing has blossomed into fruition in recent years thanks to advances in computing technologies. Cuckoo Search (CS) is one of the modern metaheuristic algorithms that has gained wide attention. CS has been used in many fields, and has been improved and adapted into different platforms and applications. Among the different platforms is the parallel architecture of NVidia called Compute Unified Device Architecture, CUDA for short. CUDA is a free toolkit that a developer uses to execute programs on a graphical processing unit (GPU) or a dedicated side-load processor. GPU’s have a manycore architecture making them an ideal parallel processing platform. Many popular metaheuristic algorithms have been successfully implemented on CUDA thanks to the distributed nature of the algorithms. In literature, some works have shown CS to be a versatile algorithm in serial and parallel execution. However, there is little work done on demonstrating parallel CS on CUDA. Therefore, in this thesis we have focused our efforts onto testing the efficacy of implementing CS in CUDA. We discover that this algorithm has a low efficacy for implementing on CUDA, and that it is best described as satisficing for such a massively parallel platform. We have attempted a straight-forward implementation based on the algorithm’s original publication following the authors’ recommendations, with small modifications to adjust for CUDA requirements. Afterwards, we modify the algorithm further to yield proper results fit of a parallel system. In addition, inadvertently and not a part of our initial focus, we have shown the reason the default algorithm does not perform well in a real-world scale up to 32 dimensions. We demonstrate that this shortcoming is attributed to CS’s Levy Flights susceptibility to the issue of dimensionality. We hence deduce that the design of CS is not balanced for large scale parallel computing. To accommodate these systems, the algorithm needs to be redesigned. Otherwise, it is very fit for small scale optimization.||URI:||http://hdl.handle.net/123456789/736|
|Appears in Programs:||0612 Computer Engineering|
Show full item record
Files in This Item:
|20181101 - Thesis.pdf||5,28 MB||Adobe PDF||View/Open Request a copy|
checked on Dec 9, 2019
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.