SOMA v CUDA

Abstract

The topic of this thesis is implementation of SOMA in CUDA. More precisely, it is the discrete version called DSOMA. I created a reference DSOMA algorithm in C# using parallelization. Based on this algorithm I implemented additional application for CUDA, which used all benefits of programming on graphic card. In the testing phase I first compared both implemented algorithms. Subsequently I verified the effectiveness of a code in CUDA. The application has been tested on various problems, followed by searching for the best solution of the particular problem. I examined the possibility of confusion of population size with number of migrations. Finally, I tested the use of computer resources at runtime. I found out by testing that when size of population is bigger than 1000 and Jmin = 1, CUDA application is faster than C# application. I also discovered that when a bigger population is used and small number of migrations is performed, DSOMA achieves better results and in comparison with C# achieves a higher calculation speed.