6745

6287

كارشناسي ارشد

تبديل انرژي

اصفهان: دانشگاه صنعتي اصفهان، دانشكده مكانيك

1390

هشت، 101ص.: مصور، نمودار

ص.ع. به فارسي و انگليسي

محمود اشرفي زاده

ابراهيم شيراني

17/3/91

احمدرضا پيشه ور، احمد صداقت

1396/10/06

كتابنامه

مكانيك

مهندسي مكانيك

ID6287

به فارسي و انگليسي: قابل رويت در نسخه ديجيتالي

Implementation of a Parallel Incompressible Flow Solver on GPU Alireza Hajji Malayeri A Hajji@me iut ac ir Date of Submission 2012 02 18 Department of Mechanical Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language FarsiSupervisor M Ashrafizaadeh Mahmud@cc iut ac irAbstract With the advancement of science and the need to study complex physical phenomena the problem oflong simulation times in various sciences is highlighted more than ever Weather Forecast from themeteorological model and the simulation of air flow around aircrafts are examples of such cases For suchcases it is important to use fine grids for a better accuracy of the numerical simulation On the other hand the available computer resources might be limited Hence these limitations may result in using coarsergrids which in turn would reduce the accuracy One way to overcome this problem is to use the parallel processing technology on recent processorswhich in turn are constantly being improved In 2006 new generation of graphics cards were introduced tothe market that was able to process non graphical data in addition to the graphical data This newgeneration founded a new method of parallel processing that is known as parallel computing on graphicsprocessors In the present thesis this technology is used to accelerate the simulation of incompressibleflows The governing flow equations have been discretized on unstructured grids using a coupled method To benchmark the new parallel flow solver three test cases have been considered namely the square liddriven cavity the skewed lid driven cavity and a backward step flow The first two problems have beensolved on both rectangular and triangular grids whereas the last problem has only been solved onrectangular grid In addition a simpler heat conduction problem has been solved to validate and check thedeveloped parallel codes and investigate the performance gain when the GPU is used Results show that theeven a rather low end graphics processor such as a GeForce 9800 graphics processor may run as fast as 48 56 times compared to a typical CPU The overall results indicate that the graphics processor can be asource of low cost yet very practical and effective hardware to reduce the time of flow simulations Theresults also show that according to the specific architecture of these processors the highest parallelefficiency in every case occurs for some specific grid sizes Our results show that when the grid size is a multiple of 16 in all dimensions and for the specifichardware used the parallel efficiency and the speed up of the GPU code is maximized Even a small changein the grid size might have a significant negative effect on the parallel efficiency The main reason for thisis due to the number of threads in each block for the GPU To obtain the maximum performance of a GPUone needs to occupy all threads in a block and if some are left idle there could be a high penalty whichlowers the overall parallel efficiency This study shows that the new technique of parallel processing on graphics processing units is a lowcost solution to speed up the flow simulation and seems to be a promising approach for the next generationCFD codes Keywords Incompressible flow Parallel processing GPU Unstructured rectangular and triangulargrids

محمود اشرفي زاده

ابراهيم شيراني

احمدرضا پيشه ور، احمد صداقت