You are here

Pedro Miranda defends master dissertation entitled "Efficient computation of CPW2000 using a CPU-GPU hybrid platform"

Pedro Miguel Linhares Miranda has requested the defence of his dissertation, entitled "Efficient computation of CPW2000 using a CPU-GPU hybrid platform". The defence will be held in meeting room of the Departamento de Informática, on December 15, 2011, 4:00pm. The public is invited.

The examining committee is the following:

  • João M. Fernandes (UMinho) - president
  • João M. P. Cardoso (FEUP) - opponent
  • Alberto Proença (UMinho) - supervisor
  • José L. Martins (UMinho) - supervisor

Abstract

The modelling and simulation of complex systems in natural science usually require powerful and expensive computational resources. The study of the plane wave properties in crystals, based on quantum mechanics pose challenging questions to computer scientists to improve the efficiency of the numerical methods and algorithms. Numerical libraries had a significant boost in recent years, taking advantage of multi-threaded environments.

This dissertation work addresses efficiency improvements in a plane wave package, CPW2000, developed by a physicist scientist, targeted to a heterogeneous platform with multicore CPU and CUDA enabled GPU devices. The performance bottlenecks were previously identified as being the module functions with FFT computations, and the study started with the application analysis and profiling. This study shows that (i) over 90% of the code execution time was spent in two functions, DGEMM and FFT, (ii) code efficiency of current numerical libraries is hard to improve, and (iii) DGEMM function calls were spread in the code, while FFT was concentrated in a single function.

These features were adequately explored to develop a new code version where parts of the code are computed on a multicore CPU with others taking advantage of the GPU multistreaming and parallel computing power. Experimental results show that CPU-GPU combined solutions offer near 10x speedup on the program routines that we proposed to improve, giving us a promising future work.

02.12.2011

home contacts RSS Feed last update: 16-Apr-2020 share facebook
Drupal theme by pixeljets.com D7 ver.1.1