Difference between revisions of "CoE Cluster April 2012"

From Teknologisk videncenter
Jump to: navigation, search
m (Results)
m
Line 22: Line 22:
 
[[File:cuda_mandelbrot.png]]
 
[[File:cuda_mandelbrot.png]]
  
Above: Mandelbrot Set drawn in 0,12 seconds, using a single NVidia CUDA capable graphics card (GeForce GTX 460) running 32 blocks and 128 threads in each block.  Total image dimension 512 x 512 pixels.
+
''Above: Mandelbrot Set drawn in 0,12 seconds, using a single NVidia CUDA capable graphics card (GeForce GTX 460) running 32 blocks and 128 threads in each block.  Total image dimension 512 x 512 pixels.''
  
 
[[File:mpi_mandelbrot.png]]
 
[[File:mpi_mandelbrot.png]]
  
Above: Mandelbrot Set drawn in 1,2 seconds, using a Master-Worker MPI pattern with 16 worker nodes and 1 master node (which collected calculated results from different parts of the image and output them to an X Window).  Total image dimension 800 x 800 pixels.
+
''Above: Mandelbrot Set drawn in 1,2 seconds, using a Master-Worker MPI pattern with 16 worker nodes and 1 master node (which collected calculated results from different parts of the image and output them to an X Window).  Total image dimension 800 x 800 pixels.''
  
 
=Slides=
 
=Slides=

Revision as of 08:19, 1 May 2012

{{#img: image=Super-computer-artw.jpg | page=Linux Cluster til Center of Excelence/Beskrivelse til CoE West | width=200px | title=Linux Supercomputer projekt }}

Assignments

Results

Having investigated CUDA C programming for Nvidia graphics cards and the CUDA architecture, we made some performance measurements using a range of numbers of blocks and threads, executing in parallel. The test program calculated values of the Mandelbrot Set on a pixel-by-pixel basis. The Mandelbrot Set was a good choice for this test beacuse:

  • It is relatively easy to program
  • Every pixel has to calculated individually - there is no correlation between values of neighbouring pixels
  • The image can be separated into parts which can be calculated separately and in parallel
  • The time taken to calculate a complete image without parallelization is long enough to allow the performance gains from parallelization to be clearly seen
  • The resulting images are very pretty (if a little strange)! :-)

The results of the benchmarking are here, as an interactive chart (hover your mouse to find out which values are represented by each line, drag to zoom and hide/reveal curves relating to the number of blocks in the legend beneath):

Along with the CUDA C programming Guide (see Literature List), the CUDA Occupancy Calculator provides a good basis for understanding and explaining the results seen. In general, it was found that more than 32 blocks and/or 128 threads per block did not provide noticable performance gains, whereas there were significant performance gains for every step increase below these values.

Cuda mandelbrot.png

Above: Mandelbrot Set drawn in 0,12 seconds, using a single NVidia CUDA capable graphics card (GeForce GTX 460) running 32 blocks and 128 threads in each block. Total image dimension 512 x 512 pixels.

Mpi mandelbrot.png

Above: Mandelbrot Set drawn in 1,2 seconds, using a Master-Worker MPI pattern with 16 worker nodes and 1 master node (which collected calculated results from different parts of the image and output them to an X Window). Total image dimension 800 x 800 pixels.

Slides

Litteratur Liste

Evaluering