|
|
SSP 1995 project summary: |
|
|
|
To re-implement the Canny Edge detector on a MIMD parallel machine, using a radial parallel image distribution.
One of the problems with parallel independent processor (MIMD) image processing algorithms it that often only a few processors have work to do, because of a non-ideal image data distribution. Recently, we have hypothesized that a radial image distribution (where an image is partitioned like slices of a pie) might have advantages over the usual 2D grid distribution.
The key idea is that, if the centre of the pie is placed at the image centre of mass, then it is more likely that each processor will have a similar amount of data to process. Also, since each slice only has 2 neighbours, communication is simpler. However, as the number of processors grows, the slices become thinner, so any initial performance gains may become lost. Because of problems with many slices meeting at a point, the distribution will probably contain a central circular slice of some radius.
The present implementation shows that the radial decomposition has a worse speedup than a regular grid decomposition due to the administration overhead.
After studying the exact times we found that this application is too small to find significant differences between the decompositions.
Akos
Frohner worked on this project.
Compressed PostScript of the project's final report is available here (319724 bytes) .
|
Webpage maintained by
mario@epcc.ed.ac.uk
|
|