[ITK] [ITK-users] Registration performance regression on machines with large number of cores

Anchit Dhar Anchit.Dhar at mathworks.com
Tue May 3 17:28:35 EDT 2016


Hello,


We are registering two images using the 'MultiResolutionImageRegistrationMethod' in ITK v4.8 and are seeing a performance degradation on machines with large number of cores.

We use the MultiResolutionImageRegistrationMethod with RegularStepGradientDescent optimizer and MeanSquares metric.

Two separate issues are observed here:

1. Performance degradation of image registration on machines with higher number of cores as compared to ones with lower number.

A Dual-core machine showed better performance with the image registration than one with 20 cores. The performance degradation observed on a machine with 20 physical cores and also on one with 6 physical (12 logical cores).

2. Performance degradation of image registration on subsequent runs on the same machine (with higher number of cores)

The performance of registration continuously regresses on subsequent runs within an application session. This issue gets resolved on setting the number of cores to a lower number (setting to number of physical cores rather than logical cores)

I have gone ahead and created the a bug for this issue.
https://issues.itk.org/jira/browse/ITK-3427

Has anyone else had similar experiences with the multi-threaded MultiResolutionImageRegistrationMethod?


Thanks.
-Anchit
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://public.kitware.com/pipermail/community/attachments/20160503/49904a2e/attachment.html>
-------------- next part --------------
_____________________________________
Powered by www.kitware.com

Visit other Kitware open-source projects at
http://www.kitware.com/opensource/opensource.html

Kitware offers ITK Training Courses, for more information visit:
http://www.kitware.com/products/protraining.php

Please keep messages on-topic and check the ITK FAQ at:
http://www.itk.org/Wiki/ITK_FAQ

Follow this link to subscribe/unsubscribe:
http://public.kitware.com/mailman/listinfo/insight-users


More information about the Community mailing list