FEATURE
Share 
Follow 
 
AboutDownloadsDocumentsForumsWikiSource CodeIssuesNews
    Trackers
Date:
2010-07-27 19:38 		Priority:
5
State:
Closed
Submitted by:
Mike Wong (mikewong899) 		Assigned to:
Mike Wong (mikewong899)
Summary:
Inconsistent behavior for featurize on desktop Linux

Detailed description
Revision: 701

Description:

A new regression test has been introduced that creates feature vectors for 50 PDB files. The output for these files is consistent between Mac OS X and "cluster" Linux (FC11). It is inconsistent between Mac OS X and "desktop" Linux (FC8).

How to Repeat:

1. Checkout and build r701 on Mac OS X and desktop Linux
2. Run "make regression"

Repeatability: 100%

Notes:

The OS platform response pattern suggests that it could be a rounding error in featurize, but this is not known for certain.

Add A Comment: Notepad

Message  ↓
Date: 2010-08-16 02:10
Sender: Teague Sterling

It appears this has been sorted out. The remaining errors
seem to be related to memory usage in the test system.
Featuredev passed all tests and koomgang only fails when
memory is not available.

Closing bug as issue appears resolved.

Date: 2010-08-13 17:16
Sender: Teague Sterling

The remaining errors causing regression tests to fail on the koomgang
and featuredev test servers seem to be related to excess precision in
floating point arithmetic causing faulty in casts. This only seems to
manifest on 32 bit operating systems (possibly only while on 64 bit
architectures). This could likely be fixed by explicitly adjusting the
floating point operation control register to force appropriate behavior
on these systems.

As a work around we are forcing the removal of excess precision by
storing intermediate results in memory prior to integer casts in
troublesome locations. Further investigation is needed to know if this is
required throughout the code. If so, modifying the FPU rounding mode
will be preferable to forcing all intermediate results to memory as it
affects performance.

To note: upgrading to a more recent compiler may not fix this error as it
seems to be an OS/hardware issue.

Date: 2010-07-30 01:47
Sender: Mike Wong

Only 2 of the 50 PDBs have unusual behavior. One of the PDBs (1b3u) has just two lines that have differing answers. This could be due to a rounding error as we've seen before. The other (1eon) has control characters (starting line 1873). Both the PDB file and the generated pointfile from atomselector.py do not have control characters. See attached files.

Date: 2010-07-30 01:47
Sender: Mike Wong

Only 2 of the 50 PDBs have unusual behavior. One of the PDBs (1b3u) has just two lines that have differing answers. This could be due to a rounding error as we've seen before. The other (1eon) has control characters (starting line 1873). Both the PDB file and the generated pointfile from atomselector.py do not have control characters. See attached files.

Date: 2010-07-29 18:38
Sender: Teague Sterling

Just curious but is there any way we could single out the
tests which were causing the discrepancies so we can test
potential fixes quicker?

Field Old Value Date By
status_idOpen2010-08-16 02:10teague
close_date2010-08-16 02:102010-08-16 02:10teague

SimTK is maintained through Grant R01GM124443 01A1 from the National Institutes of Health (NIH). It was initially developed as part of the Simbios project funded by the NIH as part of the NIH Roadmap for Medical Research, Grant U54 GM072970.

Our Pledge and Your Responsibility

  About   Join
Version 4.2.2. Website design by Viewfarm. Icons created by SimTK team using art by GraphBerry from www.flaticon.com under a CC BY 3.0 license. Forked from FusionForge 6.0.5.
Feedback