Hi All,
I've noticed some pmemd.cuda bugs have appeared for systems I'm running
after the most recent GPU patch for COM restraints. My system does not
have COM restraints, but is kinda unusual in that there are three
zero-charge, zero-LJ anchor atoms which are used to orient a host-guest
system in the solvent box.
Bugs:
1) GNU/CUDA 5.5/Amber14 pmemd.cuda chokes at startup with all 37
restraints. (see mdin.37rest in repro)
2) I found that if I removed 10 restraints, GNU/CUDA5.5/Amber14/pmemd.cuda
no longer choked (see mdin.27rest). However after completing the run it
spit out a glibc error.
3) The is no glibc error if I remove all restraints. But if there are two
distance restraints in the disang file and one of them is to an anchor
atom, there is a glibc error (mdin.2rest-err). If there are two distance
restraints but both of them are on host atoms only, there is no glibc error
(mdin.2rest-noerr).
4) Intel compilers give the same results EXCEPT bug 1) does not occur but
instead results in the glibc error after successfully finishing the MD job.
None of these errors occur with pre-Update.10 Amber14 code. They do not
occur for CPU code either.
They DO occur with GIT development code. They do occur for GNU/CUDA 6.5.
They do occur for GTX 580, 680, and 780s.
REPRO: https://www.dropbox.com/s/k8r6ab8o7j5f2u9/gpu-rest-err.tar.gz
Let me know if you need more information from me. Should I submit this to
bugzilla?
Thanks,
--Niel
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Feb 05 2015 - 23:00:02 PST
