On Wed, Nov 07, 2018, Bill Ross wrote:
>I don't have cuda+amber experience, but I would make sure you have the
>highest cuda version that your amber version supports. In general it
>sounds like it could be a cuda bug or version mismatch, based on
>'losing' a fixed amount of memory per cycle.
I don't agree with Bill here. There is generally something wrong with the
configuration of the system, leading to dynamic instabilities, when you get
this error message. (Other explanations, including Bill's are also possible,
but seem less likely to me, given our experience.)
The debug step requested before is not to run similar systems more than once,
but to try the *exact* same system more than once. That may are may not help,
depending on what happens.
Try to save a restart closer in time to when the problem occurs. At least use
the checkstructure command in cpptraj (and the checkvalidity command in
parmed) to see if there are any problems. Try backing down on how
aggressively you are changing the force field with aMD. If you can get a case
that fails with less that 95ns of simulation, comparisons with the CPU version
are also helpful.
....dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Nov 08 2018 - 06:00:04 PST
