From: Jason Swails <>
Date: Thu, 08 Jan 2015 08:16:37 -0500

On Thu, 2015-01-08 at 14:49 +0800, Yeng-Tseng Wang wrote:
> Dear all,
> These are my input file ( and my results. Could tell how to fix the
> problem ?
> Thanks
> Yeng-Tseng
> --------------------------------------------------------------
> &cntrl
> ntx=1, irest=0,
> ntc=2, ntf=2,
> nstlim=100000000,
> ntpr=5, ntwx=10000,
> ntwr=5,

You are writing output and restart files WAYYYYYYYYYYY too often. This
is not the cause of your problem, but your mdout file will be huge and
you will severely hurt the performance of pmemd.cuda by attempting to
write a full restart file every 5 steps (which amounts to several times
per second) -- not to mention the amount of disk thrashing you'll be
doing in the process.

Writing trajectory frames and output every step is helpful for debugging
problems (like this one), but are bad for production simulations.

> dt=0.002, cut=8.,
> ntt=1, tautp=10.0,rostat

This file should not even work with the phantom "rostat" string here.

> temp0=310.0,tempi=0.0,

I would suggest against starting from an initial temperature of 0. It
can cause the barostat to behave strangely at the beginning of the

> ntb=2, ntp=1, barostat=1,
> ioutfm=1,
> /

Some general suggestions I'll make for your input file -- use ntt=3 to
use Langevin dynamics for temperature control rather than the Berendsen
thermostat. There are known problems with the weak-coupling algorithms
for both temperature and pressure control (but moreso for temperature
control). Also, I would suggest using barostat=2, which samples
rigorously from the isothermal-isobaric ensemble (unlike the Berendsen
barostat, which is not guaranteed to) -- but more importantly in this
case is that it is also faster with pmemd.cuda.

However, none of these issues seem to me to be the root cause of your
problem. You need to make sure that you have properly minimized your
system first. I suspect there are some bad contacts in your structure
that is causing the simulation to blow up.

I would suggest going through the first tutorial to get a basic idea of
the steps used in typical MD studies:

If you have done all that, then run a very short simulation (since the
simulation dies by the 5th step, you can run a 10-step simulation)
printing out energies and trajectory frames every step (ntpr=1, ntwx=1)
and inspect those carefully to see if you can detect what the problem


Jason M. Swails
Rutgers University
Postdoctoral Researcher
AMBER mailing list
Received on Thu Jan 08 2015 - 05:30:02 PST
Custom Search