Re: [AMBER] Question

From: Ross Walker <ross.rosswalker.co.uk>
Date: Tue, 21 Jan 2014 19:01:28 -0800

Hi Eric,

Firstly don't use ibrun with the pmemd.cuda code - ibrun has no idea about
the number of threads, gpus needed etc. My suggestion is to request a
single cpu task and a single gpu - on public machines it generally best to
request an entire node to yourself. Then just run pmemd.cuda with
$AMBERHOME/bin/pmemd.cuda.

Second, I would avoid doing minimization with the GPU code. The latest
version of the GPU code uses the SPFP precision model which employs fixed
precision for the accumulation. The net result of this is that it only
works for systems that are close to equilibrium. Anything with spikes in
the energy will overflow the accumulators and cause linmin failures in the
minimizer. The options are either to use the SPDP or DPDP precision models
(requiring you to compile them specifically if not available) or, the
recommended approach, which is to first minimize using the CPU code and
then run the dynamics with the GPU code.

In your case on step 0 we can see that your VDW energy is massively high -
this implies an initial structure which is bad, atoms too close to each
other, so it is no wonder the dynamics explode. This was probably present
in your pre minimized structure as well and the SPFP minimizer couldn't
remove it. Start with the restart file from the CPU minimization and
things should be good.

Short version: the GPU code does not suffer bad initial structures well -
if your initial structure is hokey run minimization and heating with the
cpu code first then switch to GPU.

Finally make sure the code on the machine is fully up to date. Grok the
output for the following section:


|--------------------- INFORMATION ----------------------
| GPU (CUDA) Version of PMEMD in use: NVIDIA GPU IN USE.
| Version 12.3.1
|
| 08/07/2013

If it does not report version 12.3.1 then demand that the operators of the
machine update their AMBER installation before going any further.

All the best
Ross



On 1/21/14, 3:26 PM, "Eric Smoll" <ericsmoll.gmail.com> wrote:

>Hello Amber users,
>
>I am new to Amber and MD simulations in general so forgive me if I omit
>important information.
>
>I am using Amber12 on the Lonestar cluster at the University of Texas at
>Austin (
>https://www.tacc.utexas.edu/user-services/user-guides/lonestar-user-guide)
>.
>I was able to successfully run the calculations described in the Amber
>ionic liquid tutorial (
>http://ambermd.org/tutorials/advanced/tutorial15/Tutorial2.xhtml) using
>Lonestar's installed version of sander but my attempts to run the same
>minimization and md simulations using Lonestar's installed pmemd.cuda
>program have failed.
>
>I execute the sander run with:
>sander *ARGUMENTS*
>Similary, I execute pmemd.cuda with:
>ibrun pmemd.cuda *ARGUMENTS*
>
>Comparing the the minimization mdout files for sander and pmemd.cuda, I
>see
>several of these lines in the pmemd.cuda mdout:
>1283 .... RESTARTED DUE TO LINMIN FAILURE ...
>Also, the final results section reports values that are similar only in a
>crude sense (correct to 1 sigfig).
>
>Comparing the mdout files for the sander / pmemd.cuda molecular dynamics
>runs, the pmemd.cuda simulation has a massive initial energy and then it
>returns nonsense:
>247 NSTEP = 0 TIME(PS) = 0.000 TEMP(K) = 300.27 PRESS
>=
>12560.4
>248 Etot = 252311958.9026 EKtot = 3356.3765 EPtot =
>252308602.5261
>249 BOND = 39.0765 ANGLE = 6146.3431 DIHED =
> 94.2438
>250 1-4 NB = 129.2998 1-4 EEL = -736.3719 VDWAALS =
>252311026.9504
>251 EELEC = -8097.0157 EHBOND = 0.0000 RESTRAINT =
> 0.0000
>252 EKCMT = 236.0287 VIRIAL = -17122.5958 VOLUME =
>64008.0002
>253 Density =
> 0.7330
>254
>
>--------------------------------------------------------------------------
>----
>255
>256
>257 NSTEP = 3000 TIME(PS) = 3.000 TEMP(K) =********* PRESS
>=
>11252.9
>258 Etot = ************** EKtot = ************** EPtot =
> 460713.7796
>259 BOND = 0.0000 ANGLE = 431728.6536 DIHED =
> 5312.6252
>260 1-4 NB = -0.0001 1-4 EEL = -1.7548 VDWAALS =
>34273.4955
>261 EELEC = -10599.2398 EHBOND = 0.0000 RESTRAINT =
> 0.0000
>262 EKCMT = 12582912.0000 VIRIAL = -5530558.9855 VOLUME =
> 74552168.2535
>263 Density =
> 0.0006
>264
>
>--------------------------------------------------------------------------
>----
>The sander simulation does not show this behavior.
>
>I have scanned chapter of the manual and the
>http://ambermd.org/gpus/website but I don't anything that would
>suggest that i am doing something
>wrong. I am using
>the same parameters as the tutorial. Any insights are much appreciated.
>
>Best,
>Eric
>_______________________________________________
>AMBER mailing list
>AMBER.ambermd.org
>http://lists.ambermd.org/mailman/listinfo/amber



_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Jan 21 2014 - 19:30:03 PST
Custom Search