Re: [AMBER] pmemd.cuda.MPI with npt

From: Ross Walker <ross.rosswalker.co.uk>
Date: Thu, 18 Apr 2013 23:35:00 -0700

Hi Himanshu,


For now if you plan to run NPT runs use the single GPU and NOT the MPI
version. We'll have a patch out that addresses a number of bugs with
pmemd.cuda.MPI for NPT shortly. That will likely fix your problem. That
said given the performance improvement is so minimal running across
multiple GPUs my advice to everyone is to save the heartache and just run
single GPU runs. If you have 4 GPUs in a node you'll get a LOT more
science done by running 4 independent copies of your system (each with
different random seeds) than you will getting 15 or 20% better throughput
for a single run.

All the best
Ross

On 4/18/13 11:01 PM, "HIMANSHU JOSHI" <himanshuphy87.gmail.com> wrote:

>Dear Adrian,
>I would like to update it point wise
>
>1. The system is running well without MPI i.e. pmemd.cuda
>2. The density of system decreases and finally killing the
>simulation with NaN in restart file.
>
>
>Any comment on the mpi problem ?
>
>
>On Thu, Apr 18, 2013 at 10:54 AM, HIMANSHU JOSHI
><himanshuphy87.gmail.com>wrote:
>
>> Hello Friends,
>>
>> I ran this system with single gpu with pmemd.gpu (no mpi),
>> here its running well. What could be the problem with MPI version ?
>> Kindly let me know.
>>
>>
>> On Wed, Apr 17, 2013 at 11:56 PM, Adrian Roitberg
>><roitberg.ufl.edu>wrote:
>>
>>> Hi
>>>
>>> could you retry your runs WITHOUT the mpi? just run one gpu please and
>>> let us know.
>>>
>>> thanks
>>> adrian
>>>
>>> On 4/17/13 2:24 PM, HIMANSHU JOSHI wrote:
>>> > Dear Gustavo,
>>> > Thanks for your kind reply.
>>> >
>>> > I am already using ioufm = 1 (for binary trajectories )and iwrap = 1
>>>.
>>> >
>>> > And the **** in the energy values is also there, eg .
>>> > 1-4 NB = ************** 1-4 EEL = ************** VDWAALS =
>>> > **************
>>> >
>>> > And most importantly the same job is running well in cpu version of
>>> pmemd.
>>> > Can it be a bug with pmemd.cuda.MPI ?
>>> >
>>> >
>>> >
>>> > On Wed, Apr 17, 2013 at 11:40 PM, Gustavo Seabra
>>> > <gustavo.seabra.gmail.com>wrote:
>>> >
>>> >> Hi,
>>> >>
>>> >> what you are seeing in the restart file is likely *not* an error:
>>>the
>>> >> numbers are just too large for the output format. As you run longer
>>> >> calculations these days, this problem will just get more and more
>>> frequent.
>>> >> Its also likely that you will find some "***" in the mdcrd file as
>>> well.
>>> >>
>>> >> Some possible solutions are:
>>> >> 1. Use iwrap=1
>>> >> 2. Use binary (NetCDF) files.
>>> >>
>>> >> Please see the manual for the details.
>>> >>
>>> >> Cheers,
>>> >>
>>> >> Gustavo Seabra
>>> >> Professor Adjunto
>>> >> Departamento de Química Fundamental
>>> >> Universidade Federal de Pernambuco
>>> >> Fone: +55-81-2126-7450
>>> >>
>>> >>
>>> >> On Wed, Apr 17, 2013 at 3:03 PM, HIMANSHU JOSHI <
>>> himanshuphy87.gmail.com
>>> >>> wrote:
>>> >>> Dear friends,
>>> >>>
>>> >>> I am running pmemd.cuda.MPI for a system with approximately 1/2
>>> million
>>> >>> atoms with dna and wate, After energy min and equilibration, I
>>>tried
>>> to
>>> >> do
>>> >>> production run with constant pressure (flag : ntb =2 ntp =1) but
>>>its
>>> >>> giving error *** in the restart file and energy values in output
>>>file
>>> >> after
>>> >>> running few steps (~ 30 ps). The same job with same input files is
>>> >> running
>>> >>> well in cpu version of pmemd (pmemd.mpi).
>>> >>>
>>> >>> Earlier the same system I ran with constant volume simulation (ntb
>>>=
>>> 1 )
>>> >> in
>>> >>> gpu with same pmemd.cuda.MPI and it ran well.
>>> >>>
>>> >>> So there is some problem with pmemd.cuda.MPI with npt simulation
>>>!!!
>>> >>>
>>> >>> Can anyone assure me.
>>> >>>
>>> >>> I have applied the latest bugfixes to amber 12 amber (Version 12.2)
>>> dated
>>> >>> (01/10/2013) for both cpu and gpu.
>>> >>>
>>> >>> Looking forward for some constructive comments from amber
>>>community .
>>> >>> Thanks for your kind attention.
>>> >>>
>>> >>>
>>> >>> --
>>> >>> *With Regards,
>>> >>> HIMANSHU JOSHI
>>> >>> Graduate Scholar, Center for Condense Matter Theory
>>> >>> Department of Physics IISc.,Bangalore India 560012*
>>> >>> ॐ सर्वे भवन्तु सुखिनः सर्वे सन्तु निरामयः।
>>> >>> सर्वे भद्रणिपश्यन्तु मा कश्चिद्दुःख भाग भवेत्॥
>>> >>> <
>>> >>>
>>> >>
>>>
>>>http://www.rediffmail.com/cgi-bin/red.cgi?red=http%3A%2F%2Fsigads%2Eredi
>>>ff%2Ecom%2FRealMedia%2Fads%2Fclick%5Fnx%2Eads%2Fwww%2Erediffmail%2Ecom%2
>>>Fsignatureline%2Ehtm%40Middle%3F&isImage=0&BlockImage=0&rediffng=0
>>> >>> _______________________________________________
>>> >>> 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
>>> >>
>>> >
>>> >
>>>
>>> --
>>> Dr. Adrian E. Roitberg
>>> Professor
>>> Quantum Theory Project, Department of Chemistry
>>> University of Florida
>>> roitberg.ufl.edu
>>>
>>>
>>> _______________________________________________
>>> AMBER mailing list
>>> AMBER.ambermd.org
>>> http://lists.ambermd.org/mailman/listinfo/amber
>>>
>>
>>
>>
>> --
>> *With Regards,
>> HIMANSHU JOSHI
>> Graduate Scholar, Center for Condense Matter Theory
>> Department of Physics IISc.,Bangalore India 560012*
>>
>
>
>
>--
>*With Regards,
>HIMANSHU JOSHI
>Graduate Scholar, Center for Condense Matter Theory
>Department of Physics IISc.,Bangalore India 560012*
>_______________________________________________
>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 Fri Apr 19 2013 - 00:00:02 PDT
Custom Search