Re: [AMBER] setup time with pmemd.cuda

From: Scott Brozell <sbrozell.rci.rutgers.edu>
Date: Thu, 22 Sep 2016 09:17:00 -0400

Hi,

Well, it is encouraging that the windows-fix did not work :)

Perhaps it is time for profiling - start with cpu profiling since it
is easy; i don't know much about gpu profiling:
https://developer.nvidia.com/cuda-profiling-tools-interface

other search hits:
https://devtalk.nvidia.com/default/topic/696488/first-cuda-function-call-very-slow-more-than-a-minute-on-gtx-680-only/

scott

On Thu, Sep 22, 2016 at 11:10:34AM +0100, Hannes Loeffler wrote:
> Hi Ross,
>
> Ok, I have turned off ... the network driver (ifconfig eth0 down)...
>
> Not clear to me still though what is going on here. DNS resolution
> works fine with the host utility or gethostbyname(2).
>
>
> Many thanks,
> Hannes.
>
>
> On Wed, 21 Sep 2016 15:23:39 -0700
> Ross Walker <ross.rosswalker.co.uk> wrote:
>
> > Hi Hannes,
> >
> > Okay, well I am fresh out of ideas other than to say:
> >
> > "Have you tried turning it off and on again?"
> >
> > All the best
> > Ross
> >
> > > On Sep 21, 2016, at 1:17 PM, Hannes Loeffler
> > > <Hannes.Loeffler.stfc.ac.uk> wrote:
> > >
> > > Hi Ross,
> > >
> > > the tool reports that it has set the GPU to persistent mode but the
> > > wall clock time is still as bad as it was.
> > >
> > > Cheers,
> > > Hannes.
> > >
> > >
> > > On Wed, 21 Sep 2016 13:08:11 -0700
> > > Ross Walker <ross.rosswalker.co.uk> wrote:
> > >
> > >> Hi Hannes,
> > >>
> > >> Okay - then my next guess is the NVIDIA driver taking ages to load
> > >> for some reason. That can happen if there is a misbehaving GPU for
> > >> example. Can you try putting the driver in persistence mode and see
> > >> if that changes anything. As root:
> > >>
> > >> nvidia-smi -pm 1
> > >>
> > >> Then try again.
> > >>
> > >> All the best
> > >> Ross
> > >>
> > >>> On Sep 21, 2016, at 1:01 PM, Hannes Loeffler
> > >>> <Hannes.Loeffler.stfc.ac.uk> wrote:
> > >>>
> > >>> Hi Ross,
> > >>>
> > >>> Input is read from local hard disk and binary is on SSD. This is
> > >>> all on my workstation and I don't have any issues with these.
> > >>> Serial run looks fine but I do see that hostname is always
> > >>> 'Unknown' (also for serial and MPI CPU). I see in the code
> > >>> (master_setup.F90) that the HOSTNAME environment variable is
> > >>> queried but this doesn't work (inerr==0).
> > >>>
> > >>> Many thanks,
> > >>> Hannes.
> > >>>
> > >>>
> > >>> On Wed, 21 Sep 2016 12:16:40 -0700
> > >>> Ross Walker <ross.rosswalker.co.uk> wrote:
> > >>>
> > >>>> Hi Hannes,
> > >>>>
> > >>>> That behavior sounds really weird. My initial guess would be file
> > >>>> I/O issues meaning the reading of the input files is ridiculously
> > >>>> slow. The second might be some kind of crazy long DNS timeout.
> > >>>> PMEMD calls hostname so it record the hostname of the machine
> > >>>> running the calculation in mdout. If this is timing out for some
> > >>>> reason that might be why the setup time is always 240s.
> > >>>>
> > >>>> Both of those should be independent of GPU though. Can you try
> > >>>> the serial CPU code with the exact same input setting and see if
> > >>>> you get the same behavior?
> > >>>>
> > >>>> All the best
> > >>>> Ross
> > >>>>
> > >>>>> On Sep 21, 2016, at 11:06 AM, Hannes Loeffler
> > >>>>> <Hannes.Loeffler.stfc.ac.uk> wrote:
> > >>>>>
> > >>>>> Hi,
> > >>>>>
> > >>>>> I have a strange issue with the GPU variant of pmemd and
> > >>>>> seemingly overlong runtimes. I have benchmarked a 30000 atoms
> > >>>>> sytems which is reported to make about 32 ns/day. But the wall
> > >>>>> clock time just doesn't match with that. After running with
> > >>>>> different nstlim I realized that the setup time is constant at
> > >>>>> 240 s and top shows me that the CPU only gets busy after that
> > >>>>> time (not sure what the GPU does if it does anything).
> > >>>>>
> > >>>>> What could be the cause of this behaviour? The card is a Quadro
> > >>>>> M4000, pmemd16 compiled with gcc4.9.2/cuda 7.5.17, driver is
> > >>>>> currently 370.28 but I have also tried the current long term
> > >>>>> version.

_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Sep 22 2016 - 06:30:03 PDT
Custom Search