[Dune-devel] Long MPI startup

Ed Scott Wilson Garcia edscott at imp.mx
Fri Jan 15 15:00:26 CET 2021


That looks like a memory cache issue. What is the memory configuration for desktop versus laptop? There may be more than one cache involved (disk cache, Ram cache, CPU cache).

________________________________
From: Dune-devel <dune-devel-bounces at lists.dune-project.org> on behalf of Flemisch, Bernd <bernd.flemisch at iws.uni-stuttgart.de>
Sent: Tuesday, December 15, 2020 4:59 AM
To: dune-devel at lists.dune-project.org <dune-devel at lists.dune-project.org>
Subject: Re: [Dune-devel] Long MPI startup


Just to let you know that you're not alone. I experience the same issue on my rather new laptop with Ubuntu 20.04 and 20.10. Both for OpenMPI and MPICH. And only for the first of several consecutive runs, try

./a.out && ./a.out && ./a.out


I don't have that problem on my desktop which still runs on openSUSE 15.0. Kind of hard to bisect this.


Kind regards

Bernd


--
_________________________________________________________________

Bernd Flemisch
IWS, Universität Stuttgart               phone: +49 711 685 69162
Pfaffenwaldring 61              email: bernd at iws.uni-stuttgart.de
D-70569 Stuttgart           url: www.iws.uni-stuttgart.de/en/lh2/<http://www.iws.uni-stuttgart.de/en/lh2/>
_________________________________________________________________
________________________________
Von: Dune-devel <dune-devel-bounces at lists.dune-project.org> im Auftrag von Oliver Sander <oliver.sander at tu-dresden.de>
Gesendet: Dienstag, 15. Dezember 2020 11:34:48
An: dune-devel at lists.dune-project.org
Betreff: Re: [Dune-devel] Long MPI startup

Hi Markus,

thanks for your help.  Indeed, your test program also requires
about 2.5 seconds to terminate, so there really is nothing
Dune-specific about it.

It is still annoying though.  My machine is a recent-ish ThinkPad Yoga,
with 8 cores, running Debian Testing.  I suppose I'll just have to
suck it up.

Best,
Oliver

On 15.12.20 10:19, Markus Blatt wrote:
> Hi,
>
>
> On Fri, Dec 11, 2020 at 12:19:56PM +0100, Oliver Sander wrote:
>>
>> I recently noticed that on my system (current Debian testing), the call to
>>
>>    MPIHelper::instance(argc, argv);
>>
>> takes between 2 and 3 seconds to complete.  This seems like a long time to me.
>> Is it to be expected?  Or is there a bug somewhere?
>
> I don't think that we do anything specical during startup (besides calling
> MPI_Init). Hence I doubt that this is a Dune problem.
>
> Maybe you cound try a minimal MPI program to compare?
>
> #include <mpi.h>
> #include <iosteam>
>
> int main(int argc, char** argv){
>    MPI_Init(&argc, argv);
>    int size, rank;
>    MPI_Comm_size(&size, MPI_COMM_WORLD);
>    MPI_Comm_rank(&rank, MPI_COMM_WORLD);
>    //Above should be all that DUNE does
>    std::cout<<"Hi, I am rank "<<rank<<" of "<<size<<"."<<std::endl;
>    MPI_Finalize();
>    return 0;
> }
>
> BTW on my multicore machine (32 cores) it takes some time, too.
> Mabe you have a new machine? What MPI?
>
> HTH,
>
> Markus
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.dune-project.org/pipermail/dune-devel/attachments/20210115/5420baec/attachment.htm>


More information about the Dune-devel mailing list