[gmx-developers] BlueGene

[Berk Hess]

Hi,

Mathias Puetz wrote optimized kernels for the blue gene
which are about a factor of two faster than the normal c kernels.

Berk.

Axel Kohlmeyer wrote:
> On Thu, 31 Jul 2008, Roland Schulz wrote:
>
> RS> Hi,
> RS> 
> RS> I already commited them to CVS. This timing is after fixing the include
> RS> statements. Without the intrinsic (the default) non-bonded kernel I get only
> RS> .3Gflop/s (/2 of with).
>
> roland,
>
> have you compared to the fortran kernel? last time i compiled gmx on
> a bg/l the fortran kernel was significantly faster than the c kernel.
>
> i also remember having to make a few adjustments that i posted here,
> but didn't show up in the cvs. we've stopped using gmx on the bg/l
> (too depressingly slow) since, so i didn't care much anymore.
>
> i'm certain, i have the scripts and patches somewhere stashed away.
> so if you are curious, give me a holler and i'll start digging.
>
> cheers,
>    axel.
>
> RS> 
> RS> Because this is a compile time problem it should as well not compile on
> RS> BlueGene/L with the special kernel enabled.
> RS> The Gromacs Wiki on BlueGene explained how to compile with the default
> RS> non-bonded kernel. Thus I'm expecting that people were running with the
> RS> default (non-instrinsic) kernel. I added an remark to the wiki that one
> RS> should compile with --enable-bluegene.
> RS> 
> RS> The configure had a bug too. The configure help said that the option is
> RS> colled --ennable-bluegene-asm but it was --enable-bluegene. This I also
> RS> commited.
> RS> 
> RS> Roland
> RS> 
> RS> On Thu, Jul 31, 2008 at 3:35 AM, <hessb at mpip-mainz.mpg.de> wrote:
> RS> 
> RS> > Hi,
> RS> >
> RS> > So there is really something wrong in the include statements?
> RS> > Is this only for bluegene/P, or probably also for bluegene/L?
> RS> > I know people who are running Gromacs on a bluegene/L and have
> RS> > not heard about these issues.
> RS> > Could you mail the fixes?
> RS> >
> RS> > Berk.
> RS> >
> RS> > > Hi,
> RS> > >
> RS> > > it was mostly wrong names in the #include statements.
> RS> > >
> RS> > > It works on BlueGene/P and is much faster than the C non-bonded kernel
> RS> > > (about 2x). With this intrinsic C kernel one gets 18% of peak. Because
> RS> > > BlueGene doesn't have SIMD commands specific for single precision (thus
> RS> > it
> RS> > > does only 2 at a time instead of 4) the performance difference to Opteron
> RS> > > is
> RS> > > greater than for double precision code. The performance difference to a
> RS> > > Barcelona 2.1Ghz core is 5.7x (at 20% peak).
> RS> > >
> RS> > > Roland
> RS> > >
> RS> > > On Wed, Jul 30, 2008 at 12:29 PM, Roland Schulz <roland at utk.edu> wrote:
> RS> > >
> RS> > >> Hi,
> RS> > >>
> RS> > >> it seems that the blugene non-bonded kernel in the CVS version is
> RS> > >> incomplete. E.g. the nb_kernel_bluegene_test_asm.h is missing as far as
> RS> > >> I
> RS> > >> can see. The gromacs wiki for bluegene explains how to get the kernel
> RS> > >> generated by mknb to work. Also on the mailinglist all the messages I
> RS> > >> found
> RS> > >> about BlueGene were about mknb and not abou the bluegene specifik
> RS> > >> kernel.
> RS> > >>
> RS> > >> Could it be that it is incomplete do one has to do somethng to generate
> RS> > >> these files?
> RS> > >>
> RS> > >> Roland
> RS> > >>
> RS> > >>
> RS> > >> On Wed, Jul 30, 2008 at 11:45 AM, Berk Hess
> RS> > >> <hessb at mpip-mainz.mpg.de>wrote:
> RS> > >>
> RS> > >>> Hi,
> RS> > >>>
> RS> > >>> Mathias Puetz from IBM wrote non-bonded kernels for BlueGene/L.
> RS> > >>> I don't know if these are also used for BlueGene/P.
> RS> > >>> If there are problems, he would be the person to contact.
> RS> > >>>
> RS> > >>> Berk.
> RS> > >>>
> RS> > >>>
> RS> > >>> Roland Schulz wrote:
> RS> > >>>
> RS> > >>>> Hi,
> RS> > >>>>
> RS> > >>>> has someone looked at the non-bonded kernels for BlueGene/P? Based on
> RS> > >>>> the
> RS> > >>>> assembler produced by the compiler, it does no SIMD, runs out of
> RS> > >>>> registers
> RS> > >>>> and does unnecessary rounding.
> RS> > >>>>
> RS> > >>>> Are there plans for a PPC450 assembler kernel?
> RS> > >>>>
> RS> > >>>> Roland
> RS> > >>>> This email was Anti Virus checked by Astaro Security Gateway.
> RS> > >>>> http://www.astaro.com
> RS> > >>>>
> RS> > ------------------------------------------------------------------------
> RS> > >>>>
> RS> > >>>> _______________________________________________
> RS> > >>>> gmx-developers mailing list
> RS> > >>>> gmx-developers at gromacs.org
> RS> > >>>> http://www.gromacs.org/mailman/listinfo/gmx-developers
> RS> > >>>> Please don't post (un)subscribe requests to the list. Use the www
> RS> > >>>> interface or send it to gmx-developers-request at gromacs.org.
> RS> > >>>>
> RS> > >>>
> RS> > >>> _______________________________________________
> RS> > >>> gmx-developers mailing list
> RS> > >>> gmx-developers at gromacs.org
> RS> > >>> http://www.gromacs.org/mailman/listinfo/gmx-developers
> RS> > >>> Please don't post (un)subscribe requests to the list. Use the www
> RS> > >>> interface or send it to gmx-developers-request at gromacs.org.
> RS> > >>>
> RS> > >>
> RS> > >>
> RS> > >
> RS> > > This email was Anti Virus checked by Astaro Security Gateway.
> RS> > > http://www.astaro.com
> RS> > >
> RS> > > _______________________________________________
> RS> > > gmx-developers mailing list
> RS> > > gmx-developers at gromacs.org
> RS> > > http://www.gromacs.org/mailman/listinfo/gmx-developers
> RS> > > Please don't post (un)subscribe requests to the list. Use the
> RS> > > www interface or send it to gmx-developers-request at gromacs.org.
> RS> >
> RS> >
> RS> > _______________________________________________
> RS> > gmx-developers mailing list
> RS> > gmx-developers at gromacs.org
> RS> > http://www.gromacs.org/mailman/listinfo/gmx-developers
> RS> > Please don't post (un)subscribe requests to the list. Use the
> RS> > www interface or send it to gmx-developers-request at gromacs.org.
> RS> >
> RS> 
>
>