Bad Performance on Power4/AIX

[jschmidt]

Hi Axel,

thanks for your answer. I'm running interactively on one frame (I'm
only granted acces to this frame when running in interactive mode).
Since the sopt version is Ok, I don't expect that there is a problem
with the memory. If I run the parallel version with 1 processor, I
expect a similar (or at least not much different) behaviour than what
I find in serial. Is this correct? In my case, the parallel with 1
processor is about a factor of 100 slower than the serial version,
although I'm running on the same machine.

I'll keep trying, but this is very suprising to me, because I've never
had such a problem on an opteron cluster.

Regards,
Jochen


On 29 Okt., 21:56, Axel <akoh... at gmail.com> wrote:
> On Oct 29, 9:28 am, jschmidt <jochenc... at googlemail.com> wrote:
>
> > Hi,
>
> > I've just compiled CP2K on a system with Power4-processors
> > ("Regatta"), running under AIX, and there is a problem with the
> > parallel version. The sdbg and sopt is fine, but as soon as I switch
> > to pdbg or popt (using the arch-files from the cvs) the timings are
> > really bad. A system with one water molecule needs about 100 secs
>
> hi,
>
> have you checked that a) you are not running across multiple
> regatta frames (=> network overload) or b) you actually have
> requested and been granted access to multiple cpus to run in
> parallel on and c) have enough memory to not run into permanent
> swapping/paging. the latter would easily explain a 100x increase
> in runtime.
>
> cheers,
>     axel.
>
> for
>
> > initialization and again about 100 secs per MD step. Using the sopt-
> > version this is only about 1 sec. Here is the arch-file that I used:
>
> > PERL     = perl
> > CC       = cc
> > CPP      = /usr/ccs/lib/cpp
> > FC       = mpxlf95_r -qsuffix=f=f90
> > LD       = mpxlf95_r
> > AR       = ar -r
> > DFLAGS   = -D__AIX -D__ESSL -D__FFTSG -D__FFTESSL\
> >            -D__parallel -D__BLACS -D__SCALAPACK
> > CPPFLAGS = -C $(DFLAGS) -P
> > FCFLAGS  = -C -g -q32 -qarch=pwr4 -qflttrap=overflow:zerodivide:enable
> > \
> >            -qfullpath -qinitauto=FF -qnointlog -qnosave -qsigtrap -
> > qtune=pwr4
> > LDFLAGS  = $(FCFLAGS) -bnoquiet
> > LIBS     = -L /afs/ipp/@sys/lib -lscalapack\
> >            /afs/ipp/@sys/lib/blacsF77init_MPI-POWER4-0.a\
> >            /afs/ipp/@sys/lib/blacs_MPI-POWER4-0.a -lessl\
> >           -L/afs/rzg/@sys/lib -llapack-essl -lessl -lpessl -lmass -
> > lhmd
>
> > OBJECTS_ARCHITECTURE = machine_aix.o
>
> > Is anyone using the latest cvs-version on a such a system?
>
> > Thanks for your help!
> > Jochen