Hi Arman,
Thanks a lot! This is highly useful information.
Incidentally, together with Mateusz Ruzkowski we came across another
problem on Altix 4700, just two days ago. In this case it was NASA's
Columbia machine.
The code always crashed in the initialization of pm_nonperiodic.c. We
traced it to a call of the erfc() function. For an argument larger than
about ~30.0, this function would not return zero as it should (with
underflow flagged), but instead it would simply crash the application.
Replacing erfc() with a call of the GSL math library, i.e.
gsl_sf_erfc(), solved this.
Potentially, the above has some relation to your problem, but I'm not sure.
Volker
Arman Khalatyan wrote:
> Hello everyone who uses Gadget2 on Altix 4700 machines. I would like to
> report the bug in Intel compiler 9.0> at LRZ (Leibniz Reichnung Zentrum).
> I was fighting with question: Why Haloes produced on Altix machine at LRZ
> and haloes on other machined are differ in properties form 10-20%.
> Finally I found that the PM module (pm_periodic.c/pm_nonperiodic.c) does not
> produces correct "P[i].GravPM[dim] = acc_dim;" it seems is truncated in
> memory from double to float or something else. In single PM step the values
> for P[i].GravPM[dim] are differ in order of 1e-11 up to 1e-5. If we add
> exponential growth of errors in cosmological simulations we will end up with
> unacceptable results.
>
> I was testing Gadget-2 over several supercomputers/clusters: at AIP(
> P4HT3.6Gh, AMD), JUMP(IBM-p690+), Marenostrum(IBM,p970dualcore), UAM(SGI
> ALTIX 3000), LRZ(SGI ALTIX 4700,CPU type is ITANIUM2 Madison 9M).
>
> There results where exactly same except on ITANIUM-2.
> Important thing to point that FFTW alone produces correct results.
> Gadget-2 without PM also correct.
>
> To solve the problem .
> Only combination of this 4 options can produce correct code:
> icc -no-gcc -ansi -ansi_alias -ipo
> NOTE:
> -mp1 or -mp does not effect anything,these options force compiler do not do
> any precision optimizations.
>
> I was testing also all possible compilers (intel8.1 to intel10.0b)at LRZ
> here is tests:
>
> module load gsl
> module load fftw/mpi
>
> icc 8.1 + gcc 3.3.4 running ok (probably there is no aggressive
> optimizations for ITANIUM-2 cpus)
>
> icc 9.0 and above + gcc 4.01 or 4.1.0 does not running correctly with PM
> module in Gagdet2 in file pm_periodic.c/pm_nonperiodic.c the P[i].GravPM
> wrong
>
> solution: -O3 -no-gcc -ansi -ansi_alias -ipo
> or for debugging:
> -O0 -g -traceback (-O0 is important to switch off all possible
> optimizations)
> I found good to know.
>
> Arman.
>
> +++++++++++++++++++++++++++++++++
> Arman Khalatyan
> Address:
> Astrophysical Institute Potsdam,
> An der Sternwarte 16,
> D-14482 Potsdam, GERMANY
> Phones: (+049) 0331 7499517(work),
> (+049) 030 26558524(priv).
> +++++++++++++++++++++++++++++++++
>
>
>
> ------------------------------------------------------------------------
>
>
> -----------------------------------------------------------
>
> If you wish to unsubscribe from this mailing, send mail to
> minimalist_at_MPA-Garching.MPG.de with a subject of: unsubscribe gadget-list
> A web-archive of this mailing list is available here:
> http://www.mpa-garching.mpg.de/gadget/gadget-list
Received on 2007-02-22 12:02:48