[MITgcm-devel] [altMITgcm/MITgcm66h] Bugfix/scratch files (#11)

Jean-Michel Campin jmc at mit.edu
Fri Aug 4 11:01:48 EDT 2017 

Hi Martin,

The changes you made seems complicated:
This part: line 155-160
      IF ( .NOT.doReport ) THEN
C     called from eeboot_minimal.F before myProcId is set, so we have to
C     use scratch files and keep our fingers crossed
       OPEN(UNIT=scrUnit1,STATUS='SCRATCH')
       OPEN(UNIT=scrUnit2,STATUS='SCRATCH')
      ELSE
is not needed + it relies on opening unit with STATUS='SCRATCH' that we would
like to avoid when USE_FORTRAN_SCRATCH_FILES is undef (and with this 
IF ( .NOT.doReport ) THEN .. the procId argument that I added few days ago is 
of no use).

But I would not change anything regarding the SINGLE_DISK_IO block (there is a 
stop there, for good reasons, and it already open scrUnit 1 & 2
as real file, i.e, STATUS='UNKNOWN').

Cheers,
Jean-Michel

On Fri, Aug 04, 2017 at 04:03:38PM +0200, Martin Losch wrote:
> Hi Jean-Michel,
> I checked in a new eeset_parms.F While I think that this version will not break any tests, it is probably not very good in terms of some special cases (e.g. it will break SINGLE_DISK_IO, because I forgot add a proper flag for the declaration of scratchFile1 and 2).
> It???s Friday afternoon and my brain seems to be in weekend mode already, that???s why I am reluctant to check in anything without consulting with you. Here???s what I think I should do:
> (1) remove the SINGLE_DISK_IO block, because now you always pass something meaningfull in ???procID" to eeboot_minimal.
> (2) replace it with a
> #ifdef SINGLE_DISK_IO
>        IF ( procID .EQ. 0 ) THEN
> #else
>        IF ( .TRUE. ) THEN
> #endif
>        ELSE
> ???
>        ENDIF
> 
> at the beginning of the default (if !defined USE_FORTRAN_SCRATCH_FILES) block.
> I think that should work, what do you think?
> 
> Martin
> 
> > On 3. Aug 2017, at 15:10, Jean-Michel Campin <jmc at mit.edu> wrote:
> > 
> > Hi Martin,
> > 
> > Yes, last changes are good, and you can proceed with next step
> > when you want.
> > 
> > Cheers,
> > Jean-Michel
> > 
> > On Thu, Aug 03, 2017 at 12:54:56PM +0200, Martin Losch wrote:
> >> Hi Jean-Michel,
> >> 
> >> I know you have been busy with other stuff, but it does not look like there are any problems with my changes to eeset_parms.F
> >> Should I now do the second step and change the default as suggested (just to eeset_parms.F, if it works, I can add the stuff to all namelists)?
> >> 
> >> Martin
> >> 
> >>> On 28. Jul 2017, at 14:57, Martin Losch <Martin.Losch at awi.de> wrote:
> >>> 
> >>> OK,then Iet???s wait until Monday,
> >>> 
> >>> Martin
> >>> 
> >>>> On 28. Jul 2017, at 14:50, Jean-Michel Campin <jmc at mit.edu> wrote:
> >>>> 
> >>>> Hi Martin,
> >>>> 
> >>>> These experiments were already failing before, in the same way,
> >>>> so I am not worried too much. 
> >>>> Now some tests are not running everyday (I alternate -fast and -devel), 
> >>>> so it might be good to wait at least an other day (to pass more -devel tests).
> >>>> 
> >>>> Cheers,
> >>>> Jean-Michel
> >>>> 
> >>>> On Fri, Jul 28, 2017 at 09:58:35AM +0200, Martin Losch wrote:
> >>>>> Hi Jean-Michel,
> >>>>> 
> >>>>> it looks like some forward tests actually do fail since my change to eeset_parms.F, e.g. here:
> >>>>> svante linux_amd64_pgf77+mth.fast ( the corresponding linux_amd64_pgf77+mth.dvlp looks OK)
> >>>>> 
> >>>>> Y Y Y N .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . N/O   aim.5l_cs
> >>>>> Y Y Y N .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . N/O   aim.5l_cs.thSI
> >>>>> Y Y Y N .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . N/O   aim.5l_Equatorial_Channel
> >>>>> Y Y Y N .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . N/O   aim.5l_LatLon
> >>>>> 
> >>>>> Y Y N N .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . N/O   hs94.cs-32x32x5
> >>>>> Y Y N N .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . N/O   hs94.cs-32x32x5.impIGW
> >>>>> 
> >>>>> Y Y N N .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . N/O   short_surf_wave
> >>>>> 
> >>>>> The comile time error (hs94.cs-32x32x5, short_surf_wave) does not look related to me:
> >>>>> 
> >>>>> pgf77 -byteswapio -Ktrap=fp -mp -tp k8-64 -pc=64 -O2 -Mvect=sse  -c ini_dynvars.f
> >>>>> PGFTN-F-0007-Subprogram too large to compile at this optimization level  (ini_dynvars.f)
> >>>>> PGFTN/x86-64 Linux 16.9-0: compilation aborted
> >>>>> Makefile:1653: recipe for target 'ini_dynvars.o' failed
> >>>>> make[1]: *** [ini_dynvars.o] Error 2
> >>>>> make[1]: Leaving directory '/net/fs09/d0/jm_c/test_svante/MITgcm_pgiMth/verification/hs94.cs-32x32x5/build'
> >>>>> Makefile:1561: recipe for target 'fwd_exe_target' failed
> >>>>> make: *** [fwd_exe_target] Error 2
> >>>>> 
> >>>>> but the aim.* experiments loose their threads. 
> >>>>>>>> Error: _mp_pcpu_reset: lost thread
> >>>>> Can that be related to closing some files?
> >>>>> 
> >>>>> Martin
> >>>>> 
> >>>>>> On 27. Jul 2017, at 00:22, Jean-Michel Campin <jmc at mit.edu> wrote:
> >>>>>> 
> >>>>>> Hi Martin,
> >>>>>> 
> >>>>>> two things:
> >>>>>> 1) I've checked that MPI_COMM_RANK is not blocking (can be called
> >>>>>> by only a subset of procs) so I added this call in the OASIS block
> >>>>>> and add argument "procId" to EESET_PARMS as suggested before.
> >>>>>> This should make your coming set of changes simpler.
> >>>>>> 2) the set of changes you propose seems good to me. And for now,
> >>>>>> I would set this USE_FORTRAN_SCRATCH_FILES in CPP_EEOPTIONS.h 
> >>>>>> and not worry about genmake_local.
> >>>>>> 
> >>>>>> Cheers,
> >>>>>> Jean-Michel
> >>>>>> 
> >>>>>> On Wed, Jul 26, 2017 at 10:16:45AM +0200, Martin Losch wrote:
> >>>>>>> Hi Jean-Michel,
> >>>>>>> 
> >>>>>>> I suggest to test this now as you say, i.e. check in an eeset_parms.F where only the appropriate close statements are ammended with STATUS=???DELETE??? (which in my opinion should always work, since this option is F77 standard, but you never know ???), but also have (at least) one testreport-verification-experiment use the USE_FORTRAN_SCRATCH_FILES flag, so that it is always tested (that???s a bit annoying, since it would be the only experiment with it???s own CPP_EEOPTIONS.h file, or can this be put into some genmake_local?)
> >>>>>>> 
> >>>>>>> Martin
> >>>>>>> 
> >>>>>>>> On 25. Jul 2017, at 18:17, Jean-Michel Campin <jmc at mit.edu> wrote:
> >>>>>>>> 
> >>>>>>>> An other thing:
> >>>>>>>> Are we 100% sure that closing a scratch unit file with status "delete" 
> >>>>>>>> is completly standard on all platforms & compilers ? If not, we could
> >>>>>>>> test just this independently (i.e., check-in and see how daily test run). 
> >>>>>>>> The reason is that when someone chose to use USE_FORTRAN_SCRATCH_FILES,
> >>>>>>>> (which is not going to be the default and therefore not tested) we need to be 
> >>>>>>>> sure that the close instruction is OK.
> >>>>>>> 
> >>>>>>> 
> >>>>>>> _______________________________________________
> >>>>>>> MITgcm-devel mailing list
> >>>>>>> MITgcm-devel at mitgcm.org
> >>>>>>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
> >>>>>> 
> >>>>>> _______________________________________________
> >>>>>> MITgcm-devel mailing list
> >>>>>> MITgcm-devel at mitgcm.org
> >>>>>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
> >>>>> 
> >>>>> 
> >>>>> _______________________________________________
> >>>>> MITgcm-devel mailing list
> >>>>> MITgcm-devel at mitgcm.org
> >>>>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
> >>>> 
> >>>> _______________________________________________
> >>>> MITgcm-devel mailing list
> >>>> MITgcm-devel at mitgcm.org
> >>>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
> >>> 
> >>> 
> >>> _______________________________________________
> >>> MITgcm-devel mailing list
> >>> MITgcm-devel at mitgcm.org
> >>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
> >> 
> >> _______________________________________________
> >> MITgcm-devel mailing list
> >> MITgcm-devel at mitgcm.org
> >> http://mailman.mitgcm.org/mailman/listinfo/mitgcm-devel
> > _______________________________________________
> > MITgcm-devel mailing list
> > MITgcm-devel at mitgcm.org
> > http://mailman.mitgcm.org/mailman/listinfo/mitgcm-devel
> 
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mailman.mitgcm.org/mailman/listinfo/mitgcm-devel

More information about the MITgcm-devel mailing list