[mpas-developers] Proposed Changes to trunk

[Doug Jacobsen]

Hi Michael,

I ran it with openmpi, as well as whatever mpi implementation is installed
on the clusters here, probably openmpi as well though. The 0.00 for
efficiency is just what it gets initialized for, so there aren't any NaN's
printed out.

It might be better to just move the two timers that are called prior to
mpi_init to be after the call to mpas_framework_init. Since the call to
mpi_init is inside mpas_dmpar. Otherwise we would have to add more compile
options. I think it should be the default timer, since system_clock has
other issues like only being 32bit.

It probably makes sense to move these timers too, since the framework
finalize is outside of the timers as well. If you want to give that a try,
and let me know if that works or if that's a suitable solution I can add
that as part of my commit.

Doug


On Wed, Jan 4, 2012 at 5:49 PM, Michael Duda <duda at ucar.edu> wrote:

> Hi, Doug.
>
> I think that adding additional directories for documentation sounds fine.
>
> For the new timer code, when I run the non-hydrostatic atmosphere model,
> I'm getting only 0.00 for efficiency for all timers; is there something
> we need to do to enable this extra information?
>
> I don't think it's related to the new timing code, since I recall
> getting these messages previously on our IBM, but it appears that
> certain MPI implementations are unhappy with calls to MPI_Wtime that
> occur before a call to MPI_Init; I'm seeing the following fatal errors:
>
> ERROR: 0032-150 MPI is not initialized  in MPI_Wtime, task 7
> ERROR: 0032-150 MPI is not initialized  in MPI_Wtime, task 5
> ERROR: 0032-150 MPI is not initialized  in MPI_Wtime, task 0
> ERROR: 0032-150 MPI is not initialized  in MPI_Wtime, task 1
> ERROR: 0032-150 MPI is not initialized  in MPI_Wtime, task 2
> ERROR: 0032-150 MPI is not initialized  in MPI_Wtime, task 6
> ERROR: 0032-150 MPI is not initialized  in MPI_Wtime, task 3
> ERROR: 0032-150 MPI is not initialized  in MPI_Wtime, task 4
>
> Until now, I'd just been moving the first two calls to mpas_timer_start()
> after the call to mpas_framework_init(), but perhaps you'd know whether
> there is some fix that can be made in the timing module. Of course, I'd
> completely accept that either the IBM MPI implementation is too fussy,
> or that we just need to be careful about starting timers until after MPI
> has been properly initialized. Would something like changing "#elif _MPI"
> to "#elif (_MPI && USE_WTIME)" in the timing module be acceptable?
>
> Thanks!
> Michael
>
>
> On Wed, Jan 04, 2012 at 02:13:04PM -0700, Doug Jacobsen wrote:
> > Hello Everyone,
> >
> > I have two changes to the trunk that I would like to propose.
> >
> > First, we would like to add a directory to store design documents. The
> > proposed path is trunk/documents/[core], where each core would have a
> > directory and we could add a shared directory for global design documents
> > (similar to the multiple blocks/pio design document). The structure
> inside
> > of these directories could be decided on by the core developers, or
> people
> > working on the design documents for that directory.
> >
> > Second, since I never got around to it last year, I would like to propose
> > the commit of my previously mentioned timers. Essentially the only added
> > features are that timers are synced across processors, and a load
> > efficiency is output in addition to the other timer information.
> >
> > I have attached a copy of the timers for anyone to explore, so any
> comments
> > would be appreciated. If I don't hear anything on these I will commit
> them
> > at some time tomorrow.
> >
> > Thanks,
> > Doug
>
>
> > _______________________________________________
> > mpas-developers mailing list
> > mpas-developers at mailman.ucar.edu
> > http://mailman.ucar.edu/mailman/listinfo/mpas-developers
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.ucar.edu/pipermail/mpas-developers/attachments/20120104/9df42a85/attachment.html