status

[Rob Ross]

Hi John,

(Sorry if people get two copies of this message; I have no idea who is and 
isn't on the PnetCDF list at this point!)

On Fri, 8 Aug 2003, John Tannahill wrote:

> Some of my e-mails to ANL have not been cced to
> parallel-netcdf at mcs.anl.gov, so I thought that I would send a short one
> just to update my status.  As of Monday of this week, all of my
> installation/test code problems had been taken care of.  I am running on
> NERSC's IBM SP, seaborg, and currently using 0.8.9.  By test code, I
> mean the C & Fortran parallel netCDF routines that I had been working
> on, not the test code in the parallel-netcdf tar file itself.  Many
> thanks to everyone who has helped to get me this far, especially to
> Jianwei for getting me over the last couple of hurdles.  I am happy to
> hear that further progress is being made on making the Fortran interface
> fully functional.

Glad to hear everything is in good shape!

> I am continuing to develop additional test codes and look at timings.  
> I am on vacation next week, but will continue this work when I return.

Excellent.  We should look at integrating some of these test codes into 
the PnetCDF package as appropriate.

> Additional things that would be helpful to me / Questions:
>    * At the end of the install, automatically run some test code(s)
>      to validate the installation.

This is a good idea.  However, it can be problematic on some systems -- it 
requires that the user correctly tell us how to run MPI programs, which 
can sometimes be a little tricky.  We should probably investigate it 
though and see if we can come up with a good solution.

>    * The primary machines that we currently run on are:
>         - IBM SP
>         - Compaq TC2000
>         - Intel clusters
>      The IBM SP seems to have been taken care of; not sure about the
>      other two?  How quickly will the library be updated as major new
>      machines appear?

Intel clusters should be taken care of, at least for some Fortran 
compilers (others might require some tweaks, we'll see).  We can address 
issues on the Compaq as bug reports are made :).

I don't know how to quantitatively respond to your "how quickly" question;  
however, all the things that we are doing (cleaning up warnings, fixing
casting, testing for mismatches in type sizes, making the configure, make,
and install more portable, etc.) lead us towards a package that should
mostly work on new systems as they are tried.

So I am optimistic about this.  The IBM and SX platforms are unique enough 
that in combination with the Intel platform I think we've seen a wide 
variety of unusual configurations already :).

>    * Any idea when the first official non-beta release will be?

Soon; some time after we get these last Fortran binding issues knocked 
out and maybe get some additional tests added to the library.

>    * I would eventually like to replace our current serial netCDF
>      functionality in our model with your parallel one.  I really
>      would like to get rid of the serial stuff altogether, but don't
>      know if it's too early in the game to be considering this?

As soon as any remaining Fortran binding issues are knocked out, I think 
this is a good idea.  I think we're very close.

>    * Your plans are definitely to fully support the parallel-netcdf
>      capability?

I'm not sure what you mean here.  Yes, we're going to support this code 
:).

>    * How much coordination is going on with Russ Rew and his efforts?
>      Hopefully, I won't have to do a complete re-implementation when
>      they complete their project?

I think that the netCDF4 effort is still relatively new.  We've been in
contact with them, and we've worked with the HDF5 group before too, and I
believe that this work will provide them with some good data and examples
to think about when designing their solution.

*Maybe* they will decide that this interface is what they would like to
use; that would be ideal for everyone I think if it makes sense.  The
potential of some "instant users" of their new code base would be very
tempting to me if I were in their position -- one of the hardest parts of
introducing new interfaces is getting those early adopters (thanks again
BTW!).

Russ et. al. -- if you guys have any suggestions as to things that you 
*know* you would want to see in your interface that are missing from what 
we have now, we'd very much like to talk about them!

>      Any hope of getting your library somehow integrated with UCAR's 
>      serial one?

We're going to do some clean up to allow both serial and parallel 
libraries to be used in the same code.  I doubt that the UCAR guys want to 
support our code, so probably we'll be a separate library.  I think that's 
fine.

Development here has been very quick, and I think that the UCAR group
deserves some thanks for having a library that served as such a good
starting point!

Regards,

Rob