[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: cfortran on Cray T3E (fwd)



> I thought I'd update you on what's going on with netCDF on our T3E.
>
> First, Burkhard Burow has offered to look into getting "cfortran"
> working on the T3E.  He asked for, and I sent him, the section in the
> Cray manuals regarding "Interlanguage Communications".  Pending his
> review of that documentation, his musings are:

That's great.

>   > Strange.
>   > I don't know the cray models, but I make the following guess.
>   > Is it possible that cfortran is not 'thread-safe' or something similar
>   > on the T3E?
>   > [I've never had anyone complain to me that cfortran.h is not thread-safe,
>   >  but I personally have not examined the ssue in cfortran.h.]
>   > A way to test this hypothesis would be to restrict the execution of your
>   > application to just one processor on the T3E.
>   >
>   > If this gives the same problem,
>   > then the problem can probably be found by running the
>   > cfortran tests on the T90 and on the T3E.
>   >
>   > I will send you a script tomorrow
>   > that can run the tests.
> Do these comments suggest anything to you?  Anyhow, as soon as I get
> anything from him, we'll give it a go.


It's probably simpler than this, but it is hard to tell.

It is probably true that the netcdf library and the tests are
not automatically magically multprocessor safe.

(I have run the test of the C code on a cray-t3d with the number
of processors set to 8, with success. This may be just dumb luck.
That machine has problems with the fortran to C interface as well.)

For testing, is there an environment variable I should set to restrict
the programs to a single processor?
MPP_NCPUS=1 ?


> Second, Ramesh Rao (one of local Cray reps) has been looking at your
> 3.3 stuff on the T3E was wondering if you could provide any more info
> about how and where things went wrong during the build/test.  That,
> combined with his knowledge of the T3E, may give us something more to
> go on.  If you'd like to contact him directly, his Email address
> is address@hidden.
>
> Third, I haven't been able to connect to the "cfortran" server
> (zebra.desy.de), so I haven't been able to download the "package" which
> you mentioned contained some tests.  Hopefully, the stuff Burkhard
> sends us will serve the same purpose.
>
> That's all for now...
> John

The /ftmp scrubber just got my make.out, so, I'll have to build again
to tell you. We are packaging up the netcdf-3 full release today, so,
I'll pull that over and give it a shot.

I'll let Ramesh know where it is and what I've tried.

-glenn