parallel netcdf - first experiences

Robert Latham robl at mcs.anl.gov
Mon Jul 28 10:57:32 CDT 2003


On Mon, Jul 28, 2003 at 04:44:54PM +0200, rene.redler wrote:
> 
> Hello,
> 
>    I recently (today) subscribed to parallel-netCDF mailing list, so I 
> apologies in advance for taking your time if I rise any topic that has 
> already been discussed.

we've got some people actively testing our code on AIX, but you are
the first NEC user.  No need to worry aobut duplicate issues.

> I started with the latest version 0.8.6 and tried to install the library on 
> a NEC SX6. Doing this on a SX6 itself would have resulted in some changes 
> of the Makefile.in due to different behavior of make (-C is not supported 
> there and install, and some other odd things.) 

The -C in the Makeifles is my fault and I've gotten rid of it (but not
in time for 0.8.6 ).  Could you tell me more about the 'other odd
things'?  It may drive me crazy, but we're going to try to get the
Makefiles to work with as many make implementations as possible.

> I then started working with 
> the cross compiler which was not possible with release 0.8.4 which I 
> tested last week due to the fact that the previous release needed to run 
> executables during the configure (or make) phase.
> 
> I needed to define the following environment variables 
> 
> setenv MPIF77 sxmpif90
> setenv MPICC sxmpic++
> setenv AR sxar
> setenv RANLIB :
> 
> Running configure and make worked fine.

fantastic.  glad to hear it. 

> When trying to create the test cases I recognized that mpinetcdf.c tries 
> to call  MPI_Type_create_vector which I guess is not part of the MPI2 
> standard. Are there any plans to to add this function to the MPI2 in 
> future? I also could not find this function in the mpich-1.2.5 release. 
> Comments are welcome.

I'm going to defer to someone more familiar with the differnce between
MPI-1 and MPI-2.. We *should* be using the MPI_VERSION macro to call the
right type function, like this:

#if (MPI_VERSION < 2)
            MPI_Type_vector(subcount[0], varp->xsz, ncp->recsize,
                            MPI_BYTE, &filetype);
#else       
            MPI_Type_create_vector(subcount[0], varp->xsz, ncp->recsize,
                                    MPI_BYTE, &filetype);
#endif      

Does NEC's mpi provide MPI_VERSION?  What value does it have?

> Nevertheless trying to test_read the netCDF files failed, and the ncmpi
> routines returned and invalid id or dimension. Did anyone encounter a 
> similar behavior on any other architecture?

On ia32 linux, these tests all pass.  Could you send which tests you
ran and how you ran them?

Thanks for helping us shake out problems in our software!

==rob

-- 
Rob Latham
Mathematics and Computer Science Division    A215 0178 EA2D B059 8CDF
Argonne National Labs, IL USA                B29D F333 664A 4280 315B



More information about the parallel-netcdf mailing list