Parallel-NetCDF 1.0.0 pre-release

[Robert Latham]

On Wed, Apr 13, 2005 at 10:03:24AM -0500, Robert Latham wrote:
> Hi Rene.  Thanks for the report about the issues.  
> 
> On Wed, Apr 13, 2005 at 10:41:50AM +0200, Rene Redler wrote:
> > When trying to run this we encounter some (minor) problems.
> > 1.) nfmpi_inq_libvers is missing in the Fortran interface
> 
> We have a script that generate our Fortran interface from the C
> interface.  Our script must have missed ncmpi_inq_libvers.  It is
> the only ncmpi_* routine that returns char * instead of int.  

Ah, turns out we have to write our own wrappers for routines that
return char * (strerror and inq_libvers).  I've now done this for
inq_libvers .

> > 2.) nfmpi_set_fill is not implemented
> 
> You caught us there.  In a parallel implementation of netcdf, it is
> very difficult to implement anything other than NC_NOFILL efficently.
> We should probably implement the routine for parity with serial
> netcdf, and return with an error if anything besides NC_NOFILL is
> unsupported

We now do this.  We have ncmpi_set_fill (and nfmpi_set_fill) which
return NC_INVALID if anything besided NC_NOFILL is passed in.  

> > 3.) nfmpi_def_dim required the input length to be declared as
> >     integer(kind=MPI_OFFSET_KIND)
> 
> This is a relatively recent change so that we could better handle
> large files.  We need to better document our approach for handling 64
> bit parameters

Well, I've made a start at documenting this behavior in a top-level
README.

> > 4.) NF_UNLIMITED, often used as input parameter to nfmpi_def_dim
> >     is only declared as default integer, typically different from
> >     MPI_OFFSET_KIND. Using NF_UNLIMITED causes nfmpi_def_dim
> >     to return an error.
> 
> Thanks for catching that.

We haven't fixed this one yet to my knowledge.

> > 5.) Our file names contain the time in a format hh:mm:ss . These
> >     colons are misinterpreted by pnetcdf. Apparently you use the
> >     colon to look after information concerning the file system.
> 
> We pass the file name down to ROMIO, which, if one is given,  uses a
> prefix to determine the file system type.  Typically that would be
> something like "pvfs2:testfile" or "nfs:/path/to/data".  One way
> around that is to provide the file system prefix in addition to your
> timestamp.  So if your file is "10:32:43.data", using
> "<FSTYPE>:10:32:43.data" should make everything happy.  Again, this
> needs to be better documented.  

Added this to that README i mentioned above.

> > Some hints for the installation of Parallel-NetCDF 1.0.0 under
> > Suse Linux.
> > 
> > yacc and lex are not included in the standard installation.
> > To have yacc available the bison package is required. To have
> > lex available the flex package must be installed.
> 
> We'll add this to the documentation

Added this too.

Thanks again for reporting these issues.  
==rob

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