parallel-netcdf 0.9.4 released

[Robert Latham]

On Wed, Jul 21, 2004 at 03:33:08PM +0200, Joachim Worringen wrote:
> Problems I found (on SX-6):
> - configure probing for the sizeof(MPI_Offset): linking is done with $CC 
> instead of $MPICC. This does not work on SX (don't ask me why; I know it 
> should as no MPI function is referenced). This works alright on IA32. Anyway, 
> in configure.in, line 245, it should read:
> if $MPICC -c conftest.c && $MPICC -o conftest conftest.o ; then

Ok, that change makes sense.  Jianwei, is there a platform (maybe the
IBM? ) where we should use CC instead of MPICC to link?

> - In the same context, the declaration of 'a' in line 236 can be omitted (our 
> compiler gives ugly warning for unused variables. 

fixed this to use 'sizeof(a)'

> Other notes:
> - I used the settings as described in README.SX. It seems that the additional 
> iteration (calling make again in src/libf after the initial make) is no 
> longer necessary, at least not on the system I tested on which was upgraded 
> to Super-UX 14.1 from 12.1 in the meantime. As the additional invocation 
> causes no harm, we should better leave the README as is.

good to know.  I'll make a note that the extra step might not be
necessary, but doesn't hurt things.

> - Although crossconfiguring does not work (checking for working ftruncate is 
> not possible), configuring on the target host and then do a cross-compile 
> works fine when setting up necessary aliases on the configure host (cc -> 
> sxcc etc.) or vice versa on the compile host (sxcc -> cc). Might be changed 
> in README.SX.

I've updates README.SX

> Test results for SX: 
> _ nf_test: the int1 and int2 tests fail with "value read not that expected" 
> for get operations and "Range error: No error" for put operations. I'll have 
> to look into this more closely (if it has to do with the fact that integer*1 
> is not supported on SX)

Most likely.  We *shouldn't* be running int1 tests if there is no int1
support:

#if defined(NF_INT1_T)
        external        test_nfmpi_get_var_int1
#endif
...
#if defined(NF_INT1_T)
        call test('nfmpi_get_var1_int1', test_nfmpi_get_var1_int1)
#endif

so this might not be working like w

> - nc_test: passes with the expected two errors 

great!

> - nc_test2: fails with a number of errors like this (also fails like this on 
> IA32):
>         FAILURE at line 637 of util.c: ncmpi_put_vara_text_all: Edge+start 
> exceeds dimension bound
> I assume this doesn't really matter as only nc_test and nf_test are included 
> in the 'testing' target of test/Makefile?

nc_test2 still uses ptrdiff_t instead of MPI_Offset.  I'll add it to
the 'testing' target to make sure I don't make such a mistake again.

> Final note: is the introduction of a ncmpi_delete() function (wrapper for 
> MPI_File_delete) still planned? 

It should be in 0.9.4:

int
ncmpi_delete(char *filename, MPI_Info info)

I'm not sure if we'll ever need the 'info' parameter, but it's there
just in case.

Thanks for trying out the new release!

==rob

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