On Thu, 11 Apr 2013, Matthew Knepley wrote:

On Thu, Apr 11, 2013 at 9:48 AM, Satish Balay <[email protected]> wrote:

...

On Thu, 11 Apr 2013, Matthew Knepley wrote:

...

On Thu, Apr 11, 2013 at 9:38 AM, Satish Balay <[email protected]> wrote:

...

On Thu, 11 Apr 2013, Matthew Knepley wrote:

...

On Thu, Apr 11, 2013 at 9:17 AM, Satish Balay <[email protected]> wrote:

...

On Thu, 11 Apr 2013, Matthew Knepley wrote:

...

On Wed, Apr 10, 2013 at 7:53 PM, Satish Balay <[email protected]

...

wrote:

...

...

...

...

How do we handle these packages with complicated dependencies

[they

...

...

are complicated as petsc dependencies?]

In this report hdf5 was built with szip-2.1 and zlib-1.2.7. So how does petsc configure automatically detect this?

And I've recently used the following for hdf5/necddf5 on fusion [with the extra implicit depenceny of hdf5 on zlib, and netcdf5 on hdf5 from within configure]

'--with-hdf5-lib=-L/soft/hdf5/1.8.6-parallel/lib -lhdf5_hl -lhdf5 -lgpfs',

'--with-netcdf-lib=-L/soft/netcdf/4.1.1-parallel/lib -lnetcdff -lnetcdf -L/usr/kerberos/lib64 -lcurl -ldl -lgssapi_krb5 -lkrb5 -lk5crypto -lcom_err -lidn -lssl -lcrypto -lz',

Yeah - things work if the user knows the dependencies and the link command for those dependencies - and specify it to petsc configure as above. But I'm not sure how to autodetect this.

Also currently -lz is handled in package.py with 'self.needsCompression' similar to 'self.needsMath' with the detection of -lz in libraires.py:checkCompression() - but one can't specify a --with-zlib-lib option this way. I guess this part can be fixed by migrating it a standalone package z.py. [And somehow handle the optional part of this dependency for hdf5]

We already have a mechanism for this. Lots of packages depend on other packages. This is just screwed up in the case of libz because someone (maybe me) did not want to write an entire package file for it, and instead copped out with the needsCompression.

Currently we handle 'mandatory' dependencies properly - and optional

...

depencencies for --download-pakcage somewhat - but not for --with-package-dir

For hdf5 - handling just zlib.py is not sufficient. Do we add szip.py, zlib.py, gpfs.py [and perhaps more] for hdf5? And somehow use these as optionaly when testing for user provided --with-hdf5? Which package currently implements this usage [to check for this optionaly dependency - and give the correct error]?

i.e In this case - when the user specified --with-hdf5-dir option. We had to print an error.: * please provide --with-zlib-dir and --with-szip-dir options. [For a different user - with the same usage - there would be no error]

For netcdf above one would have kerbros.py, curl.py and perhaps more?

We have had this same discussion a million times, and we always eventually get to the same answer and forget. Do we need our own error message for petsc-dev?

We cannot possibly guess everything that can go wrong. What we need is a system where the user can specify what he actually wants and we check it. So, zlib and everything else are optional dependencies for hdf5. If something is not specified, it fails, and the user goes back and specifies the right thing.

Sure - but specify how? specify everything with --with-hdf5-lib,--with-hdf5-include options?

Or somehow he knows that he has to specify --with-zlib-include/-with-zlib-lib options?

This. It would be nice if HDF5 would tell us about its dependencies, but if it does not, then the user has to, and this is our system (packages).

So we would start adding zlib.py, szip.py and manymore [and support --download-zlib --download-szip]. And if user gets the error:

--with-hdf5-dir=/foo/bar does not work.

So he/she send us configure.log

Then we look at it and respond saying: you need to specify --with-zlib-dir --with-szip-dir options aswell.

Does this look right?

Yes, that is what I think makes sense.

We can always try to add another layer that 'greps' for libz in the hdf5 error message and guesses that the user forgot it, but hopefully before we get to it, a sane storage library appears, or HDF5 gets its act together.

In this case - why bother with libz.py and szip.py? Why not tell the user to specify hdf5 completely with --with-hdf5-include and --with-hdf5-lib options? Just printing the link errors on the screen might be sufficient to clue in the user. Currently its hidden inside configure. [And its not clear which of the 10 tests configure did for this case should be printed on the screen as the appropriate message] Satish