Bug#904020: libpetsc-real3.9-dev: dependency on libopenmpi-dev not strict enough

[Drew Parsons]

On Wed, 2018-07-18 at 13:52 +0200, Paul Gevers wrote:
> Package: libpetsc-real3.9-dev
> Version: 3.9.3+dfsg1-2
> 
> Dear Drew,
> 
> On 17-07-18 13:20, Graham Inggs wrote:
> > Hi Drew
> > 
> > > On Tue, 17 Jul 2018, Drew Parsons wrote:
> > > > The configuration of ci.debian.org is not consistent.   A
> > > > failing test
> > > > of deal.ii is marked on
> > > > https://ci.debian.net/packages/d/deal.ii/testing/amd64/
> > > > as triggered by petsc/3.9.3+dfsg1-2, but the test log shows
> > > > that in
> > > > fact libpetsc-real3.8-dev is used for the test
> > > > (i.e. petsc/3.8.4+dfsg1-2+b2).  So unsurprisingly the test
> > > > fails, which
> > > > disrupts migration of the new petsc 3.9 (throwing it out to 10
> > > > days
> > > > instead of 5). The failure itself occurs since the different
> > > > petsc were
> > > > built against different openmpi.
> > > From the times in the logs, it appears that the test was run
> > > before
> > 
> > deal.ii's binNMU happened [1].
> > So the delay is only temporary.
> > 
> > As soon as the test is successful, the migration delay will
> > disappear.
> > 
> > I've requested a retry for that test, so it should happen soon.
> 
> Although Graham was right with his timing regarding binNMU of
> deal.ii,
> there is more going on. I really don't want to offend you, but I
> wonder
> if you read the actual error message carefully? (It took Graham and
> me
> quite some time to figure this out and to understand it).
> 
> '''
> PETSc was configured with one OpenMPI mpi.h version but now appears
> to
> be compiling using a different OpenMPI mpi.h version
> '''
> 
> This message comes from petsc, so, whatever header file got installed
> was blessed by petsc.
> 
> Apart from petsc/3.8.4+dfsg1-2+b2 also petsc/3.9.3+dfsg1-2 packages
> were
> installed (so both of them), so I think the trigger did it's job.
> 
> I believe that the current failure is really pointing at a bug in the
> dependencies of libpetsc-real3.9-dev which probably need tighter
> restrictions on libopenmpi-dev looking at the message. If
> libpetsc-real3.9-dev would require libopenmpi-dev from unstable, I
> believe everything would be all right. I'll trigger a test like the
> one
> that currently fails, but with libopenmpi-dev from unstable to prove
> my
> hypothesis.


Thanks for the extra analysis Paul.  You're right, I picked up on the
appearance of petsc 3.8 in the failing tests but actually petsc 3.9 is
installed there too (which is a bit weird).

In regards to the petsc/openmpi dependency, we've already addressed
that. It used to depend on the subminor release version, we've relaxed
it to match major.minor 
https://bitbucket.org/petsc/petsc/commits/ca70f86ee9db8e69523e0e69f12289c6cab9b4cb . 

That relationship is expressed in the libpetsc-real3.9-dev dependency,
libopenmpi-dev (>= 3.1), and that's for both libpetsc-real3.8-dev and
libpetsc-real3.9-dev at the moment.


OpenMPI was libopenmpi-dev 3.1.0-7 in the failing deal.ii tests, and 
libopenmpi-dev 3.1.1.real-3 in the successful tests. Either way that's
still openmpi 3.1, so is supposed to be compatible.  

So looks like you're right, there may be a PETSc bug when it tests for
mpi compatibility. openmpi 3.1.0 and 3.1.1. are both supposed to be
fine, they should not be triggering the incompatibility warning.

Drew