[Pkg-openmpi-maintainers] Bug#545919: Bug#545919: openmpi: Please add support for BLCR

[Alan Woodland]

2009/9/10 Alan Woodland <awoodland at debian.org>:
> 2009/9/10 Manuel Prinz <manuel at debian.org>:
>> Hi Alan!
>>
>> Am Donnerstag, den 10.09.2009, 00:33 +0100 schrieb Alan Woodland:
>>> BLCR is now in main. It would be nice if openmpi were built using this
>>> where it is available.
>>>
>>> I've attached a short patch adding options to configure, and build-depends
>>
>> Thanks a lot for the patch!
>>
>> I did think about adding it but BLCR was not packaged yet. I will add it
>> in the next upload.
>>
>> Unfortunately, I do not have any possibility to test Open MPI with BLCR
>> at the moment. Can you confirm it works as expected?
>
> I checked it configured and built last night on my machine, but I've
> not had a chance to test any functionality properly yet.

Ok, I've had a bit more of a fiddle with it now. Several things have come up :)

Firstly a dependency on libcr0 is now generated automatically for
libopenmpi1.3, which is technically accurate, but not really strictly
needed. If libcr0 isn't installed mpirun still works exactly as before
anyway. The only time you notice that libcr0 isn't installed is when
you pass the "-am ft-enable-cr" flag to mpirun which causes it to try
and dlopen() the library that is really linked against libcr0.

Secondly it seems that we're going to need to install ompi-checkpoint
and ompi-restart in order for this to be useful. This could be done a
number of ways I think:
1) add to debian/openmpi-bin.install (Simple, gets the extra utilities
even on platforms blcr doesn't support though. Might not be a bad
thing anyway if they're useful for the 'self' checkpoint mechanism
anyway?)
2) add a conditional in install-arch and arrange for them to only be
installed on amd64,i386,ppc,armel where BLCR  builds and works
currently
3) add an extra package (e.g. openmpi-checkpoint, build it only on
supported arches. Suggest/recommend this package, probably from
openmpi-bin? This solution also fits nicely with what I said earlier
about not needing libcr0 unless you actually want to use blcr
checkpoints, because we could suppress the automatic dependency for
libcr0 and manually add it to this package instead, so avoiding
forcing all openmpi users to install libcr0 unless they actually care
about checkpoints)

What are your thoughts on this?

I'll try and convert the I've got into a repeatable, automated test
and then add it to this bug report in a bit too.

Alan