[Pkg-openmpi-maintainers] Open Mpi 1.2.6 and several dozen new Debian man page woes

[Jeff Squyres]

On May 21, 2008, at 8:53 AM, Manuel Prinz wrote:

> Another issue that I'd like to address: The only patch that we  
> currently
> apply to our OpenMPI is a rather trivial one and I'd like to ask you  
> for
> inclusion in OpenMPI. (Patch attached.) Because of some user request
> about libibverbs warnings being issued if InfiniBand is not available,
> we added a line to openmpi-mca-params.conf so it can be easily  
> disabled
> by the user without reading much documentation. We also documented  
> that
> in the Debian specific documentation. I do not know how often this
> questions arises but I think it won't hurt to include that two lines
> (one, actually) as it does not change the default behavior of OpenMPI.
> (That was important to us.) If you could include that patch, we could
> ship OpenMPI unmodified which I think is a Good Thing. :)


Ah - I had to think about why you wanted that for a minute: you're  
shipping libibverbs in Debian, but the majority of machines that  
people run on don't have IB or iWARP.  So OMPI builds support for  
OpenFabrics, but then complains at run-time because it can't find any  
hardware to use for the openib BTL.  Now I get it.  :-)

==> Clarification: is it libibverbs that complains?  I know that OMPI  
complains, but the text of the patch says that it's ibverbs that  
complains.

I can see your rationale for this for the v1.2 series.  I doubt any of  
the other vendors will object, but I'll ask them if they care if we  
put this in for a future v1.2 release (if it ever happens).

I'd like to propose a change for this behavior for the upcoming v1.3  
series: either have OMPI not complain if it can't find any OpenFabrics  
hardware or have a specific MCA parameter for turning of the  
complaint.  I need to check with the other vendors on which way they  
would prefer the default -- I'm of split mind on this (the  
implications caused by Debian [and others] shipping libibverbs kinda  
changes the game here; I need to mull on it a bit).

I'll start a thread on the OMPI devel list about both of these items;  
I expect little/no resistance for the v1.2 change.

What do you think?

-- 
Jeff Squyres
Cisco Systems