[ewg] EWG/OFED meeting minutes for Jan 25 2010

[Jeff Squyres]

On Jan 25, 2010, at 4:54 PM, Roland Dreier (rdreier) wrote:

> Please do *NOT* add ummunotify to OFED.  The upstream kernel has
> explicitly rejected it (at least for the moment).  It adds a huge
> maintenance and compatibility headache if OFED ships ummunotify and then
> upstream takes an incompatible version.

I think the rationale here is that we're now several years into the problem with no upstream solution -- can't we have *something* that works, even if it's a stopgap until upstream finally gets a better solution?

We continue to have problems and issues with injecting MPI-embedded ptmalloc2 into applications.  ummunotify, even if it's a temporary solution, would be a welcome relief.

> Effort would be much better spent getting ummunotify or something like
> it into shape where upstream will merge it.

That's hard to disagree with.  But it just seems like there is an unbounded amount of time to get upstream to have a final solution.  ummunotify works fine now, even if it's an OFED-specific solution.  

...but then again, would it be horrible if the distros picked up ummunotify, too?  Any upstream solution is likely to be entirely different (i.e., a different kernel module name), so it should be relatively easy for applications to determine at run time which, if either, they want to use.  Once there's a final solution upstream, applications can fade away their support for ummunotify.

-- 
Jeff Squyres
jsquyres at cisco.com