[ofa-general] RE: [ewg] RFC: Do we wish to take MPI out of OFED?

[Jeff Squyres]

On Jun 15, 2009, at 9:52 AM, Pavel Shamis (Pasha) wrote:

> > We think that the simple & clear answer is: "take the MPI packages  
> out
> > of OFED"
>
> It is not so simple and clear for me after all this discussion on  
> the thread. Some OFED member want to remove MPIs and some strongly  
> against it (the same correct for OFED user community too).
>

It speaks volumes to me that Red Hat has told us -- repeatedly -- that  
this is bad software engineering, not what the rest of the Linux  
community does, and has specifically asked us not to do it.  But yet  
we still do.  The main objection so far seems to have been "but we  
want to include MPI" (the co-development arguments do not make sense  
for OFED *distribution*).

> As I mentioned previously I sure that this step will push users from  
> OFED distributions towards vendors packages, like it was before  
> first OFED release.
>

Perhaps it's time that we start discussion of how the every vendor has  
incompatible/different OFED distributions is far worse than including  
MPI.  As I indicated earlier, we're back to the bad old days of  
mVAPI.  This compatibility stuff is confusing enough for those of us  
who are in the industry and working with this stuff every day -- can  
you imagine being a customer just trying to get a bug fix?

> What will you answer on user's question: "How can I install MPI with
> OFED support ?"  Will you send user to read 10 pages of MPI  
> Installation FAQ ? :-)
>


Many customers are in the habit of downloading/installing MPI  
manually, anyway, since they want to be on the bleeding edge of  
releases (as has been stated on this list).  For customers who just  
want an MPI that works, a simple installation guide, and/or getting  
the MPI's to distribute binaries, and/or having customers download  
MPI's from the distros are all viable solutions.  In short: these are  
all solvable problems.

More specifically: wasn't OFED created to solve the problem of mVAPI ! 
= mVAPI != mVAPI?  If so, it has failed.  There are two obvious  
solutions (and probably others):

1. Test together and only use the distros to distribute new versions, or
2. Use better packaging such that vendors can *contribute* their own  
technology without effectively making (potentially incompatible) forks  
of OFED.

-- 
Jeff Squyres
Cisco Systems