[moody20@llnl.gov: Re: Startup-Protocol of MVAPICH-1.0]

[Pete Wyckoff]

frank.mietke at informatik.tu-chemnitz.de wrote on Wed, 02 Apr 2008 17:46 +0200:
> If they would freeze this interface it would be great. This interface is
> abstract and if there is the need to exchange new or more information they can do
> this without changing the adopted files. The version variable in ib.c is the
> version of this collective interface in MVAPICH-1.0. So, if there is a change in
> this abstract interface then they change this version variable as well. But
> these changes should be rare. I think with this interface they made it more
> maintainable for us.

I've checked in your patch, including the four files copied from
mvapich and a license file.

It is a bit unfortunate to have to copy those files into mpiexec
almost exactly (modulo your compile warning fixes), but this is
probably the most easily maintainable approach.  The new code is not
integrated fully into mpiexec, meaning that error messages will not
look consistent, nor will it use the same malloc and read_full
helper routines, but instead does its own.  However, keeping it less
integrated means we could re-copy the pmgr* files from mvapich again
in the future if they change the protocol.  We can re-evaluate this
later if things seem stable.

Adam:  if you would be kind enough to let us know if there are
changes in pmgr*, we would definitely appreciate some advance
warning there.  And please continue to bump the version number as
you do so that mpiexec can support all versions of mvapich.

There are three big changes now in SVN:  mvapich 1.0, totalvew for
mpich2, infinipath.  Probably it is time for another release
already.  I'll give interested parties a couple of weeks to test
it out, then will plan to spin the release hopefully May 5 ish.

Thanks to all, as usual!

		-- Pete