On Aug 21, 2009, at 07:33 , Ralph Castain wrote:
Hi Rich
On Aug 21, 2009, at 5:14 AM, Graham, Richard L. wrote:
I have several questions here - since process migration is an open
research question,
and there is more than one way to address the issue -
- Is this being implemented as a component, so that other
approaches can be used ?
Absolutely - we have several organizations involved, all with
competing approaches
This become a recurrent statement, every time something smelly should
be defended.
- If so, what sort of component interface is being considered ?
Still being determined. One reason for exposing the frameworks at
this time is that much of the ongoing effort may occur in separate,
hidden tmp branches for proprietary reasons. The eventual source
code for those components may well never be committed, but the
frameworks need to be in the system so that MCA will pickup the
binary modules.
This is a wrong reason. We did multi-institution work in the past
starting from a tmp branch. The only overhead is that somebody has to
keep the copy in sync with the trunk, but this approach at least has
the merit to keep our trunk [more or less] clean.
I deliberately left the frameworks "inactive" so that changes in the
APIs can be done as the work progresses without impacting the OMPI
community. The participants need to develop a little further before
we fully understand what the APIs need to be - a little hard to
openly discuss them without exposing potentially proprietary algos.
The hope is that, as people develop their components, they can
identify missing needs in the API so at least that much can be
safely communicated and resolved.
- What is the impact (or expected impact) on the rest of the system ?
For anyone who doesn't explicitly build it and turn it on, nothing.
For those who do, there will be no impact on MPI procs themselves as
they won't load nor activate these frameworks. There will be an
increased orted footprint and activity level, which could
potentially reduce performance by stealing cycles from MPI procs -
obviously, that depends on #procs vs cores and other factors.
If nobody knows how to do it, this _clearly_ should be a good enough
reason to do the work in a temp branch before polluting the trunk.
george.
I'm speaking solely of the RTE impact and issues here, of course -
Josh would have to address the MPI perspective.
HTH
Ralph
Thanks,
Rich
On 8/20/09 6:57 PM, "Ralph Castain" <r...@open-mpi.org> wrote:
Hmmm...I'm afraid I cannot entirely substantiate your concerns.
Everything compiles just fine for me under both Linux and OSX. True,
the files are not completely instantiated on the trunk at this time,
but they also are not activated on the trunk for precisely that
reason.
Can you provide an example of where something isn't building? I can't
find it on any platform available to me.
Thanks
Ralph
On Aug 20, 2009, at 4:32 PM, George Bosilca wrote:
Ralph,
I'm a little bit concerned about the commits stated below as their
quality doesn't match the usual quality standards of the trunk.
There are several issues: mostly empty files, names coming from
other frameworks, failure to compile on several platforms (including
Linux and MAC OS X). I'm more than happy to see research code in the
trunk, and I will be really interested to see the proof that
preemptive migration works. However, the quality of the trunk should
not suffer.
Moreover, we have mercurial and svn temporary repositories for such
kind of work. And we did force people in the past to work on
temporary branches before such large commits on the trunk.
Please reconsider your patches.
Thanks,
george.
https://svn.open-mpi.org/trac/ompi/changeset/21849
https://svn.open-mpi.org/trac/ompi/changeset/21850
https://svn.open-mpi.org/trac/ompi/changeset/21848
https://svn.open-mpi.org/trac/ompi/changeset/21847
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
