Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

Subject: Re: [OMPI devel] IOF 2.0
From: Jeff Squyres (jsquyres_at_[hidden])
Date: 2008-10-17 11:30:46


Thanks Ralph! I'll give the hg branch a shot on my systems today.

I'm particularly excited about orte-iof(1); "mpirun --screen ..." is
now one [major] step closer! :-)

On Oct 17, 2008, at 10:38 AM, Ralph Castain wrote:

> Hello all
>
> As probably most of you know, we have been actively rewriting the I/
> O forwarding subsystem to reduce its complexity (so it is easier to
> maintain), and to enable some new features. This work has been
> completed and will be committed to the trunk later today.
>
> NOTE: it will -not- be included in the 1.3 distribution, but is
> intended for 1.4.
>
> Some of the new features include:
>
> 1. ability to tag output using process name and stream (stdout,
> stderr) using a new mpirun cmd line option --tag-output
>
> 2. ability to direct stdin to any specified rank, or to all ranks,
> or to no ranks using a new mpirun cmd line option --stdin. This
> defaults to rank=0t o preserve today's behavior
>
> 3. addition of a new MPI_Info key "ompi_stdin_target" that allows
> comm_spawn'd jobs the same ability as described in #2. This defaults
> to "none" to preserve today's behavior
>
> 4. a new tool "orte-iof" that takes cmd line options indicating the
> pid of the mpirun you want to get stdout/err from, and the rank of
> the procs from which you want that info. The tool then connects to
> that mpirun and prints the requested info to its own stdout/err.
> Again, the --tag-output cmd line option is available.
>
> You will hopefully find the code much less complex than the earlier
> version. In addition, we have completely eliminated IOF-related
> wireup communications. We retain flow-control for stdin to help
> constrain orted memory footprint - this is now done via an xon/xoff
> protocol as opposed to the prior technique of exchanging an "ack"
> after every message.
>
> I have tested the new IOF on several systems and it appears to be
> functioning fine. However, I do not promise that we won't see a
> hiccup or two once this is introduced into the system.
>
> Please keep me advised of any issues you observe. Hopefully, this
> will transition smoothly.
>
> Ralph
>
> _______________________________________________
> devel mailing list
> devel_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/devel

-- 
Jeff Squyres
Cisco Systems