[mpich-discuss] Joining a job from a process without mpiexec parent

Dmitriy Lyubimov dlieu.7 at gmail.com
Mon Mar 5 12:19:01 CST 2018


Mike: exactly. That's about where our conversation ended circa 2 years ago.
IIRC I was told along the lines that it was not supported, and i pointed
out that MPI standard  recommends (if not requires) this way of doing
things. I figured 3.3 is now getting closer to release now that i try to
see if something has changed for it, but I guess I have to assume the
developing parties are not interested in that user experience themselves,
so I just have to conclude that nothing unfortunately for me has  changed
in that regard since 3.2 in terms of UI experience.

Which is a shame really, because this is one of the easiest ways for
integrating 3rd party process management including 3rd party clouds,
containers, resource managers etc. etc.

On Sat, Mar 3, 2018 at 2:30 PM, <Michael.Stokes at uah.edu> wrote:

> Dmitriy,
>
> I've written a server and client code example and compiled it using MPICH
> 3.2 using MPI_Comm_{accept,connect}() to see if MPICH 3.2 will form a
> intercommunicator between the two processes without using a launcher
> {mpiexec,mpirun}.  Before I started, I read the instructions ( MPI Specs
> 3.1 section 10.5.2). It states
>
> "A high-quality implementation will allow any process (including those not
> started with a
> “parallel application” mechanism) to become an MPI process by calling
> MPI_INIT. Such
> a process can then connect to other MPI processes using the
> MPI_COMM_ACCEPT and
> MPI_COMM_CONNECT routines, or spawn other MPI processes. MPI does not
> mandate
> this behavior, but strongly encourages it where technically feasible."
>
> Case 1) With the server and client run directly from the command-line the
> server executes down to the MPI_Comm_accept() call as anticipated, but when
> the port name is entered into the client process, it dies on
> MPI_Comm_connect() with the error ...
>
> match_arg (utils/args/args.c:159): unrecognized argument pmi_args
> [mpiexec at Que] HYDU_parse_array (utils/args/args.c:174): argument matching
> returned error
> [mpiexec at Que] parse_args (ui/mpich/utils.c:1596): error parsing input
> array
> [mpiexec at Que] HYD_uii_mpx_get_parameters (ui/mpich/utils.c:1648): unable
> to parse user arguments
> [mpiexec at Que] main (ui/mpich/mpiexec.c:153): error parsing parameters
>
> Case 2) When I run the server directly from the command-line and launch
> the client using mpirun, the same sort of error messages emerge but this
> time from the server.  Since the server blocks on the accept() call, the
> client must be successful in sending something to the server.  But
> apparently not what it is expecting.
>
> So my conclusion is that MPICH is attempting to support the Singleton
> process, but either has a bug or there might be an installation issue on my
> end.
>
> Hopefully someone can help sort this.
>
> --Mike
>
>
> _______________________________________________
> discuss mailing list     discuss at mpich.org
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/discuss
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpich.org/pipermail/discuss/attachments/20180305/ef88d5bf/attachment.html>
-------------- next part --------------
_______________________________________________
discuss mailing list     discuss at mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss


More information about the discuss mailing list