Open MPI logo

Open MPI User's Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Open MPI User's mailing list

Subject: Re: [OMPI users] on-the-fly nodes liberation
From: Ralph Castain (rhc_at_[hidden])
Date: 2009-03-09 08:34:47


I'm afraid not - once started, the orted must stay alive until mpirun
terminates. The problem is that the orteds are used to route messages,
and there is currently no way to remove an orted without breaking this
network.

I know people are investigating this possibility in support of fault
tolerance for when an orted unexpectedly fails, but there is nothing
currently in the code base for this capability.

Ralph

On Mar 9, 2009, at 6:28 AM, Marcia Cristina Cera wrote:

> Hi,
>
> May I sign one orted daemon to finish its execution on-the-fly?
>
> Context: I intend to use OpenMPI in a dynamic resource environment
> as I did with LAM/MPI helped by lamgrow and lamshrink commands.
>
> To perform grow operations (increase the amount of nodes/resources
> on-the-fly) OpenMPI enable an incremental resource utilization. All
> nodes that can be used are listed in the hostifile file (inform as
> mpirun parameter) and according to they are firstly used through
> MPI_Comm_spawn one orded daemon is created in each node. According
> to some first tests, this feature is enough to satisfy our goals.
>
> In the other hand, performing shrink operations, we need to liberate
> nodes to be eventually used by other application/jobs. In other
> words, we must terminate all applications processes and also the
> orted daemon. In the application side, the solution is easy once we
> can notify the processes (by a message or signal) to safety finish
> its execution and perform MPI_Finalize. In the orted side, we must
> finish its execution in the target node and also update its status
> to 'INVALID'.
> How may I do it? Is there a specific signal or procedure to this?
>
> Thank you in advance!
> márcia.
> _______________________________________________
> users mailing list
> users_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/users