Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

From: Ralph Castain (rhc_at_[hidden])
Date: 2006-04-20 13:28:08


Looks reasonable - let me see what can be done.

Thanks
Ralph


Greg Watson wrote:
Ok, thanks.

For clarification, the model we're using at the moment looks roughly like this:

orte_init();

forever () {
    if (do_our_stuff() == GAME_OVER)
        break;
    opal_event_loop(OPAL_EVLOOP_ONCE);
}

orte_finalize();

The simplest change for us would be something like:

orte_init();

forever () {
    if (do_our_stuff() == GAME_OVER)
        break;
    if (opal_event_loop(OPAL_EVLOOP_ONCE) != ORTE_SUCCESS) {
        clean_up_our_stuff();
        break;
    }
}

orte_finalize();

Greg


On Apr 20, 2006, at 10:21 AM, Ralph Castain wrote:

You make a good point about the library not calling exit(). I'll have to recruit some help to look at the notion of opal_even_loop returning an error value - it isn't entirely clear who it would return it to in our system,. Even though I understand how someone in your situation would handle it, I have to ensure that it doesn't cause the base system problems, or force a major code revision that would need to be scheduled into the project.

We'll have to get back to you on this - most of the folks are at a workshop this week, so it will probably be next week before we can discuss it.

Ralph


Greg Watson wrote:
The simplest thing for us would be for opal_event_loop() to return an error value. That way we can detect the situation and clean up our system. At the moment we're not trying to restart orted, so clean recovery of orte is not that important, though ultimately I would think it is desirable. Other alternatives are to pass you an error handler that you call, or you could send a signal that we can trap.

>From our perspective, we're simply calling a library that does stuff. Having the library call exit() at any point is a major problem for applications trying to do more than run a single job.

Greg

On Apr 20, 2006, at 9:40 AM, Ralph Castain wrote:

Well, I actually don't know much about opal_event_loop and/or how it is intended to work. My guess is that:

(a) your remote orted is acting as the seed and your local process (the one in Eclipse) is running as a client to that seed - at least, that was the case last I talked to Nathan

(b) when the seed orted dies, it is the oob in your local client that actually detects socket closure and decides that - since it is the seed that has lost contact - the local application must abort.

(c) the errmgr.abort function does exactly what it was supposed to do - it provides an immediate way of killing the local process.

I'd be a little hesitant to recommend overloading the errmgr.abort function as you really do want the local processes to die when losing connection to the seed (at least, until we develop a recovery capability for the seed orted - which is some ways off), and (given the way you are running) I'm not sure you can have a different errmgr for your process while leaving the other one for everyone else.

Probably the best solution for now would be for us to insert a (yet another) MCA parameter into the errmgr that would (if set) have errmgr.abort do something other than exit. The question then is: what would you want it to do?? We need to have it tell the rest of the system to stop trying to send messages etc - right now, I don't think the infrastructure exists to do that short of killing orte.

We could try to have errmgr.abort do an orte_finalize - that would kill the orte system without impacting your host program, I suspect. You would then have to re-initialize, so we'd have to find some way to let you know that we had finalized. I can't swear this will work, though - we might well generate a segfault since this is happening deep down inside the system. We could try it, though.

Would any of that be of help? Do you have any suggestions on how we might let you know that we had finalized?

Ralph


Brian Barrett wrote:
On Apr 19, 2006, at 4:15 PM, Greg Watson wrote:
We've just run across a rather tricky issue. We're calling opal_event_loop() to dispatch orte events to an orted that has been launched separately. However if the orted dies for some reason (gets a signal or whatever) then opal_event_loop() is calling exit(). Needless to say, this is not good behavior us. Any suggestions on how to get around this problem?
Is the orted you are connecting to the "seed" daemon? I think the only time we should be exiting like that is if the orted was the seed daemon. I'm not sure what we want to do if that's the case -- it looks like we're calling errmgr.abort() when badness happens. I wonder if your application can provide its own errmgr component that provides an abort that doesn't actually abort? Just some off the cuff ideas -- Ralph could probably give a better idea of exactly what is happening... Brian
_______________________________________________
devel mailing list
devel@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel


_______________________________________________
devel mailing list
devel@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel