Open MPI logo

Open MPI User's Mailing List Archives

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

Subject: Re: [OMPI users] error performing MPI_Comm_spawn
From: Ralph Castain (rhc_at_[hidden])
Date: 2009-12-17 15:48:54


Will be in the 1.4 nightly tarball generated later tonight...

Thanks again
Ralph

On Dec 17, 2009, at 4:07 AM, Marcia Cristina Cera wrote:

> very good news!!!!
> I will wait carefully for the release :)
>
> Thanks, Ralph
> márcia.
>
> On Wed, Dec 16, 2009 at 10:56 PM, Ralph Castain <rhc_at_[hidden]> wrote:
> Ah crumb - I found the problem. Sigh.
>
> I actually fixed this in the trunk over 5 months ago when the problem first surfaced in my own testing, but it never came across to the stable release branch. The problem is that we weren't serializing the comm_spawn requests, and so the launch system gets confused over what has and hasn't completed launch. That's why it works fine on the trunk.
>
> I'm creating the 1.4 patch right now. Thanks for catching this. Old brain completely forgot until I started tracking the commit history and found my own footprints!
>
> Ralph
>
> On Dec 16, 2009, at 5:43 AM, Marcia Cristina Cera wrote:
>
>> Hi Ralph,
>>
>> I am afraid I have been a little hasty!
>> I remake my tests with more care and I got the same error also with the 1.3.3 :-/
>> but in such version the error happens after some successful executions... because of that I did not realize before!
>> Furthermore, I increased the number of levels of the tree (that means have more concurrently dynamic process creations in the lower levels) and I never arrive to execute without error, unless I add the delay.
>> Perhaps the problem might even be a race condition :(
>>
>> I test with LAM/MPI 7.1.4 and in a first moment it works fine. I have work with LAM for years, but I migrate o OpenMP last year once LAM will be discontinued...
>>
>> I think that I can continue the development of my application adding the delay, while I wait for a release... and I leave the performance tests to be made in the future :)
>>
>> Thank you again Ralph,
>> márcia.
>>
>>
>> On Wed, Dec 16, 2009 at 2:17 AM, Ralph Castain <rhc_at_[hidden]> wrote:
>> Okay, I can replicate this.
>>
>> FWIW: your test program works fine with the OMPI trunk and 1.3.3. It only has a problem with 1.4. Since I can replicate it on multiple machines every single time, I don't think it is actually a race condition.
>>
>> I think someone made a change to the 1.4 branch that created a failure mode :-/
>>
>> Will have to get back to you on this - may take awhile, and won't be in the 1.4.1 release.
>>
>> Thanks for the replicator!
>>
>> On Dec 15, 2009, at 8:35 AM, Marcia Cristina Cera wrote:
>>
>>> Thank you, Ralph
>>>
>>> I will use the 1.3.3 for now...
>>> while waiting for a future fix release that break this race condiction.
>>>
>>> márcia
>>>
>>> On Tue, Dec 15, 2009 at 12:58 PM, Ralph Castain <rhc_at_[hidden]> wrote:
>>> Looks to me like it is a race condition, and the timing between 1.3.3 and 1.4 is just enough to trip it. I can break the race, but it will have to be in a future fix release.
>>>
>>> Meantime, your best bet is to either stick with 1.3.3 or add the delay.
>>>
>>> On Dec 15, 2009, at 5:51 AM, Marcia Cristina Cera wrote:
>>>
>>>> Hi,
>>>>
>>>> I intend to develop an application using the MPI_Comm_spawn to create dynamically new MPI tasks (or processes).
>>>> The structure of the program is like a tree: each node creates 2 new ones until reaches a predefined number of levels.
>>>>
>>>> I developed a small program to explain my problem as can be seen in attachment.
>>>> -- start.c: launches (through MPI_Comm_spawn, in which the argv has the level value) the root of the tree (a ch_rec program). Afterward spawn, a message is sent to child and the process block in an MPI_Recv.
>>>> -- ch_rec.c: gets its level value and receives the parent message, then if its level is less than a predefined limit, it will creates 2 children:
>>>> - set the level value;
>>>> - spawn 1 child;
>>>> - send a message;
>>>> - call an MPI_Irecv;
>>>> - repeat the 4 previous steps for the second child;
>>>> - call an MPI_Waitany waiting for children returns.
>>>> When children messages are received, the process send a message to its parent and call MPI_Finalize.
>>>>
>>>> Using the openmpi-1.3.3 version the program runs as expected but with openmpi-1.4 I get the following error:
>>>>
>>>> $ mpirun -np 1 start
>>>> level 0
>>>> level = 1
>>>> Parent sent: level 0 (pid:4279)
>>>> level = 2
>>>> Parent sent: level 1 (pid:4281)
>>>> [xiru-8.portoalegre.grenoble.grid5000.fr:04278] [[42824,0],0] ORTE_ERROR_LOG: Not found in file base/plm_base_launch_support.c at line 758
>>>>
>>>> The error happens when my program try to launch the second child immediately after the first spawn call.
>>>> In my tests I try to put an sleep of 2 second between the first and the second spawn, and then the program runs as expected.
>>>>
>>>> Some one can help me with this version 1.4 bug?
>>>>
>>>> thanks,
>>>> márcia.
>>>>
>>>> <spawn-problem.tar.gz>_______________________________________________
>>>> users mailing list
>>>> users_at_[hidden]
>>>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>>>
>>>
>>> _______________________________________________
>>> users mailing list
>>> users_at_[hidden]
>>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>>>
>>> _______________________________________________
>>> users mailing list
>>> users_at_[hidden]
>>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>>
>>
>> _______________________________________________
>> users mailing list
>> users_at_[hidden]
>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>>
>> _______________________________________________
>> users mailing list
>> users_at_[hidden]
>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>
>
> _______________________________________________
> users mailing list
> users_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/users
>
> _______________________________________________
> users mailing list
> users_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/users