On Oct 11, 2011, at 1:40 PM, George Bosilca wrote:
> Unfortunately, this issue appears as having been introduced by a certain
> change that was supposed to make the ORTE code more debugging-friendly
> [obviously]. That particular change duplicated the epoch-tainted error
> managers into their "default" version, more stable and also providing less
> features.
Interesting - I didn't see any "comm failure" calls back to the errmgr.
However, it is always possible I missed them...I'll check again.
>
> The patches (25245, 25248) proposed so far as a solution to this problem
> should be removed, as they do not really solve the problem, instead they
> alleviate the symptoms. From here there are two possible fixes:
>
> 1. Put back the code dealing with the daemons leaving the job in the
> "default" version of the orted error manager.
>
> Here are the lines to be added in update_status in
> orte/mca/errmgr/default_orted/errmgr_default_orted.c:
>
> if (0 == orte_routed.num_routes() &&
> 0 == opal_list_get_size(&orte_local_children)) {
> orte_quit();
> }
Thanks for looking at this more closely. I'll restore those lines, and see if
we are actually getting there. Could be the system I'm using behaves
differently.
>
> 2. Remove the "default" versions and fall back to a single sane set of error
> managers.
No thanks, I'd rather only go thru this recovery process once :-)
>
> george.
>
> PS: I feel compelled to clarify a statement here. The failure detection at
> the socket level, is working properly under all circumstances, how we deal
> with it at the upper levels suffered some mishandling lately.
>
>
> On Oct 11, 2011, at 13:17 , George Bosilca wrote:
>
>>
>> On Oct 11, 2011, at 01:17 , Ralph Castain wrote:
>>
>>>
>>> On Oct 10, 2011, at 11:29 AM, Ralph Castain wrote:
>>>
>>>>
>>>> On Oct 10, 2011, at 11:14 AM, George Bosilca wrote:
>>>>
>>>>> Ralph,
>>>>>
>>>>> If you don't mind I would like to understand this issue a little bit
>>>>> more. What exactly is broken in the termination detection?
>>>>>
>>>>>> From a network point of view, there is a slight issue with the commit
>>>>>> 25245. A direct call to exit will close all pending sockets, with a
>>>>>> linger of 60 seconds (quite bad if you use static ports as an example).
>>>>>> There are proper protocols to shutdown sockets in a reliable way, maybe
>>>>>> it is time to implement one of them.
>>>
>>> I should have clarified something here. The change in r25245 doesn't cause
>>> the daemons to directly call exit. It causes them to call orte_quit, which
>>> runs thru orte_finalize on its way out. So if the oob properly shuts down
>>> its sockets, then there will be no issue with lingering sockets.
>>>
>>> The question has always been: does the oob properly shut down sockets? If
>>> someone has time to address that issue, it would indeed be helpful.
>>
>> Please, allow me to clarify this one. We lived with the OOB issue for 7 (oh
>> a lucky number) years, and we always manage to patch the rest of the ORTE to
>> deal with.
>>
>> Why suddenly someone would bother to fix it?
>>
>> george.
>>
>>>
>>>
>>>>
>>>> Did I miss this message somewhere? I didn't receive it.
>>>>
>>>> The problem is that the daemons with children are not seeing the sockets
>>>> close when their child daemons terminate. So the lowest layer of daemons
>>>> terminate (as they have no children). The next layer does not terminate,
>>>> leaving mpirun hanging.
>>>>
>>>> What's interesting is that mpirun does see the sockets directly connected
>>>> to it close. As you probably recall, the binomial routed module has
>>>> daemons directly callback to mpirun, so that connection is created. The
>>>> socket to the next level of daemon gets created later - it is this later
>>>> socket whose closure isn't being detected.
>>>>
>>>> From what I could see, it appeared to be a progress issue. The socket may
>>>> be closed, but the event lib may not be progressing to detect it. As we
>>>> intend to fix that anyway with the async progress work, and static ports
>>>> are not the default behavior (and rarely used), I didn't feel it worth
>>>> leaving the trunk hanging while continuing to chase it.
>>>>
>>>> If someone has time, properly closing the sockets is something we've
>>>> talked about for quite awhile, but never gotten around to doing. I'm not
>>>> sure it will resolve this problem, but would be worth doing anyway.
>>>>
>>>>
>>>>>
>>>>> Thanks,
>>>>> george.
>>>>>
>>>>> On Oct 10, 2011, at 12:40 , Ralph Castain wrote:
>>>>>
>>>>>> It wasn't the launcher that was broken, but termination detection, and
>>>>>> not for all environments (e.g., worked fine for slurm). It is a
>>>>>> progress-related issue.
>>>>>>
>>>>>> Should be fixed in r25245.
>>>>>>
>>>>>>
>>>>>> On Oct 10, 2011, at 8:33 AM, Shamis, Pavel wrote:
>>>>>>
>>>>>>> + 1 , I see the same issue.
>>>>>>>
>>>>>>>> -----Original Message-----
>>>>>>>> From: [email protected] [mailto:[email protected]]
>>>>>>>> On Behalf Of Yevgeny Kliteynik
>>>>>>>> Sent: Monday, October 10, 2011 10:24 AM
>>>>>>>> To: OpenMPI Devel
>>>>>>>> Subject: [OMPI devel] Launcher in trunk is broken?
>>>>>>>>
>>>>>>>> It looks like the process launcher is broken in the OMPI trunk:
>>>>>>>> If you run any simple test (not necessarily including MPI calls) on 4
>>>>>>>> or
>>>>>>>> more nodes, the MPI processes won't be killed after the test finishes.
>>>>>>>>
>>>>>>>> $ mpirun -host host_1,host_2,host_3,host_4 -np 4 --mca btl sm,tcp,self
>>>>>>>> /bin/hostname
>>>>>>>>
>>>>>>>> Output:
>>>>>>>> host_1
>>>>>>>> host_2
>>>>>>>> host_3
>>>>>>>> host_4
>>>>>>>>
>>>>>>>> And test is hanging......
>>>>>>>>
>>>>>>>> I have an older trunk (r25228), and everything is OK there.
>>>>>>>> Not sure if it means that something was broken after that, or the
>>>>>>>> problem
>>>>>>>> existed before, but kicked in only now due to some other change.
>>>>>>>>
>>>>>>>> -- YK
>>>>>>>> _______________________________________________
>>>>>>>> devel mailing list
>>>>>>>> [email protected]
>>>>>>>> hxxp://www.open-mpi.org/mailman/listinfo.cgi/devel
>>>>>>>
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> devel mailing list
>>>>>>> [email protected]
>>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> devel mailing list
>>>>>> [email protected]
>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> devel mailing list
>>>>> [email protected]
>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>>>>
>>>
>>>
>>> _______________________________________________
>>> devel mailing list
>>> [email protected]
>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>>
>>
>> _______________________________________________
>> devel mailing list
>> [email protected]
>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>
>
> _______________________________________________
> devel mailing list
> [email protected]
> http://www.open-mpi.org/mailman/listinfo.cgi/devel
