Open MPI logo

Open MPI User's Mailing List Archives

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

From: Mike Houston (mhouston_at_[hidden])
Date: 2007-03-26 16:20:17


Sadly, I've just hit this problem again, so I'll have to find another
MPI implementation as I have a paper deadline quickly approaching.

I'm using single threads now, but I had very similar issues when using
multiple threads and issuing send/recv on one thread and waiting on a
posted MPI_Recv on another. The issue seems to actually be with
MPI_Gets. I can do heavy MPI_Put's and things seem okay. But as soon
as I have a similar communication pattern with MPI_Get's things get
unstable.

-Mike

Brian Barrett wrote:
> Mike -
>
> In Open MPI 1.2, one-sided is implemented over point-to-point, so I
> would expect it to be slower. This may or may not be addressed in a
> future version of Open MPI (I would guess so, but don't want to
> commit to it). Where you using multiple threads? If so, how?
>
> On the good news, I think your call stack looked similar to what I
> was seeing, so hopefully I can make some progress on a real solution.
>
> Brian
>
> On Mar 20, 2007, at 8:54 PM, Mike Houston wrote:
>
>
>> Well, I've managed to get a working solution, but I'm not sure how
>> I got
>> there. I built a test case that looked like a nice simple version of
>> what I was trying to do and it worked, so I moved the test code
>> into my
>> implementation and low and behold it works. I must have been doing
>> something a little funky in the original pass, likely causing a stack
>> smash somewhere or trying to do a get/put out of bounds.
>>
>> If I have any more problems, I'll let y'all know. I've tested pretty
>> heavy usage up to 128 MPI processes across 16 nodes and things seem to
>> be behaving. I did notice that single sided transfers seem to be a
>> little slower than explicit send/recv, at least on GigE. Once I do
>> some
>> more testing, I'll bring things up on IB and see how things are going.
>>
>> -Mike
>>
>> Mike Houston wrote:
>>
>>> Brian Barrett wrote:
>>>
>>>
>>>> On Mar 20, 2007, at 3:15 PM, Mike Houston wrote:
>>>>
>>>>
>>>>
>>>>
>>>>> If I only do gets/puts, things seem to be working correctly with
>>>>> version
>>>>> 1.2. However, if I have a posted Irecv on the target node and
>>>>> issue a
>>>>> MPI_Get against that target, MPI_Test on the posed IRecv causes a
>>>>> segfaults:
>>>>>
>>>>> Anyone have suggestions? Sadly, I need to have IRecv's posted.
>>>>> I'll
>>>>> attempt to find a workaround, but it looks like the posed IRecv is
>>>>> getting all the data of the MPI_Get from the other node. It's like
>>>>> the
>>>>> message tagging is getting ignored. I've never tried posting two
>>>>> different IRecv's with different message tags either...
>>>>>
>>>>>
>>>>>
>>>> Hi Mike -
>>>>
>>>> I've spent some time this afternoon looking at the problem and have
>>>> some ideas on what could be happening. I don't think it's a data
>>>> mismatch (the data intended for the IRecv getting delivered to the
>>>> Get), but more a problem with the call to MPI_Test perturbing the
>>>> progress flow of the one-sided engine. I can see one or two places
>>>> where it's possible this could happen, although I'm having trouble
>>>> replicating the problem with any test case I can write. Is it
>>>> possible for you to share the code causing the problem (or some
>>>> small
>>>> test case)? It would make me feel considerably better if I could
>>>> really understand the conditions required to end up in a seg fault
>>>> state.
>>>>
>>>> Thanks,
>>>>
>>>> Brian
>>>>
>>>>
>>>>
>>> Well, I can give you a linux x86 binary if that would do it. The
>>> code
>>> is huge as it's part of a much larger system, so there is no such
>>> thing
>>> as a simple case at the moment, and the code is in pieces an largely
>>> unrunnable now with all the hacking...
>>>
>>> I basically have one thread spinning on an MPI_Test on a posted IRecv
>>> while being used as the target to the MPI_Get. I'll see if I can
>>> hack
>>> together a simple version that breaks late tonight. I've just played
>>> with posting a send to that IRecv, issuing the MPI_Get,
>>> handshaking and
>>> then posting another IRecv and the MPI_Test continues to eat it,
>>> but in
>>> a memcpy:
>>>
>>> #0 0x001c068c in memcpy () from /lib/libc.so.6
>>> #1 0x00e412d9 in ompi_convertor_pack (pConv=0x83c1198, iov=0xa0,
>>> out_size=0xaffc1fd8, max_data=0xaffc1fdc) at convertor.c:254
>>> #2 0x00ea265d in ompi_osc_pt2pt_replyreq_send (module=0x856e668,
>>> replyreq=0x83c1180) at osc_pt2pt_data_move.c:411
>>> #3 0x00ea0ebe in ompi_osc_pt2pt_component_fragment_cb
>>> (pt2pt_buffer=0x8573380) at osc_pt2pt_component.c:582
>>> #4 0x00ea1389 in ompi_osc_pt2pt_progress () at
>>> osc_pt2pt_component.c:769
>>> #5 0x00aa3019 in opal_progress () at runtime/opal_progress.c:288
>>> #6 0x00ea59e5 in ompi_osc_pt2pt_passive_unlock (module=0x856e668,
>>> origin=1, count=1) at osc_pt2pt_sync.c:60
>>> #7 0x00ea0cd2 in ompi_osc_pt2pt_component_fragment_cb
>>> (pt2pt_buffer=0x856f300) at osc_pt2pt_component.c:688
>>> #8 0x00ea1389 in ompi_osc_pt2pt_progress () at
>>> osc_pt2pt_component.c:769
>>> #9 0x00aa3019 in opal_progress () at runtime/opal_progress.c:288
>>> #10 0x00e33f05 in ompi_request_test (rptr=0xaffc2430,
>>> completed=0xaffc2434, status=0xaffc23fc) at request/req_test.c:82
>>> #11 0x00e61770 in PMPI_Test (request=0xaffc2430,
>>> completed=0xaffc2434,
>>> status=0xaffc23fc) at ptest.c:52
>>>
>>> -Mike
>>> _______________________________________________
>>> 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
>
>