Open MPI logo

Open MPI User's Mailing List Archives

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

Subject: Re: [OMPI users] problem when mpi_paffinity_alone is set to 1
From: Ralph Castain (rhc_at_[hidden])
Date: 2008-08-22 09:51:02


I believe I have found the problem, and it may indeed relate to the
change in paffinity. By any chance, do you have unfilled sockets on
that machine? Could you provide the output from something like "cat /
proc/cpuinfo" (or the equiv for your system) so we could see what
physical processors and sockets are present?

If I'm correct as to the problem, here is the issue. OMPI has (until
now) always assumed that the #logical processors (or sockets, or
cores) was the same as the #physical processors (or sockets, or
cores). As a result, several key subsystems were written without
making any distinction as to which (logical vs physical) they were
referring to. This was no problem until we recently encountered
systems with "holes" in their system - a processor turned "off", or a
socket unpopulated, etc.

In this case, the local processor id no longer matches the physical
processor id (ditto for sockets and cores). We adjusted the paffinity
subsystem to deal with it - took much more effort than we would have
liked, and exposed lots of inconsistencies in how the base operating
systems handle such situations.

Unfortunately, having gotten that straightened out, it is possible
that you have uncovered a similar inconsistency in logical vs physical
in another subsystem. I have asked better eyes than mine to take a
look at that now to confirm - if so, it could take us a little while
to fix.

My request for info was aimed at helping us to determine why your
system is seeing this problem, but our tests didn't. We have tested
the revised paffinity on both completely filled and on at least one
system with "holes", but differences in OS levels, processor types,
etc could have caused our tests to pass while your system fails. I'm
particularly suspicious of the old kernel you are running and how our
revised code will handle it.

For now, I would suggest you work with revisions lower than r19391 -
could you please confirm that r19390 or earlier works?

Thanks
Ralph

On Aug 22, 2008, at 7:21 AM, Camille Coti wrote:

>
> OK, thank you!
>
> Camille
>
> Ralph Castain a écrit :
>> Okay, I'll look into it. I suspect the problem is due to the
>> redefinition of the paffinity API to clarify physical vs logical
>> processors - more than likely, the maffinity interface suffers from
>> the same problem we had to correct over there.
>> We'll report back later with an estimate of how quickly this can be
>> fixed.
>> Thanks
>> Ralph
>> On Aug 22, 2008, at 7:03 AM, Camille Coti wrote:
>>>
>>> Ralph,
>>>
>>> I compiled a clean checkout from the trunk (r19392), the problem
>>> is still the same.
>>>
>>> Camille
>>>
>>>
>>> Ralph Castain a écrit :
>>>> Hi Camille
>>>> What OMPI version are you using? We just changed the paffinity
>>>> module last night, but did nothing to maffinity. However, it is
>>>> possible that the maffinity framework makes some calls into
>>>> paffinity that need to adjust.
>>>> So version number would help a great deal in this case.
>>>> Thanks
>>>> Ralph
>>>> On Aug 22, 2008, at 5:23 AM, Camille Coti wrote:
>>>>> Hello,
>>>>>
>>>>> I am trying to run applications on a shared-memory machine. For
>>>>> the moment I am just trying to run tests on point-to-point
>>>>> communications (a trivial token ring) and collective operations
>>>>> (from the SkaMPI tests suite).
>>>>>
>>>>> It runs smoothly if mpi_paffinity_alone is set to 0. For a
>>>>> number of processes which is larger than about 10, global
>>>>> communications just don't seem possible. Point-to-point
>>>>> communications seem to be OK.
>>>>>
>>>>> But when I specify --mca mpi_paffinity_alone 1 in my command
>>>>> line, I get the following error:
>>>>>
>>>>> mbind: Invalid argument
>>>>>
>>>>> I looked into the code of maffinity/libnuma, and found out the
>>>>> error comes from
>>>>>
>>>>> numa_setlocal_memory(segments[i].mbs_start_addr,
>>>>> segments[i].mbs_len);
>>>>>
>>>>> in maffinity_libnuma_module.c.
>>>>>
>>>>> The machine I am using is a Linux box running a 2.6.5-7 kernel.
>>>>>
>>>>> Has anyone experienced a similar problem?
>>>>>
>>>>> Camille
>>>>> _______________________________________________
>>>>> 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