Open MPI logo

Open MPI User's Mailing List Archives

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

Subject: Re: [OMPI users] affinity issues under cpuset torque 1.8.1
From: Ralph Castain (rhc_at_[hidden])
Date: 2014-06-19 14:44:02


Sorry, I should have been clearer - I was asking if cores 8-11 are all on one socket, or span multiple sockets

On Jun 19, 2014, at 11:36 AM, Brock Palen <brockp_at_[hidden]> wrote:

> Ralph,
>
> It was a large job spread across. Our system allows users to ask for 'procs' which are laid out in any format.
>
> The list:
>
>> [nyx5406:2][nyx5427:2][nyx5506:2][nyx5311:3]
>> [nyx5329:4][nyx5398:4][nyx5396:11][nyx5397:11]
>> [nyx5409:11][nyx5411:11][nyx5412:3]
>
> Shows that nyx5406 had 2 cores, nyx5427 also 2, nyx5411 had 11.
>
> They could be spread across any number of sockets configuration. We start very lax "user requests X procs" and then the user can request more strict requirements from there. We support mostly serial users, and users can colocate on nodes.
>
> That is good to know, I think we would want to turn our default to 'bind to core' except for our few users who use hybrid mode.
>
> Our CPU set tells you what cores the job is assigned. So in the problem case provided, the cpuset/cgroup shows only cores 8-11 are available to this job on this node.
>
> Brock Palen
> www.umich.edu/~brockp
> CAEN Advanced Computing
> XSEDE Campus Champion
> brockp_at_[hidden]
> (734)936-1985
>
>
>
> On Jun 18, 2014, at 11:10 PM, Ralph Castain <rhc_at_[hidden]> wrote:
>
>> The default binding option depends on the number of procs - it is bind-to core for np=2, and bind-to socket for np > 2. You never said, but should I assume you ran 4 ranks? If so, then we should be trying to bind-to socket.
>>
>> I'm not sure what your cpuset is telling us - are you binding us to a socket? Are some cpus in one socket, and some in another?
>>
>> It could be that the cpuset + bind-to socket is resulting in some odd behavior, but I'd need a little more info to narrow it down.
>>
>>
>> On Jun 18, 2014, at 7:48 PM, Brock Palen <brockp_at_[hidden]> wrote:
>>
>>> I have started using 1.8.1 for some codes (meep in this case) and it sometimes works fine, but in a few cases I am seeing ranks being given overlapping CPU assignments, not always though.
>>>
>>> Example job, default binding options (so by-core right?):
>>>
>>> Assigned nodes, the one in question is nyx5398, we use torque CPU sets, and use TM to spawn.
>>>
>>> [nyx5406:2][nyx5427:2][nyx5506:2][nyx5311:3]
>>> [nyx5329:4][nyx5398:4][nyx5396:11][nyx5397:11]
>>> [nyx5409:11][nyx5411:11][nyx5412:3]
>>>
>>> [root_at_nyx5398 ~]# hwloc-bind --get --pid 16065
>>> 0x00000200
>>> [root_at_nyx5398 ~]# hwloc-bind --get --pid 16066
>>> 0x00000800
>>> [root_at_nyx5398 ~]# hwloc-bind --get --pid 16067
>>> 0x00000200
>>> [root_at_nyx5398 ~]# hwloc-bind --get --pid 16068
>>> 0x00000800
>>>
>>> [root_at_nyx5398 ~]# cat /dev/cpuset/torque/12703230.nyx.engin.umich.edu/cpus
>>> 8-11
>>>
>>> So torque claims the CPU set setup for the job has 4 cores, but as you can see the ranks were giving identical binding.
>>>
>>> I checked the pids they were part of the correct CPU set, I also checked, orted:
>>>
>>> [root_at_nyx5398 ~]# hwloc-bind --get --pid 16064
>>> 0x00000f00
>>> [root_at_nyx5398 ~]# hwloc-calc --intersect PU 16064
>>> ignored unrecognized argument 16064
>>>
>>> [root_at_nyx5398 ~]# hwloc-calc --intersect PU 0x00000f00
>>> 8,9,10,11
>>>
>>> Which is exactly what I would expect.
>>>
>>> So ummm, i'm lost why this might happen? What else should I check? Like I said not all jobs show this behavior.
>>>
>>> Brock Palen
>>> www.umich.edu/~brockp
>>> CAEN Advanced Computing
>>> XSEDE Campus Champion
>>> brockp_at_[hidden]
>>> (734)936-1985
>>>
>>>
>>>
>>> _______________________________________________
>>> users mailing list
>>> users_at_[hidden]
>>> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users
>>> Link to this post: http://www.open-mpi.org/community/lists/users/2014/06/24672.php
>>
>> _______________________________________________
>> users mailing list
>> users_at_[hidden]
>> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users
>> Link to this post: http://www.open-mpi.org/community/lists/users/2014/06/24673.php
>
> _______________________________________________
> users mailing list
> users_at_[hidden]
> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users
> Link to this post: http://www.open-mpi.org/community/lists/users/2014/06/24675.php