[Trillian-users] One Node, One Proc Per Node Jobs

James Pringle jpringle at unh.edu
Tue Dec 6 14:38:46 EST 2016


Dear Jie-Xiang --

    Please see the discussion on
http://trillian-use.sr.unh.edu/index.php/Main_Page ; you can indeed run up
to 32 jobs on the same node (memory and IO constraints apply, of course).
Otherwise you are wasting resources. The way to do it is not to create a
new queue.  Why? because if you are on one node, processes can interfere,
and it is better to have control of all the processes on the node you are
using.

   See also
https://www.nersc.gov/users/computational-systems/retired-systems/hopper/running-jobs/example-batch-scripts/
under the heading "Multiple Serial Jobs" but note that there is not a
serial queue.  (and even when there is, I tend to not use the serial
queue).

    Again, if your jobs are not memory or IO bound, running 1 core per node
is wasting 95% of the power of the machine.

Jamie

On Tue, Dec 6, 2016 at 1:19 PM, Yu, Jiexiang <Jiexiang.Yu at unh.edu> wrote:

> Hi there,
>
> Is that means all the jobs is divided by nodes rather than cores? Can we
> run two jobs on the differenct cores of a same node? If we can, just create
> a new queue, which contains only one or two nodes, for the serial jobs.
>
>
> Jie-Xiang
> ------------------------------
> *From:* Trillian-users <trillian-users-bounces at lists.sr.unh.edu> on
> behalf of James Pringle <jpringle at unh.edu>
> *Sent:* Tuesday, December 6, 2016 12:42 PM
> *To:* Gorby, Matthew
> *Cc:* trillian-users at lists.sr.unh.edu
> *Subject:* Re: [Trillian-users] One Node, One Proc Per Node Jobs
>
> Matt --
>
>    You are right -- in general 1 core/node jobs are a very inefficient use
> of trillian in general, and qstat is generally accurate in what it
> displays. When I have been active on trillian, I have asked people to be
> more considerate and efficient in their use of trillian. There are
> instructions on how to start multiple jobs on a single node on the home
> page.  I can give more pointers.
>
>    The only time it makes sense to have 1 core per node jobs on trillian
> is when the job is IO or memory bound. Even in that case, it is usually
> inefficient to use trillian. Its cores are much slower than even a recent
> laptop, and most of the undergraduate cluster machines on campus are faster
> per-core.
>
>    I think it is acceptable to ask running a bunch of 1 PE per node jobs
> if they have a quantified need to run on trillian the way they are.
>
> Jamie Pringle
>
>
> On Tue, Dec 6, 2016 at 12:24 PM, Gorby, Matthew <Matthew.Gorby at unh.edu>
> wrote:
>
>> Hello Trillian Users,
>>
>>
>> I'm seeing a lot of one node, one proc / node jobs running at the
>> moment.  If this is just the way "qstat -a" displays a type of run I'm not
>> familiar then please excuse this email.  If those jobs really are what they
>> appear to be then may we have one of the admins chime in on this issue?
>> I'd like to know if that is an acceptable use of Trillian's resources.  If
>> not, it would be really great to have the other 31/32 of those
>> procs available.
>>
>>
>> Thanks,
>>
>>
>> -Matt
>>
>>
>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.sr.unh.edu/pipermail/trillian-users/attachments/20161206/d7a36c9d/attachment.html>


More information about the Trillian-users mailing list