[gridengine users] Tight integration problem with mvapich2 2.0

Götz Waschk goetz.waschk at gmail.com
Wed Jan 7 12:42:59 UTC 2015


Dear Reuti,

On Wed, Jan 7, 2015 at 1:36 PM, Reuti <reuti at staff.uni-marburg.de> wrote:

> Hi,
>
> > Am 07.01.2015 um 13:18 schrieb Götz Waschk <goetz.waschk at gmail.com>:
> >
> > Hi everyone,
> >
> > I have noticed one difference in the behaviour of mvapich2 between 1.8
> and 2.0rc1: with the former, hydra_pmi_proxy and all mpi processes are in
> the same process group, but with the later version, they all have a
> different pgid. Just killing hydra_pmi_proxy will terminate all mpi
> processes with 1.8 but not with 2.0rc1.
>
> Yes, the default is to kill the process group (done by a negative pid as
> argument). But you can try to use the additional group id to look for all
> kids:
>
> $ qconf -sconf
> ...
> execd_params                 ENABLE_ADDGRP_KILL=TRUE
>
> it appears I already have this set:
execd_params                 ENABLE_ADDGRP_KILL H_MEMORYLOCKED=infinity \
                             ENABLE_BINDING

So it doesn't help here unfortunately.

Regards, Götz
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://gridengine.org/pipermail/users/attachments/20150107/26a2f7d9/attachment.html>


More information about the users mailing list