On 8/6/2011 12:59 AM, Daniel Gruber wrote:
Am 03.08.2011 um 10:28 schrieb William Hay:
On 2 August 2011 17:58, Rayson Ho<[email protected]> wrote:
It's a bug introduced by another bug fix in SGE 6.2u5, and Oracle was
first who fixed the bug in Oracle Grid Engine. Then we added a
workaround in SGE 6.2u5p1 in Open Grid Scheduler, and Son of Grid
Engine copied it. I think Univa also fixed the bug at some point, as
the fix was copied by Son of Grid Engine (and dropped the workaround).
OGS will just stick with the workaround as we don't like the
workaround or the fix...
You will just need to upgrade your SGE 6.2u5 cluster with a patched
SGE execd - either compile execd yourself or in fact you can get it
from the hwloc drop-in upgrade package:
http://gridscheduler.sourceforge.net/projects/hwloc/GridEnginehwloc.html
hwloc looks rather interesting. Do your integrations work with other
versions of Grid Engine (we're at 6.2u3)?
Just for completeness: Univa Grid Engine 8.0.1 is going to support
hwloc as well.
Are those changes already in the
https://github.com/gridengine/gridengine git repo?
Thanks,
Bill
Cheers,
Daniel
From poking around on the hwloc pages it appears to support cgroups
which can do a lot more than just bind cpus and memory.
Presumably if one had a cgroup based system one could just extend the
hwloc created cgroup with the required additional features.
William
Rayson
On Tue, Aug 2, 2011 at 8:15 AM, Jesse Becker<[email protected]> wrote:
On Mon, Aug 01, 2011 at 07:41:41PM -0400, William Deegan wrote:
Should the maxvmem column in the accounting file be the true max memory
footprint of the running process? (and children?)
I've seen problems with 6.2u5 in the accounting records. It appears to
"wrap" at 4GB, which probably indicates a 32/64 bit issue. I think
there's information about it in the mailing list.
I'm not sure about child processes.
--
Jesse Becker
NHGRI Linux support (Digicon Contractor)
_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users
