[Mauiusers] detecting and scheduling vmem
bchristiansen at clusterresources.com
Mon Mar 16 16:07:14 MDT 2009
Looking in MPBSI.c, it appears that nodememovercommitfactor is already
implemented to work with swap, except that in in _MPBSIGetSSSStatus the
setting of N->CRes.Swap is after the overcommitting work is done. How
does the following patch work for you?
@@ -6263,6 +6263,11 @@
} /* END else if (!strcmp(Name,"message")) */
} /* END while (ptr != NULL) */
+ /* NOTE: PBS totmem = swap + RAM */
+ if (TotMem > 0)
+ N->CRes.Swap = TotMem;
if (MSched.NodeMemOverCommitFactor > 0.0)
/* NOTE: both real memory and swap overcommitted */
@@ -6276,11 +6281,6 @@
/* memory factor not applied to node load */
- /* NOTE: PBS totmem = swap + RAM */
- if (TotMem > 0)
- N->CRes.Swap = TotMem;
} /* END __MPBSIGetSSSStatus() */
Gareth.Williams at csiro.au wrote:
> I passed your email on to CR in my open moab ticket. This seems like a straightforward and good idea to me.
> From: Roy Dragseth [roy.dragseth at cc.uit.no]
> Sent: Tuesday, 24 February 2009 9:37 PM
> To: mauiusers at supercluster.org
> Subject: Re: [Mauiusers] detecting and scheduling vmem
> On Friday 20 February 2009 07:08:42 Gareth.Williams at csiro.au wrote:
>> Hi All,
>> I have been working with CR to get a patch in maui and moab to improve
>> treatment of vmem. The patch is in maui-3.2.6p21-snap.1234905291 and will
>> be in upcoming moab releases. Many thanks to Brian Christiansen at CR.
>> With the patch, maui/moab correctly detects a node's available vmem as
>> being the total physical memory plus swap space. This can then be
>> scheduled/allocated by requesting vmem and job's virtual memory allocation
>> can be limited on a per process basis and periodically measured (and action
>> taken or overuse) on a per job basis.
>> I think this makes scheduling on vmem a good option. It's a more
>> predictable quantity than mem/pmem.
> This is great, but is it possible to get
> TotMem = (int)(TotMem * MSched.NodeMemOverCommitFactor);
> at line 6384? This will make the NODEMEMOVERCOMMITFACTOR work both for mem
> and swap as earlier.
> We want to use a slight overcommit to allow a user to specify -lpvmem=2gb
> instead of -lpvmem=2000mb on a machine with 2GB RAM per core. Some of the ram
> isn't available to userspace, so using 2gb will underutilize the cores on a
> node and users aren't really good at guessing the memory consumption anyway.
> mauiusers mailing list
> mauiusers at supercluster.org
More information about the mauiusers