[Users] Ovirt Memory Page Sharing or to less memory

[Itamar Heim]

On 10/17/2012 11:16 PM, Alan Johnson wrote:
> On Mon, Oct 15, 2012 at 6:51 PM, Mathok <mathok1 at googlemail.com
> <mailto:mathok1 at googlemail.com>> wrote:
>
>     It is possible to increase the level of memory page sharing on a
>     cluster above the 200% with a option and start the vms after that
>     change ? or it is possible to deactivate the memory threshold with
>     any option?
>
>
> First, let me be clear that this is complete speculation.  That said,
> my interpretation of these features based on how the GUI behaves is that
> the Cluster Memory Optimization (CMO) setting changes the
> default Physical Memory Guarantee (PMG) on new VMs under the Resource
> Allocation tab.  However, you could set that number manually to whatever
> you want.  I have been setting it to 1/4 of the Memory Size (MS) on the
> General tab as a general rule, but varying it by VM as needed.

that is correct. the default value for "reserved memory" at vm level is 
based on cluster over commit.

cluster over commit can be set to a custom value via the api/sdk/cli.

you should validate ksm is up and running.
you should also note ksm by default only kicks in around 80% host memory 
utilization. so if you want to be more aggressive on sharing memory 
pages (at the expense of a bit more cpu cycles for ksm), you can 
configure it at vdsm level to start at a lower threshold.

also, you should note that during vm startup, until vm goes to up, 
engine assumes it is taking all of its defined RAM, then moves to check 
actual ram consumption of the VM.

so between ksm configuration, and the time VMs go to up state, you may 
need to throttle the launch of multiple VMs on same host.

>
> So, to restate, when CMO is set to None, PMG=MS when creating a new VM.
>   When CMO is set to Server, PMG=MS/150%.  When CMO is Desktop,
> PMG=MS/200%.  But you can still set the PMG to any custom value any time
> the VM is not running.  That much I am sure of.  What I have not tested
> is my expectation that the engine will pile the VMs on the hosts based
> purely on their the greater of PMG and actual VM use.  It would be
> reasonably plausible that the engine limits VMs on the hosts based on
> CMO as well as PMG and actual use, but I just don't know for sure either
> way.

CMO - total potential RAM of the VMs on the host should not exceed host 
ram * CMO.

PMG - total reserved RAM of the VMs on the host, should not exceed host ram

(that's "close enough to the truth", calculation is taking some 
overheads into account, etc.)


>
> I would greatly appreciated it if some guru could school me on this point.
>
> _______________
> Alan Johnson
> alan at datdec.com <mailto:alan at datdec.com>
>
>
> _______________________________________________
> Users mailing list
> Users at ovirt.org
> http://lists.ovirt.org/mailman/listinfo/users
>