Open Stack

On 09/02/16 16:25, "Chris Friesen" <chris.friesen at windriver.com> wrote:

>On 02/09/2016 04:20 AM, Daniel P. Berrange wrote:
>
>> It not actually a huge waste - in fact it is likely to be beneficial to the
>> overall performance of your VM. You need to remember that there is no direct
>> relationship between allocation of huge pages to KVM and usage of huge pages
>> inside the guest OS - any combination is valid. In particular if you give
>> huge pages to KVM, then this has a performance benefit to the guest, even if
>> the guest OS doesn't use huge pages itself, because it increases the TLB hit
>> rate in the host for memory accesses by the guest. Using huge pages inside
>> the guest OS too, increases the TLB hit rate of the guest hardware further
>> extending this performance benefit. As such hosts which are intended to use
>> huge pages should have all their host memory allocated to the huge page pool,
>> except for what is required to run the various host OS services.
>
>It's not very efficient to use 1G hugepages with instances having less than 1G 
>of RAM, and since you have to pre-allocate 1G pages at boot time it's hard to 
>know how much to allocate to 1G pages vs 2MB pages.
>
>If all your instances have RAM in multiples of 1G then that's not a problem, of 
>course.


We went for 2M huge pages in the end (http://openstack-in-production.blogspot.fr/2015/09/ept-huge-pages-and-benchmarking.html).  I was surprised how much better this worked than THP, I guess the problem is that THP is applied after the allocation and thus pages can be mixed.

We found it difficult to get a balanced NUMA configuration. A recent blog from Mirantis suggested that the problem was to do with BIOS shadowing settings as described in https://www.mirantis.com/blog/mirantis-openstack-7-0-nfvi-deployment-guide-huge-pages/.



>
>Chris
>
>__________________________________________________________________________
>OpenStack Development Mailing List (not for usage questions)
>Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
>http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 2792 bytes
Desc: not available
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20160209/7b5f2606/attachment.bin>