drawback of transparent_hugepage=always?

[Reindl Harald]

Am 11.12.2014 um 15:04 schrieb Josh Boyer:
> On Thu, Dec 11, 2014 at 01:51:41AM +0100, Reindl Harald wrote:
>> Google around brings no clear statement
>>
>> is there any drawback of boot with "transparent_hugepage=always" which seems
>> not to be the default currently - as far as i understand it should recude
>> the overhead of memory management (especially in virtual guests) and seems
>> to have been the default for some time
>>
>> http://oracle-base.com/articles/linux/configuring-huge-pages-for-oracle-on-linux-64.php
>> at least talks about "You can check the current setting using the following
>> command, which is displaying the default value of "enabled=[always]"
>>
>> https://access.redhat.com/documentation/en-US/Red_Hat_Enterprise_Linux/6/html/Performance_Tuning_Guide/s-memory-transhuge.html
>
> We had this set to always for quite a while.  However, it isn't a win on
> all workloads.  You can run into overhead in cases where you have a
> smaller amount of RAM that is often allocated and freed frequently,
> which fragments it quite a bit and causes a lot of hugepages to be
> broken into regular pages, etc.  We discussed this a bit with the
> upstream developers and they didn't expect big wins in a typical desktop
> or laptop machine.  We switched to madvise by default so that
> applications that can benefit still do, while the overhead for
> everything else is reduced.
>
> For static workloads like databases, or virt hosts, or machines with
> larger amounts of RAM it might make sense to have it as always

thank you for feedback confirming my considerations i liked to be backed 
by people knowing more of the internals!

i am thinking about our big virtual servers with 8-16 GB RAM per VM and 
a constant workload (dbmail with large innodb and such)

maybe the following links are helpful for others too

http://www.vmware.com/files/pdf/large_pg_performance.pdf

http://kb.vmware.com/selfservice/microsites/search.do?language=en_US&cmd=displayKC&externalId=1021896