Re: [PATCH 0/4] ksm - dynamic page sharing driver for linux

[Izik Eidus <ieidus@redhat.com>]

Andrew Morton wrote:
> On Tue, 11 Nov 2008 21:18:23 +0200
> Izik Eidus <ieidus@redhat.com> wrote:
>
>   
>>> hm.
>>>
>>> There has been the occasional discussion about idenfifying all-zeroes
>>> pages and scavenging them, repointing them at the zero page.  Could
>>> this infrastructure be used for that?  (And how much would we gain from
>>> it?)
>>>
>>> [I'm looking for reasons why this is more than a muck-up-the-vm-for-kvm
>>> thing here ;) ]
>>>       
>
> ^^ this?
>
>   
>> KSM is separate driver , it doesn't change anything in the VM but adding 
>> two helper functions.
>>     
>
> What, you mean I should actually read the code?   Oh well, OK.
>   
Andrea i think what is happening here is my fault
i will try to give here much more information about KSM:
first the bad things:
KSM shared pages are right now (we have patch that can change it but we 
want to wait with it) unswappable
this mean that the entire memory of the guest is swappable but the pages 
that are shared are not.
(when the pages are splited back by COW they become anonymous again with 
the help of do_wp_page()
the reason that the pages are not swappable is beacuse the way the Linux 
Rmap is working, this not allow us to create nonlinear anonymous pages
(we dont want to use nonlinear vma for kvm, as it will make swapping for 
kvm very slow)
the reason that ksm pages need to have nonlinear reverse mapping is that 
for one guest identical page can be found in whole diffrent offset than 
other guest have it
(this is from the userspace VM point of view)

the rest is quite simple:
it is walking over the entire guest memory (or only some of it) and scan 
for identical pages using hash table
it merge the pages into one single write protected page

numbers for ksm is something that i have just for desktops and just the 
numbers i gave you
what is do know is:
big overcommit like 300% is possible just when you take into account 
that some of the guest memory will be free
we are sharing mostly the DLLs/ KERNEL / ZERO pages, for the DLLS and 
KERNEL PAGEs this pages likely will never break
but ZERO pages will be break when windows will allocate them and will 
come back when windows will free the memory.
(i wouldnt suggest 300% overcommit for servers workload, beacuse you can 
end up swapping in that case,
but for desktops after runing in production and passed some seiroes qa 
tress tests it seems like 300% is a real number that can be use)

i just ran test on two fedora 8 guests and got that results (using GNOME 
in both of them)
 9959 root      15   0  730m 537m 281m S    8  3.4   0:44.28 
kvm                                                                            

 9956 root      15   0  730m 537m 246m S    4  3.4   0:41.43 kvm
as you can see the physical sharing was 281mb and 246mb (kernel pages 
are counted as shared)
there is small lie in this numbers beacuse pages that was shared across 
two guests and was splited by writing from guest number 1 will still 
have 1 refernce count to it
and will still be kernel page (untill the other guest (num 2) will write 
to it as well)


anyway i am willing to make much better testing or everything that 
needed for this patchs to be merged.
(just tell me what and i will do it)

beside that you should know that patch 4 is not a must, it is just nice 
optimization...

thanks.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>