Re: Ballooning Queries

[Dor Laor <dor.laor@qumranet.com>]

Sukanto Ghosh wrote:
> On Sat, Aug 9, 2008 at 2:31 PM, Anthony Liguori <anthony@codemonkey.ws> wrote:
>   
>> Sukanto Ghosh wrote:
>>     
>>> I understand the idea behind ballooning as " it effectively increases
>>> or decreases the amount of physical memory given to the guest, with
>>> the help of the guest's native memory management algorithms". The
>>> benefit is obvious in cases where the hypervisor does a hard
>>> partitioning of the memory between the guests. The guest vm (virtual
>>> memory manager) will indirectly say that these are my least used
>>> pages, or the safest candidates for eviction.
>>>
>>> But in case of kvm, the guest memory is itself allocated and managed
>>> by the host linux vm. So, suppose if the guest evicts a lru page
>>> (which might even not actually reside in host physical memory at the
>>> time) and even the host chooses a lru page for eviction, the host
>>> decision is better in a sense that this page is definitely residing in
>>> the physical memory. So, are we gaining from using a balloon driver in
>>> the guest.
>>>
>>>       
>> What ballooning is useful for in KVM, is reducing the amount of memory a
>> guest uses by a large amount.  It's not terribly useful for shaving a few
>> dozen MBs from a guest, but it is useful for dropping a 1GB guest down to
>> 512GB.
>>
>> Since the guest tells the host what portions of it's memory it won't be
>> using, we can evict it completely from memory (it's not even swapped, it's
>> just deleted).
>>
>>     
>
> But for deleting these pages shouldn't it be ensured that the guest
> remains under the illusion that these pages are least recently used
> (may be, by modifying the guest pte for those pages, so the reference
> bits are always set for them) and the guest's page eviction algorithm
> won't touch these pages. Later on when it is possible to give back the
> guest those pages, you remove this illusion.
>   
The guest pins the ballooned pages in memory and it won't swap them.
It's simpler than hacking the LRU data.
> Another alternative will be, not to have the guest under any sort of
> illusion and give it back the pages when it wants to evict them. But
> this will defeat the purpose of having the balloon.
>
>
> Which way is it done in kvm/linux ?
>
>
>
>   
The guest is not under any illusion, it's para virtualization and the 
guest kernel is aware about
all these pages.
Host request the balloon to grow->guest allocates pages and does not 
touch them -> host frees these pages.
The opposite way for balloon deflate.