[fedora-virt] swap space on virtual machines

[Dor Laor]

On 07/25/2009 09:02 PM, Richard W.M. Jones wrote:
> On Fri, Jul 24, 2009 at 12:39:54PM -0400, Rich Mahn wrote:
>> How big should the swap parititons be on virtual machines
>> under qemu, qemu/kvm?
>>
>> It seems to me that if the VM actually needs swap space, it
>> would be more efficient to allocate more virual memory to it.
>> I know it's possible to allocate more virtual
>> memory than there is physical memory.  Intuitively (which
>> isn't always right) it would seem that allocating more
>> memory is more efficient than letting the VM use swap space.
>> It would seem that the host machine would have to either use
>> physical memory or swap space if it's required.
>
> A real interesting question.
>
> One place I think you're wrong is the assumption that adding more
> memory to a VM is better than having the VM use a swap disk.  The
> reason would be that the VM's memory manager will assume that the
> [from its point of view] physical memory will be much faster than
> swap, and so will arrange memory vs swap use accordingly.  But this
> assumption isn't true, this so-called physical memory is really just
> as slow as swap!
>
> You say:
>
>> Compared to the VM using a swap space that is a virtual disk space
>> that goes to real disk.
>
> but actually virtualized block I/O isn't so bad if you use the correct
> (virtio) drivers.
>

I think several terms are mixed together here:

The guest is given certain amount of ram to play with.
Various hypervisors virtualize this guest ram differently. Regardless of 
the hypervisor. The guest does assumes it is his ram.
Even if you allocate large amount of ram to the guest, it might swap 
since the guest runs processes that uses virtual memory, and thus might 
exceed the allocated ram to the guest. Guest swapping is a reasonable 
scenario that should be allowed and supported.
When using a fast virtual block device like virtio, it should even be 
almost as fast as native swapping (which is always slower then ram..).

KVM hypervisor runs a VM as standard process. As such, it allocate 
virtual userspace memory to the VM as ram. We can overcommit the host 
ram and even swap it from the host.
Host swapping works nicely but might really slow down things since both 
host & guest will swap out/in/out the same pages.


> I think the only way you'll really answer this is to conduct tests
> based on your specific workload.  I've only done this with Xen, where
> this doesn't apply because Xen guest memory _is_ physical memory.
>
> By the way, Kernel Shared Memory (KSM, [1]) complicates everything...

But make things better ;)

>
> Rich.
>
> [1] http://lwn.net/Articles/306704/
>

So bottom line, if you have free ram, give it to the guest - why save 
it? Even large guest can swap, determined by the load.

Dor