Re: [Qemu-devel] [RFC] New Migration Protocol using Visitor Interface

[Anthony Liguori <aliguori@us.ibm.com>]

On 10/05/2011 06:28 AM, Michael S. Tsirkin wrote:
> On Mon, Oct 03, 2011 at 10:00:48AM -0500, Anthony Liguori wrote:
>>> Yes, it's easy to quantify. I think the following gives us
>>> the offset before and after, so the difference is the size
>>> we seek, right?
>
> OK, Orit (Cc'd) did some research - this is a booting
> while still in grub, size probably does not get much less than that.
> windows is said to be much more aggressive in allocating memory.
>
>   start offset: 9600673
>   end offset: 9614933
>
> So we get 15K out of 9M.

So, let's do some napkin math.

Assume that it works out that most of that 15k are 4 byte integers.  If we 
assume an average name length of 6 characters, a string should be encoded in 8 
bytes.  That means for every 4 bytes, we add 8 bytes which means we're 
increasing by 200%.

That means 45k.

A 1gbit link can xmit at max, 128k in 1ms.  So that extra 30k is going to cost 
~250us in transmit time if we can get 1gbit.  A the default rate limit, it 
should cost us right around 1ms.

I guess it's liveable although with 30 network cards, I suspect it gets a heck 
of a lot worse.

Regards,

Anthony Liguori

> By the way, most of the memory here is pretty much
> all uniform I guess, because compressing it gets us:
> gzip: 1934169
> bzip2 -9: 1462551
>
> So even with aggressive compression, we probably won't be able to get
> below 1.5M for memory, two orders of magnitude above device state.
>
> Sounds convincing?
>