Re: 64 bit guest much faster ?

[Michael Tokarev <mjt@tls.msk.ru>]

Avi Kivity wrote:
> On 10/26/2009 10:58 AM, Gerd Hoffmann wrote:
>> On 10/23/09 17:54, Stefan wrote:
>>>
>>> Hello,
>>>
>>> I have a simple question (sorry I'm a kvm beginner):
>>> Is it right that a 64bit guest (8 CPUs, 16GB) is
>>> much faster than a 32bit guest (8 CPUs, 16GB PAE).
>>                                           ^^^^
>> Yes.  With *that* much memory the 32bit guest struggles with address 
>> space limitations (32bit -> 4G), whereas the 64bit guest doesn't.
>>
>> With up to 1G you shouldn't see a noticable difference.  But the more 
>> highmem the 32bit guest uses the higher is the penalty.  Especially 
>> without ept/npt as every kmap() of a high page is a roundtrip to the 
>> hypervisor then.
> 
> Oh yes, without ept/npt the slowdown should indeed be significant with 
> this much memory.

How it is with 4Gb guest/mem without PAE (I mean, with CONFIG_HIGHMEM_4G=y)?
Or even 2Gb?  In case of npt or without.

Can we construct a sort of a table of expected slowdowns (not in numbers
but just in terms "significant", "minor" etc) of running <4Gb or >4Gb
(and <1Gb and >1Gb if that makes significant diffencece) 32bit guests
with and without npt and 64bit guests, please?  I guess it's quite
interesting to many users.

 From the above it looks like it's better to run 64bit kernel in the 32bit
guest in these situations too.

I haven't measured it, just because it never occured to me that there
MAY be any difference.  But I've only non-npt hardware here at the
moment.

Thanks!

/mjt