Re: [Xen-devel] Re: Max. PV and HVM Guests

[Keir Fraser <keir.fraser@eu.citrix.com>]

On 09/11/2009 15:17, "Pasi Kärkkäinen" <pasik@iki.fi> wrote:

>> Furthermore, while I'm not
>> Linux developer and so not familiar with how Linux calculates buffering
>> and caching, I do know that my Linux systems dynamically manage buffers
>> and caches, and when memory is reduced or some application requires a
>> larger amount of physical memory, Linux reduces the amount of data in
>> buffers and caches.
>> 
> 
> Yeah, it has to do with sizing the network buffers, caches etc..
> 
> It shouldn't _crash_, so Teo is seeing some bug I believe. But it has
> always been "best practice" to limit dom0 memory - and prevent weird
> things happening later (like "memory squeeze in netback driver").

The issue is not really kernel data like network buffers and buffer cache.
It is kernel memory metadata -- primarily the per-page info structure that
the kernel maintains. The metadata doesn't get shrunk with memory size when
ballooning out, hence it increases as a proportion of memory still assigned
to the domain. That really is significant when aggressively ballooning down
a large-memory domain.

 -- Keir