-generic memory limits

Pete Graner pgraner at canonical.com
Thu Dec 18 12:19:59 UTC 2008

Andy Whitcroft wrote:
> I have been noticing bugs being reported by i386 kernel users who are
> are installing more memory and finding that they are limited to 3GB of
> ram.  Adding more memory than this is simply not detected or used.
> This makes sense as the common behaviour for modern hardware is for
> there to be a 1GB hole from 3GB to 4GB which is used for IO space and
> the AGP aperture, with the 4th GB appearing from 4-5GB.  Currently
> our i386 -generic kernels have HIGHMEM4G enabled (but not HIGHMEM64G)
> which means we are limited to 4GB of physical address space (not memory)
> and so we can only see the first 3GB of this ram.
> With 4GB of ram no longer being a truly huge ammount of ram perhaps we
> are getting to the point where we want to allow even the i386 kernels to
> support more memory than this.  This would mean enabling HIGHMEM64G
> which in turn enables PAE so there would be some cost.
> Now of course the -server kernel does support these higher memory sizes.
> I am loathed to suggest a new flavour for i386 bigmem generic (desktop)
> installs, but the server kernel is optimised differently from the generic
> kernel based on a different usage model, it has a different default IO
> scheduler and even a different tick rate so a different responsivness.
> I guess the question is should we consider either:
> 1) enabling HIGHMEM64G and thus PAE in the i386 -generic kernel, or
> 2) creating a new -genericbigmem flavour.
> -apw

Tim has been advocating folding -server into -generic, if thats the case 
wouldn't we get that buy default?

Pete Graner		<pgraner at canonical.com>
Manager			+1.828.358.0068 Office
Ubuntu Kernel Team	+1.978-987-3125 Mobile
Canonical Ltd.		http://www.canonical.com/

More information about the kernel-team mailing list