Re: Memory priorities, was Re: 1.4.0 BETA problems

[Ignatios Souvatzis <is%jocelyn.rhein.de@localhost>]

On Thu, May 13, 1999 at 01:40:13PM +0200, Michael Boehnisch wrote:
> Matthias Scheler <tron%lyssa.owl.de@localhost> wrote:
> > In article <199905071309.PAA05634%psycho4.uni-paderborn.de@localhost>,
> >       Michael Boehnisch <billy%psycho.uni-paderborn.de@localhost> writes:
> > >   Cyberstorm PPC/68060 with 80MB RAM (8MB reserved by PPC)
> > >   Fastlane Z3          with 24MB RAM
> > >   Base memory               16MB RAM fast, 2MB chip
> > > 
> > >   When Bootblock-booting with option -n2, ...
> > 
> > You want to use "-n0" anyway. My system compiles a kernel twice as fast
> > if the memory on the Z3 Fastlane is ignored. Obviously the memory is
> > accessed in a strang order, so that the slow Z3 Fastlane memory is used
> > even if CPU board fast memory is available.
> 
> Why not including a new concept "memory priority" into the kernel?
> 
> Is this possible?  I think of a modification in the memory allocation
> functions that try to get fast (high priority) RAM first before touching
> slower (low priority) memory.

We have this, but currently only do Z2mem/fastmem differently.
Actually, I suspect you'll still see better performance without Z3
memory, especially once chuc silvers Unified Buffer Cache will be
integrated.

> Another idea could be an option to the mfs device: use slow memory for
> this device, it's still far better than hard disk access but saves the
> high priority RAM for kernel/user applications.

mfs uses _virtual_ memory. it isn't allocated any specific RAM at all; it uses
whatever the pager gives it, like any user program running.

Regards,
        -is