On Sun, Nov 08, 1998 at 12:14:23PM +0100, Paul Komarek wrote:
> On 8 Nov 1998, Andi Kleen wrote:
> > break. It might work if you can guarantee that no single user process
> > ever grows bigger than 1G-epsilon, but that is not tested and a bit
> > fragile.
> >
> > -Andi
>
> I'm working with my thesis advisor to pick a good smp configuration for
> our research. We are looking at a machine with 4xP-II Xeon 450s, and 4 GB
> of ram. I've been watching this thread for a while, and am getting
> nervous that we won't be able to use the 4 GB in the way we would like.
> So I'm asking a few clarifying questions:
>
> 1) Would we have access to all 4 GB (through one or several processes)?
No. Linux currently needs all physical memory mapped + some virtual address
space for the process and possible IO mappings. With 4GB physical there
would be no virtual space left. You can probably go upto 3GB if you make
sure that no processes uses more than 1GB-epsilon, or 2GB where one process
can use 2GB-epsilon etc.
> 3) Does WinNT have similar limitations?
I believe there is a version of NT that supports 3GB physical + a very very
ugly API (similar to old EMS) to access more memory. AFAIK newer versions
of Solaris/x86 or Unixware have better support for that much memory. In Linux
the current solution is to switch to an 64bit architecture (Alpha or UltraSparc)
Changing the "all physical memory has to be mapped" requirement would be
a fundamental architectural change that is unlikely in the near future.
-Andi
