Bob Glamm wrote:
>
> Also, I've tried to find results about the expected diminishing returns of
> many cpus in a regular SMP setup. I assume this occurs because of the
> cache coherance traffic on the shared bus. Are there some architectures
> on which Linux SMP works that get around this with multiple or special
> busses? Otherwise, what kind of SMP-like support is there under Linux for
> machines that aren't strictly SMP? For instance the SGI Origin series
> isn't an SMP architecture, but the ccNUMA support from the hardware and
> a special version of Irix make it SMP-like (at least according to the
> promotional literature).
>
> I know that on a high-bandwidth backplane (e.g. Challenge or Sun's
> Enterprise servers) the reasonable limit is somewhere between 16-32
> CPUs; beyond that the bus contention limits overall machine performance.
The limit on the SUN Enterprise 10000 is 64 CPUs. Sun got Oracle to
scale to 57x single
CPU performance on a 64 CPU machine. That's almost linear. This is
possible due to
a 17GB/s backplane, reading data from other CPU's 2nd level cache and
lots of other
stuff...
Regards,
Clemens Huebner
