[GRLUG] large systems performance

[Michael Mol]

On Nov 16, 2007 1:15 AM, Bob Kline <bob.kline at gmail.com> wrote:
> On Nov 16, 2007 12:57 AM, Tim Schmidt <timschmidt at gmail.com> wrote:
> > On Nov 16, 2007 12:45 AM, Bob Kline <bob.kline at gmail.com> wrote:
> > > If running Linux,  which  version?
> So would you guess that even more
> processors would help?  With your
> load profile that is?
>
> It's not impossible that even a desktop
> will have hundreds of CPUs one day.
> I have no idea whether Linux could
> support that.  But with the concept of
> one process per processor, without much
> swapping while in typical use,    one
> would see even more improvement.
> Swapping always creates overhead,  so
> the best way to reduce overhead is to not
> swap where possible.  Including letting
> other jobs wait in some circumstances
> if they are not real time.

The O(1) scheduler brought in with the start of the 2.6 kernel series
means that you won't see a performance degradation with the addition
of more cores, if you're running independent tasks.  You might see a
limit to performance improvements if your software isn't written
right, though; It's deceptively easy to write bad multi-threaded code
that hangs up on one or two locks.

-- 
:wq