Subject: Re: Multiprocessor with NetBSD ?
To: mike stone <firstname.lastname@example.org>
From: Brett Lymn <email@example.com>
Date: 06/05/2001 23:19:43
According to mike stone:
>shows that for constant B greater than 0, speedup is far from linear
>with respect to N.
Mmmm cool ascii graph - I was impressed :-)
>. for tightly-linked
>MP systems, sharing the memory bus is an inescapable serial dependency,
>so you're pretty much stuck with sub-linear speedup.
Ahh but are you not forgetting that there is no longer a linear
connection between the processor and the main memory? There are at
least a couple of levels of cache in there that reduce the dependency
on going to main memory, this means that there will be lower memory
bandwidth utilisation than you may think and the utilisation may be
unpredictable depending on the processor work mix (ie if the processor
manages to run entirely out of cache you will get a much more linear
scaling, unlikely that it will happen though). The other thing that
is helping to linearise the curve is that larger MP systems do not
have a single procesor to memory bus. They tend to have what is, in
effect, a large crossbar switch between the processors and memory so
access to memory is not shared between all the processors.
Brett Lymn, Computer Systems Administrator, BAE SYSTEMS