On Wed, 2006-03-08 at 21:19 -0600, Brian Hurt wrote:

> > I couldn't believe it! That's why we tried the experiment:
> > did the OS really invalidate the 'whole' caches, or did our
> > trivial increment program fail, because the memory being
> > incremented wasn't coherent?

>  I'm simplifying 
> enormously here, 

I know, so was I -- I've read the AMD64 specs.

Suppose I have a thread that creates an array
which fits roughly in the cache. Then I hand it
over to another thread to fold. The first thread
has to load stuff from RAM to populate the array, 
which slows it down.

Then it has to write the whole array back to RAM,
and the second thread reloads the whole array from
RAM into the cache.

In that case the whole cache has to be not only 
flushed to RAM, it has to be reloaded into another
cache. Pretty nasty .. on a MP machine.  

On a uniprocessor, the same CPU will do the fold,
and the cache doesn't have to be either saved
or reloaded. Yes, that's an extreme case!

-- 
John Skaller <skaller at users dot sourceforge dot net>
Async PL, Realtime software consultants
Checkout Felix: http://felix.sourceforge.net