On Mon, Apr 15, 2013 at 3:10 PM, Pascal J. Bourguignon <
[email protected]> wrote:

> I think that one place where "light cone" considerations are involved is
> with caches in multi-processor systems. If all processors could have

instantaneous knowledge of what the views of the other processors are
> about memory, there wouldn't be any cache coherence problem.  But light
> speed, or information transmission speed is not infinite, hence the
> appearance of light cones or "light cones"-like phenomena.


Many people seem to jump from one extremism to another - from
"instantaneous transfer" to "unbounded delay" - without seriously
considering the useful middle (predictable, bounded delay). The middle has
many models (including cellular automata) and is capable of supporting
synchronous/real-time distributed systems. It's also where you'll find
light cones... and many interesting, efficient synchronization patterns.

Interestingly, cache coherence is not a problem if your programming model
*doesn't* assume instantaneous transfer, i.e. because you'd end up
explicitly modeling the delays and thus managing the distinct views in a
formal manner - using distinct locations in memory, and thus distinct cache
lines. (I believe this contributes to the success of modeling
multi-processor systems as distributed systems.)

Regards,

Dave
_______________________________________________
fonc mailing list
[email protected]
http://vpri.org/mailman/listinfo/fonc

Reply via email to