🚀 go-pugleaf

From article <2936@bloom-beacon.MIT.EDU>, by martillo@athena.mit.edu (Yakim Martillo):
> As far as I can tell, bandwidth like memory just gets cheaper and more
> available.  Futurebus is not too far off and gigabit bandwidth network
> backbones will probably be available in the near future, which is good
> because we will be seeing customers that desire network transparent 3d
> high resolution real-time color graphics.  Favoring a system because
> it might work better over 9600 bps serial links than another seems to
> evince a serious confusion about goals.  

You won't see *effective* bandwidth anywhere near a gigabit any time
soon, especially for small messages like most of those X shuffles
around.  And it's not the physical bandwidth of the network
dominates the cost, it's the CPU time spent packing up messages, doing
traps, changing (heavy weight) contexts, fielding interrupts, copying
(or mapping) messages into user address space, etc.

> Now I have the impression
> Sun unlike DEC is not doing much research in faster back-planes so that
> maybe such goals are reasonable for Sun.  As for context switching,
> it would seem that situations could easily exist where NeWS could
> actually require more context switching than X because in addition
> to context switching between client and server there would be context
> switching within the server as well.

But context switching within the server is (well, I'm assuming) much
faster than between client and server.  A context switch between two
lightweight processes that share memory amounts to little more than
saving one set of registers and loading another.  This should be a
couple of orders of magnitude faster than a full address-space switch.
-- 

Mike Coffin				mike@arizona.edu
Univ. of Ariz. Dept. of Comp. Sci.	{allegra,cmcl2,ihnp4}!arizona!mike
Tucson, AZ  85721			(602)621-4252