[Date Prev][Date Next] [Chronological] [Thread] [Top]

[rtl] Re: realtime network



>>>>> "Michael" == Michael Kaufmann <root@mucosa.franken.de> writes:

 Michael> On Mon, 15 Jun 1998, Paul Koning wrote:
 >> No network can guarantee delivery of traffic to a destination with
 >> a realtime bound.  On the other hand, many networks offer delivery
 >> of

 Michael> Sorry, i can't agree with you.

 >> traffic with probability P very close to 1 with a realtime bound.

 Michael> Sure, when you cut the cable in the middle of a packet
 Michael> transfer or something like this, all your calculations are
 Michael> useless. Also when you use a reduntant network you will
 Michael> never have a absolute delivery of probability (with this
 Michael> sort of view).

All real networks have non-zero bit error rates, therefore all real
networks have P < 1 delivery probability.  How can there be any
disagreement about this?  You don't need to cut a cable to lose a
packet.  (Then again, cable damage, adapter failure, and end system
failure all need to be considered in real world designs.)

If you're building a network (especially a real time network) you
ought to know what value of P you can live with (and for what time
bound, treating any late packets as lost packets).  

For most real designs (unless they cover a large distance) Ethernet
works fine.  Do the analysis with the real properties, not the popular 
myths, and you will see this.  There are a few where it isn't
sufficient; for example, with large distances FDDI may be the right
answer.   But in general, I would start a design with Ethernet (or
Fast Ethernet) at 50% or less load, redundant if necessary, and see
where that leads.  Usually it will produce a good solution.

But if you insist on P == 1, you have set a requirement that no
network can meet.  If you claim that your network does meet it, you're
fooling yourself (and your customers).  I'm not going to budge on
that.

	paul
--- [rtl] ---
For more information on Real-Time Linux see:
http://www.rtlinux.org/