Do ATM-based Exchange Points make sense anymore?

Mikael Abrahamsson swmike at swm.pp.se
Thu Aug 8 18:49:14 UTC 2002


On Thu, 8 Aug 2002, Alex Rubenstein wrote:

> I have no first-hand experience SRP/DPT.

In Sweden the highspeed exchange points have been running OC12 SRP and 
OC48 SRP for a few years. 
 
> Does it scale well to 10's, or perhaps hundreds of nodes?

Not really. The OC12 SRP has had numerous problems with overload. Fairness
works very well but still doesn't fix the simple fact that the ring is
full. The reason for it being full is that Cisco priced the OC48 cards way
too high for it to get any general acceptance and the only ISPs on the
OC48 ring are the really big ones, who usually have private peers anyway.
 
> And, is there the possibility of one member hurting the entire ring?

Yes, definately. Especially since we enabled SRR on the 24 node OC12 SRP
ring there have been two cases where a single fiber problem has forced the
whole ring into SRR mode. Not fun to get capacity/#nodes per router when
you have 24 nodes on the ring and some people put in 300-600 meg/s on the
ring normally and in SRR mode are limited to (in this case) 23 megabit.

IMHO SRP/DPT works well when you have 3-6 nodes on the ring but for 
exchange points it doesnt scale well.

Things to consider about SRP/DPT for an exchange point:

It's a ring, large number of nodes mean very high speed ring which 
increases cost for the small guys who have to get big routers and 
expensive cards even though they have little traffic.

Compare the price of a GigE interface to a dual OC48 SRP interface, plus 
that you can get GigE stuff from a lot of vendors, only one vendor (as far 
as I know) supports OC48 SRP.

No migration path. The ring uses one speed and everybody on it has to be 
at that speed. Upgrade means new ring and no intra-ring connectivity.

Personally I like the switching way of ethernet. Add a little signalling
to it that it doesnt have today (for instance, get a switch to notify
everybody when mac-addresses it has on one port goes away due to that port
going down so everybody can reset their bgp sessions) and it would work
even better than today. Anyone know of any such initiatives?

Perhaps SRP/DPT would be more viable if one could add L2 switching between
rings to it, then some people with little traffic could use OC12 and the
big guys could use higher speeds, plus you could segment the rings into
smaller rings. Since the L2 switches would have to be aware of all
signalling SRP/DPT does it should be able to handle most cases which
SRP/DPT was designed to bed good at in the first place.

-- 
Mikael Abrahamsson    email: swmike at swm.pp.se




More information about the NANOG mailing list