a question about the economics of peering

Fri Nov 30 23:43:11 UTC 2001

On Fri, 30 Nov 2001, batz wrote:

> 
> Indeed. The notion of "hopcount" is objectively meaningless. 
> You can count hops by physical wires, data segments, network hops, 
> AS's or combinations of each (MPLS, GRE, PPP etc). 

Actually, not.  The hops which "count" are places where traffic flows are
stochastically multiplexed.  Since every such place has to have buffers of
size comparable to end-to-end RTT * bandwidth (to give enough time for TCP
flow control to react), the maximal packet delay is Nhops*RTT.

Note this does not say what the median variable latency (fixed latency is
the signal propagation time) is going to be because it depends on
probability of multiple congestions along the path.  If congestion events
are independent, the result isn't that bad (with long-lived congestion
probability of 50% the median variable latency is 2*RTT).  Unfortunately,
this assumption is not likely to be realistic. I.e. in the "aggregation
tree" networks (like the most modern backbones which aggregate outgoing
exterior traffic at multiple points, where every aggregation point reduces
the total egress bandwith), the congestion is likely to occure
simultaneously at many aggregation points along the path.

The way to combat variable latency is to reduce the number of hops. This
effectively means either creation of meshes of fixed bit-rate circuits
(which is way way way ugly because of routing scalability issues), or
reducing the number of aggregators along the path by reducing the number
of routers.  The simplest way to do that is to replace router clusters
with large parallel routers featuring non-blocking switching fabrics.

Note that non-CBR virtual-circuit techniques (tunnels, MPLS, non-CBR ATM,
FR, X.25 etc) don't do anything to reduce the _real_ hopcount; they simply
obscure it.

--vadim