Actually, router hops are a problem when the packet
times (your P/C) are on the order of bit propagation
So, for 1.5 Mbps and approx. 1.5 Kbit packets,
P/C is 1 msec.
for 45 Mbps,
P/C is 33 usec.
You can see that routers for T1 and below hops introduce
a (minimum) latency on the order of prop delay.
T3 and beyond should not really be a problem.
Vadim, you are right in the context of the Sprintlink backbone,
and other DS3 and higher backbones. Of course, most large
continental NSP's have DS3 backbones. Also, these networks
are relatively small (in number of hops).
However, a lot of leaf connectivity is not at DS3 and
the more such hops you have in your path, the higher
your path delay becomes. (A separate discussion is
whether and how much this additional latency hurts.)
As an aside, the NSFnet T1 backbone introduced 6 msec.
or so of per hop latency, limiting each router to
approx. 150-160 pps. Additional NSS hops really hurt us
If a router is to switch DS3 at line speeds, it has to
process 2*45 Mbps, or 90 Mbps per adapter. At 1.5 Kbit/packet,
this is approx. 16 usec./packet. You can see why on-card
embedded systems came into vogue ...
(who kinda remembers that backbone engineer
skin thing - 1987-1992.)