Oliver Bartels;
Do you know that the current Internet backbone is operating with parallel 10Gs?
Someone may stretch the term "backbone". There is no single one.
There is 10G backbone with paralle routes operated at least one ISP in Japan.
So, large memory costs.
Not really.
Todays price, single pcs., IDT/-100 at Arrow for 64K*72 TCAM: $155 from stock.
Instllation of chips, especially those operating at highest possible speed, costs a lot more than that. If it is not on chip, you lose, a lot.
If someone operates a major backbone, this is no real cost factor compared to leased lines and wavelengths ...
If country backbone is as slow as 10G, maybe. However... In Japan, today, several Kms of dark fiber costs tens of USD/month.
Even if you multiply this with five for sales droid financing ...
Internet grows faster than any sales droid can understand.
Fast synchronous static memory is even *much* less, e.g. few $'s per Multi-Mbit chip.
As it is off the chip, you automarically lose.
Do you also know that access speed of memory (including but not limited to TCAM) degrades proportional to log or sqrt of the number of entries?
Again, it depends ...
It depends? I'm saying how it depends. It, of course, is not a problem, if and only if the routing table size has hard limit.
Both modern routers and modern CPUs are highly pipelined, which means there is some performance loss if TCAM or primary cache miss occurs.
A *cached* router is a good-weather-only product which would die e.g. if a DDoS or SQL Slammer is on the road.
That's why the global routing table must be small.
Thus modern backbone routers *do not use route caches*.
I know. Who suggested route cache? I didn't.
Secondary or third level cache of modern CPUS often have millions of entries and constructed with static RAM.
Which again tells us that 10K is not a real limit for modern routers, as well as 100K is not.
It merely means that it costs, especially when you do off chip interleaving with a lot of wiring.
Noone would buy a *new* backbone router which isn't capable to handle a n*100K table even if there is "rain and snow" inside the network ...
That was a valid argument for IPv4. Masataka Ohta