Havard Eidnes wrote: Thank you for good examples.
Something that right now confuses *me* is: If I understand this correctly, the 'default-free zone' is meant to be kept below 1000 routes, so routers can be fast.
The hard limit is, IMO, 8192.
Past history has told us that at least some technologists have failed miserably in predicting whether fast routers can be built which handle largish routing tables.
And, I happened to be a technologist who know how to build a fast (Pbps or more) router even before no one have the needs. So, the only problem on fast routers is their prices. Limiting the number of global routing table entries helps a lot. With no loops in the forwarding path of routers, the limiting factor is latency of routing table look up. While parallelism is a solution, it does not reduce the routing table size. So, to reduce the prices of routers, it is better to reduce the routing table size. Their are poeple who use science to prove that it is impossible to make flying machines. Their are people who use science to make flying machines more efficient and less expensive.
If I recall correctly, an argument similar to the above one was one of the arguments the ATM proponents used in their day: you need a short header to do fast lookups, and "fast IP routers cannot be built". History since then has at least told me that this was not entirely accurate.
At that days, my theory explained why ATM is about 10 times slower than IP. The theory is simply that, given 28 bit address with hierarchy (VPI/VCI or more), it is almost as complex as processing 32 bit IP. So, it takes about 10 times more amount of work to process 48B chunks than 500B-in-average chunks. Of course, there are other factors obscuring the essential difference.
Therefore, I tend to view the above justification and limits with a healthy dose of scepticism.
So, you can still insist that my theory is incorrect and ATM is 8192 or more times faster than IP, while I can keep saying ATM is about 10 times slower than IP with reasons. I can also say 8192 is large enough for the number of global routing table entries with reasons. You can alsy say IPv6 is no good because 128 bit address is not long enough. But, it's your limitation, not mine. Masataka Ohta PS Note that there are people who analize that large routing table increases the time for BGP convergence.