Hi Nick, On Sun, Dec 11, 2011 at 4:27 PM, Nick Hilliard <nick@inex.ie> wrote:
On 11/12/2011 14:25, Jan Zorz @ go6.si wrote:
maybe that would be real incentive to replace bgp with something that is more efficient and scales better.
bgp isn't the problem here. In fact, BGP is doing just fine: it scales linearly according the number of prefixes, and is staying well behind moore's law, as you can run bgp calculation engines on commodity CPU.
The problem we face is dealing with lookup engines which can process ever increasing numbers of prefixes at ever faster rates.
Nick
OK, simplification might help, however, may I try to add more details?
1, It would be better not to forget the limits of the equipments installed in the current network (not negligabe percentage of the installed equipments with 500k IPv4 prefix capability only.). 2, As far as I know network convergence still needs several 10s of seconds in real life, even with faster CPU in the installed equipments. 3, YES: updating and accessing the FIB stored in the memory of line cards need time, and the time needed is hard to reduce. (No hope in "Moore low"). Plus: a 40GE, or a 100GE card needs even more complicated and costly arrangements of memory banks allowing timely access of FIB at line speed. Anyhow, do we agree that forward looking statements about the technology of the future and the policy that we accept today are two different issues? Do we agree that a clear, understandable limitation of the bad effects of the PI address space would be helpfull, and reduce the conflict between PI-funs and "clean-forwarding-table" networkers? I suggested a mesure: OK. let's allow PI allocations in exceptional cases until the number of PI allocations is below of the number of the LIR in the RIPE regions. (Other regions might accept the same, therefore it is easy to extend this policy limitaton at global level, and keeping the "pollution" low, globally). OR, do you want PI allocation for every village in certain continents? Best, Géza