Christian Panigl, ACOnet/UniVie +43 1 4065822-383 wrote:
Recently, following a scheduled router-maintenance on an Ebone backbone router, I had problems to communicate with NORDUnet (haven't explicitely tried PIPEX ;-) from /24 networks for about 2 hours !
- I can't beleive that it's really necessary and reasonable to kill a network (regardless of prefix-length) for two hours if it's flapping twice maybe within half an hour but not more frequently within a month !
Hello Christian, a reply from my perspective :-) Its an incentive to increase aggregation within the community without resorting to the draconian measures used by some ISPs. (You know who ;-) The smallest blocks allocated by registries now is /19 which we would only penilise for a matter of 20 minutes max (IF they reached the flap threshold). This leaves /24s as in 192/8 and maybe multi-homed customers, possibly some Last Resort numbers also. (I'm sure there are a few more valid uses also). Can I just assure you that we would certainly not kill any prefixes within the parameters you define above. Pregressive Dampening is aimed at routes which oscillate *very* regularly, i.e a few times in a matter of minutes. Default figures are 15 minutes half life of dacay penalty. The *minimum* number of withdrawals required to effect dampening is 2 with a most probably figure of 3 given that the recovery starts straight away (Half life of penalty 15 minutes remember)
- Imagine you're SW-upgrading a router and (very unlikely, as we all know ;-} detect that you have to step back ... BINGO, it's perfectly and innocently electrocuted :-(
Relatively unlikely unless it goes very wrong - and then it does have a real effect on peers doesn't it.
- I'd suggest that dampening (regardless of prefix length) shouldn't start before AT LEAST three flaps are happening in a row (let's say within half an hour).
This is about what I recommended - see above
- Dampening should lockout real network instabilities not make worse even scheduled maintenance !
How does one tell the difference ? I would say that scheduled maintenace would not normally cause potentially serious problems and if it does then it becomes a *real* network problem. ?? Any thoughts here from anyone here ?
- Besides, I know applications where it might be perfectly reasonable to announce a single *providerindependent* /24 and where it's contraproductive and politically incorrect to include it into an ISP aggregate ! A solution could be to ask Internic/RIPE to define "PI" address-ranges which can and should be excluded from the /24 hostility acts.
Toni Li's paper on ISPAC would go some way towards this and any other such schemes. Otherwise yes, for instance the root nameservers networks soon to be using /32. We would for instance exclude these. Regards -Tony