RE: [lir-wg] Discussion about RIPE-261
Carlos,
Carlos Morgado wrote: Other people have already mentioned on the list memory/cpu isn't an issue on this day and age. You definitly don't need a c12k to carry a full routing table today.
This is not what vendors are saying. I presented just after Jeff Doyle from Juniper 3 weeks ago and he was very specific about this. I extracted that one slide here: http://arneill-py.sacramento.ca.us/ipv6mh/j.ppt Let's face it: I'm sure that Juniper would love to keep selling big routers. They are telling us they can't guarantee they will keep up increasing memory and CPU, so we do have an issue with memory/cpu. Michel.
Michel Py wrote:
Carlos,
Carlos Morgado wrote: Other people have already mentioned on the list memory/cpu isn't an issue on this day and age. You definitly don't need a c12k to carry a full routing table today.
This is not what vendors are saying. I presented just after Jeff Doyle from Juniper 3 weeks ago and he was very specific about this. I extracted that one slide here: http://arneill-py.sacramento.ca.us/ipv6mh/j.ppt
Let's face it: I'm sure that Juniper would love to keep selling big routers. They are telling us they can't guarantee they will keep up increasing memory and CPU, so we do have an issue with memory/cpu.
Haha!! A little bit a weak argument, isn't it? What if a router vendor would tell you he couldn't guarantee to keep up with PPS performance if we are going to introduce 10Gig Ethernet??? No way. Also the Internet doesn't get more instable with more prefixes. (In)stability is dependent on the number and stability of links carrying the prefixes. So you have to measure instablity in percent of the of the entire prefix pool. I don't think we are doing worse than ten years ago despite an increase of 12'000 percent in prefixes. Actually the transit network seems to be very stable. There are some problems with leaf customers who've got BGP but are imcompetent at it and shoot themselfes in the foot. But that is a competence problem. (Maybe we a BGP drivers license?) Convergence time is not dependent on the number of prefixes either. It's dependent on the min, mean and max radius/diameter of the network (in terms of hops which carry the full table modulo the intra-AS BGP reflector mesh). Ok, maybe some vendors implementation gets instable with more prefixes but that's a bug and thus an engineering problem. -- Andre
Hi, On Tue, May 27, 2003 at 07:56:15PM +0200, Andre Oppermann wrote:
Also the Internet doesn't get more instable with more prefixes.
Can you back that claim by some facts or research studies? I know that people at a couple of universities are researching into this right now, and currently the primary assumption is "yes, it does" (because the sheer size of the lists is longer, routers *need* more time to process them, delaying convergence, plus every now and then you hit a boundary that causes BGP flaps due to out-of-memory and/or necessary router upgrades). Gert Doering -- NetMaster -- Total number of prefixes smaller than registry allocations: 54837 (54495) SpaceNet AG Mail: netmaster@Space.Net Joseph-Dollinger-Bogen 14 Tel : +49-89-32356-0 80807 Muenchen Fax : +49-89-32356-299
Gert Doering wrote:
Hi,
On Tue, May 27, 2003 at 07:56:15PM +0200, Andre Oppermann wrote:
Also the Internet doesn't get more instable with more prefixes.
Can you back that claim by some facts or research studies?
No, but by mathematical logic. Can you claim otherwise?
I know that people at a couple of universities are researching into this right now, and currently the primary assumption is "yes, it does" (because the sheer size of the lists is longer, routers *need* more time to process them, delaying convergence, plus every now and then you hit a boundary that causes BGP flaps due to out-of-memory and/or necessary router upgrades).
The "assumption" that it hurts does not count because I "assume" that it does not hurt. Please provide scientific facts as you have asked me to. Longer lists are fine. You just need more processing power. That's why we have Moore's law. If I look at the advantages (see IEEE and ACM magazines) forwarding table implementations in soft- as well as hardware which have been made because of higher PPS needs/wants I don't have any worries that all these well-paid bright people come up with some optimisations which will take us well beyond one million active prefixes and triple the path. Simple engineering again. If they can do it with PPS and wirespeed they should be able to do it with BGP processing as well. A router that has a bug or runs out of memory will always happen. There is no silver bullet against that. It's the percentage of the entire prefix/path base that matters. If you are looking at absolute number you will definatly see an increase, no doubt. If I remember correctly a Juniper has got a Mobile-PII 450MHz CPU as control processor. Hardly on par with today's available processing power. You could have said your same stuff five years ago when we had 70'000 prefixes. But now we have 127'000 prefixes it still works fine as before. My "assumptions" are based on the mathematical properties of the BGP distance vector routing protocol. Of course the larger the mean AS distance grows, the more "instable" in itself it gets. But that is normal Internet growth. Or do you want stop the Internet to grow? Is it large enough? Do we need more ISPs? We see a very common occurence here. The moment an open and very competitive market has matured, the (remaining) players start (whether implicit or explicit) to hinder new entrancies into the market. This is either done by denial of (direct) access or policy barriers. Do we have enough ISPs? Do we have enough parties with routable IP address space? All the "technical" arguments I've heard so far simply define the status quo (currently available and deployed routers developed three to five years ago) as end of all means. From which the rational of no more space for significant growth (prefixes/ASs) is derived from. This is wrong. If we would have applied the same rational at a time when a T3 backbone was gigantic, then we'd never been here to have this conversation today. Do we? -- Andre Oppermann
Other people have already mentioned on the list memory/cpu isn't an issue on this day and age. You definitly don't need a c12k to carry a full routing table today.
This is not what vendors are saying. I presented just after Jeff Doyle from Juniper 3 weeks ago and he was very specific about this. I extracted that one slide here: http://arneill-py.sacramento.ca.us/ipv6mh/j.ppt
Let's face it: I'm sure that Juniper would love to keep selling big routers. They are telling us they can't guarantee they will keep up increasing memory and CPU, so we do have an issue with memory/cpu.
Other vendors are saying different things. If we stopped overloading BGP with all these other features it would most likely also last a lot longer. - kurtis -
participants (4)
-
Andre Oppermann -
Gert Doering -
Kurt Erik Lindqvist -
Michel Py