[routing-wg]On Vince's talk
Sorry I had to cut the line at the microphone. Hans Peter Holen would you care to send your comment to the list? In addition, while watching Vince's presentation, and seeing how the little IPv6 would contribute to the total sum of routes one would be seeing announced to the Internet (internal routes are each ISP's problem in this context, to some extent) I was wondering what would actually happen towards the time around which *new* IPv4 addresses would become scarce and people are forced to change the way IPv4 is used (eg, by splitting current allocations and trading those smaller chunks). Then the question might become, what would the mess look like if there is no IPv6 deployment? and does the picture Vince hinted at become any worse in the absence of IPv6 deployment, even with the less than perfect routing solutions currently available? Joao
Sorry I had to cut the line at the microphone. Hans Peter Holen would you care to send your comment to the list?
In addition, while watching Vince's presentation, and seeing how the little IPv6 would contribute to the total sum of routes one would be seeing announced to the Internet (internal routes are each ISP's problem in this context, to some extent)
Perhaps I didn't explain clearly enough... In the analysis given, IPv4 contributes approx 180K globally-visible routes while ipv6 contributes 82K (21K ASNs plus 61K more-specifics advertised for TE or other purposes). The remainder of the 350K to 530K routes are customer and infrastructure internal routes for both IPv4 and ipv6 - it is assumed that IPv4 and ipv6 contribute an equal number of such routes. Yes, such internal routes are each ISP's problem but if one believes that large ISPs will have similar numbers of them, they need to be counted for purposes of router sizing purposes. It is not clear whether GSE or any other mechanism for reducing global state would help scale internal state. In any case, even if we only consider the 180K + 82K = 260K globally-visible routes, an exponential or quadratic growth trend of in the quantity of that state will eventually cause a problem. As Geoff so eloquently stated, it isn't the raw number that is scary, it is the long-term trend. Remember that even an exponential curve can start off at a very shallow slope...
I was wondering what would actually happen towards the time around which *new* IPv4 addresses would become scarce and people are forced to change the way IPv4 is used (eg, by splitting current allocations and trading those smaller chunks).
Then the question might become, what would the mess look like if there is no IPv6 deployment? and does the picture Vince hinted at become any worse in the absence of IPv6 deployment, even with the less than perfect routing solutions currently available?
These are certainly legitimate questions to ask; the analysis done for this presentation doesn't really explore them. One might imagine that a scalable routing and addressing solution for ipv6 (or its successor) might make it more deployable and thus a migration from IPv4 more attractive. --Vince
On 4-okt-2006, at 16:10, Joao Damas wrote:
In addition, while watching Vince's presentation, and seeing how the little IPv6 would contribute to the total sum of routes one would be seeing announced to the Internet (internal routes are each ISP's problem in this context, to some extent) I was wondering what would actually happen towards the time around which *new* IPv4 addresses would become scarce and people are forced to change the way IPv4 is used (eg, by splitting current allocations and trading those smaller chunks).
Well, obviously that wouldn't be good. The question is how bad it would get? I think the current trends in routing table growth would continue more or less the same as before in the years immediately following the depletion of the free IPv4 address pools, as the underlying need remains the same and people will still find a way to get an address block, even though they probably won't get it from a RIR and it's going to be smaller than they'd like. (Although the RIRs will get address space back from people that don't need it anymore, so they'll likely be able to continue to give out small blocks. 90% of all allocations are responsible for only 10% of the address space = less than a /8 a year.)
Then the question might become, what would the mess look like if there is no IPv6 deployment? and does the picture Vince hinted at become any worse in the absence of IPv6 deployment, even with the less than perfect routing solutions currently available?
Ah, but there already is IPv6 deployment. I use it every day. A quarter or so of the traffic generated at RIPE meetings is IPv6. 0.1% of all traffic flowing over the AMS-IX is IPv6. Going out on a limb here, that figure suggests that 3% of all systems exchanging traffic over the AMS-IX are IPv6-enabled. Logic: for 97% of all systems, 100% of their traffic is IPv4, but for the other 3%, 97% would be IPv4 and 3% IPv6, for a total of 0.97 + 0.03 * 0.97 = 0,999% of all traffic being IPv4. There is probably something else that generates the majority of the 130 Mbps of IPv6 traffic on AMS-IX, but it's still an average of 130 Mbps, which is probably more than the IPv4 traffic 15 years ago. Coming back to routing: seeing recent developments in address policies and given the notion that the same people who made a mess of the IPv4 routing table will be running IPv6 at some point, I think the basic problem will remain the same. The routing table for IPv6 + IPv4 will probably be quite similar to the one for IPv4 without IPv6, at least until either the effects of massive IPv4 scarcity or retiring of IPv4 in favor of IPv6 will/would become noticeable, either of which will be a long way off. Last year the routing table increased by 16%. That gives us (in thousands): 2007 ~ 210 2008 ~ 240 2009 ~ 280 2010 ~ 325 2011 ~ 375 2012 ~ 435 2013 ~ 500 So if I were to buy an expensive router I would certainly want it to be able to carry half a million prefixes in its FIB table, and have enough RP memory for several million BGP table entries. Linecards with 512 MB RAM can be had today, and assuming that a FIB entry is less than a kilobyte seems reasonable but of course assumptions are dangerous. The problem is probably the jump to 64 bit RP CPUs because the 2 gigabyte barrier will be a problem in the forseeable future. I agree that internal routes can be a problem in large ASes, but that's just a question of proper engineering: unlike in inter-domain routing, in internal routing there are aggregation mechanisms that can be made to work effectively. I would be nice if we could create similar mechanisms for BGP. The whole notion that when someone in Nairobi starts announcing a bunch of more specifics to his neighbor, EVERY BGP router in the world must process these extra prefixes and then search through them for every packet that is forwarded, is seriously broken. As for my comments during Vince's talk: (Ugh, I think the APNIC website has an IPv6 PMTUD problem, pages take forever to load... Works better with javascript disabled.) http://www.apnic.net/news/hot-topics/index.html#ip-addressing
Hello; On Oct 5, 2006, at 8:50 AM, Iljitsch van Beijnum wrote:
On 4-okt-2006, at 16:10, Joao Damas wrote:
In addition, while watching Vince's presentation, and seeing how the little IPv6 would contribute to the total sum of routes one would be seeing announced to the Internet (internal routes are each ISP's problem in this context, to some extent) I was wondering what would actually happen towards the time around which *new* IPv4 addresses would become scarce and people are forced to change the way IPv4 is used (eg, by splitting current allocations and trading those smaller chunks).
Well, obviously that wouldn't be good. The question is how bad it would get? I think the current trends in routing table growth would continue more or less the same as before in the years immediately following the depletion of the free IPv4 address pools, as the underlying need remains the same and people will still find a way to get an address block, even though they probably won't get it from a RIR and it's going to be smaller than they'd like.
(Although the RIRs will get address space back from people that don't need it anymore, so they'll likely be able to continue to give out small blocks. 90% of all allocations are responsible for only 10% of the address space = less than a /8 a year.)
Then the question might become, what would the mess look like if there is no IPv6 deployment? and does the picture Vince hinted at become any worse in the absence of IPv6 deployment, even with the less than perfect routing solutions currently available?
Ah, but there already is IPv6 deployment. I use it every day. A quarter or so of the traffic generated at RIPE meetings is IPv6. 0.1% of all traffic flowing over the AMS-IX is IPv6. Going out on a limb here, that figure suggests that 3% of all systems exchanging traffic over the AMS-IX are IPv6-enabled. Logic: for 97% of all systems, 100% of their traffic is IPv4, but for the other 3%, 97% would be IPv4 and 3% IPv6, for a total of 0.97 + 0.03 * 0.97 = 0,999% of all traffic being IPv4. There is probably something else that generates the majority of the 130 Mbps of IPv6 traffic on AMS- IX, but it's still an average of 130 Mbps, which is probably more than the IPv4 traffic 15 years ago.
I cannot resist pointing out here that 2.2 % of the ASNs and 4.2% of the IPv4 address blocks are multicast enabled, as seen from here http://www.multicasttech.com/status/ and ~ 2% of the Abilene backbone netflow traffic is IPv4 ASM multicast, as seen here http://netflow.internet2.edu/weekly/longit/perc-multicast-octets.png Regards Marshall
Coming back to routing: seeing recent developments in address policies and given the notion that the same people who made a mess of the IPv4 routing table will be running IPv6 at some point, I think the basic problem will remain the same. The routing table for IPv6 + IPv4 will probably be quite similar to the one for IPv4 without IPv6, at least until either the effects of massive IPv4 scarcity or retiring of IPv4 in favor of IPv6 will/would become noticeable, either of which will be a long way off.
Last year the routing table increased by 16%. That gives us (in thousands):
2007 ~ 210 2008 ~ 240 2009 ~ 280 2010 ~ 325 2011 ~ 375 2012 ~ 435 2013 ~ 500
So if I were to buy an expensive router I would certainly want it to be able to carry half a million prefixes in its FIB table, and have enough RP memory for several million BGP table entries. Linecards with 512 MB RAM can be had today, and assuming that a FIB entry is less than a kilobyte seems reasonable but of course assumptions are dangerous. The problem is probably the jump to 64 bit RP CPUs because the 2 gigabyte barrier will be a problem in the forseeable future.
I agree that internal routes can be a problem in large ASes, but that's just a question of proper engineering: unlike in inter- domain routing, in internal routing there are aggregation mechanisms that can be made to work effectively.
I would be nice if we could create similar mechanisms for BGP. The whole notion that when someone in Nairobi starts announcing a bunch of more specifics to his neighbor, EVERY BGP router in the world must process these extra prefixes and then search through them for every packet that is forwarded, is seriously broken.
As for my comments during Vince's talk:
(Ugh, I think the APNIC website has an IPv6 PMTUD problem, pages take forever to load... Works better with javascript disabled.)
http://www.apnic.net/news/hot-topics/index.html#ip-addressing
participants (4)
-
Iljitsch van Beijnum -
Joao Damas -
Marshall Eubanks -
Vince Fuller