I am concerned about DFZ routers needing to have their FIBs look at
48 bits of address for every packet sent to one of the new end-user
address blocks as allocated by ARIN, AfriNIC and now RIPE.
But this is up to 16 bits worse:
* /64 - Site needing only a single subnet.
* /60 - Site with 2-3 subnets initially.
* /56 - Site with 4-7 subnets initially.
* /52 - Site with 8-15 subnets initially.
* /48 - Site with 16+ subnets initially.
A wide TCAM can handle these address bits in a single clock-cycle,
but the high end routers CRS-1, M120 and MX960 don't use TCAM in
their FIB. TCAM chews too much power, is too expensive, has long
update times and has various other problems. These high end routers
use trie-based algorithms involving lookups into gigabytes of slow,
shared, DRAM. This can only be done some number of bits at a time,
where that number may by 5 to probably 20 or maybe 24 at the
absolute maximum.
http://www.firstpr.com.au/ip/sram-ip-forwarding/router-fib/
Consider a stream of 50 IPv6 VoIP packets a second, each bearing 20
bytes of data. Let's say it has to go through 16 DFZ routers. With
the new /64 allocation size, every packet chews the resources of 20
routers each having to work its way through 64 bits of address, with
RAM lookups. That's 128 bytes of address data to be processed,
laboriously, en-route, for a lousy 20 bytes of relatively
inconsequential VoIP call - 1/50 second.
Each VoIP call between hosts in two /64 prefixes with 16 DFZ routers
en-route involves those routers collectively working through this
many bits:
50 * 2 * 64 * 16 = 120,400 bits!!
This sounds really inefficient, compared to IPv4 in which DFZ
routers in practice need to look at 24 bits, at most, of the
destination address of IPv4 packets. Since only about 1.23% of the
advertised space is for prefixes of 21 bits or more:
http://www.firstpr.com.au/ip/host-density-per-prefix/
16 to 20 bits is probably sufficient for most packets.
- Robin
_______________________________________________
Ietf mailing list
Ietf(_at_)ietf(_dot_)org
https://www1.ietf.org/mailman/listinfo/ietf