On Feb 7, 2011, at 8:30 AM, William Herrin wrote:
On Mon, Feb 7, 2011 at 9:25 AM, Jamie Bowden <jamie@photon.com> wrote:
It would help if we weren't shipping the routing equivalent of the pre DNS /etc/hosts all over the network (it's automated, but it's still the equivalent). There has to be a better way to handle routing information than what's currently being done.
Hi Jamie,
Consensus in the routing research arena is that it's a layer boundary problem. Layer 4/5 (TCP, various UDP-based protocols) intrudes to deeply into layer 3. Sessions are statically bound at creation to the layer 3 address. Unlike the dynamic MAC to IP bindings (with ARP) the TCP to IP bindings can't change during the potentially long-lived session. Thus route proliferation is needed to maintain them.
Much better routing protocols are possible, but you first either have to break layer 3 in half (with a dynamic binding between the two halves that renders the lower half inaccessible to layer 4) or you have to redesign TCP with dynamic bindings to the layer 3 address. Ideas like LISP take the former approach. Ideas like SCTP and Multipath TCP take the latter. The deployment prospects are not promising.
Modest improvements like FIB compression are in the pipeline for DFZ routing, but don't expect any earth shattering improvements.
On the other hand, when we can deprecate global routing of IPv4, we will see an earth shattering improvement as the current 10:1 prefix to provider ratio (300,000 prefixes for ~30,000 active ASNs) drops to something more like 2:1 in IPv6 due to providers not having to constantly run back to the RIR for additional slow-start allocations. Owen