On Mon, 04 Jul 2011 17:27:24 +1000 Karl Auer <ka...@biplane.com.au> wrote:
> On Mon, 2011-07-04 at 16:34 +0930, Mark Smith wrote: > > I missed this earlier - > > > > > Note that the subnet router anycast address doesn't > > > even ensure that the "nearest" router is contacted, which would arguably > > > have been useful, because the current spec deliberately builds in a > > > small random delay in responding, to avoid network congestion. > > > > > > > Are you referring to rfc4861, section 7.2.7? The thing that makes the > > first/quickest response "stick" is the Override flag being set to 0 - > > That's correct, the first/quickest response will stick. The thing that > stops that working in the sense of getting to the "nearest" router is > this: > > "Nodes that have an anycast address assigned to an interface treat > them exactly the same as if they were unicast addresses with two > exceptions. First, Neighbor Advertisements sent in response to a > Neighbor Solicitation SHOULD be delayed by a random time between 0 > and MAX_ANYCAST_DELAY_TIME to reduce the probability of network > congestion. Second, the Override flag in Neighbor Advertisements > SHOULD be set to 0, so that when multiple advertisements are > received, the first received advertisement is used rather than the > most recently received advertisement." > > Because of that random delay, the nearest router may not be the one that > "wins". It's a SHOULD, so I suppose implementations could choose to > ignore that. Unless there are a LOT of routers on the link, network > congestion is most unlikely to be a problem, and leaving out the delay > would help the nearest router "win". > Ok, so it seems a bit of variation has been added in so that the anycast "clients" are likely to be spread more across all the available anycast servers. This is probably not a bad thing. I'd been assuming it was working like PPPoE does, where the first BRAS to respond is (usually) the one chosen by the client. However I have seen a scenario where out of a group of equal capacity and performance BRASes, when one of them was placed one extra layer 2 hop away (i.e. a switch inserted between it and the switch the others were attached to), quite surprisingly customer connections drifted away from it over time, despite it being less loaded, and therefore in theory should be the fastest to respond. This extra switch hop was only adding microseconds of latency to customer's ADSL connections that had latencies in the order of 9 to 20ms. Adding some variation into the PADOs as per the method above would probably have lessened if not eliminated this customer session drift. > But I simply see no point to the subnet router anycast address. Is there > any use case that could not be equally well dealt with using > all-routers-on-link or normal routed anycast? > > Regards, K. > > -- > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ > Karl Auer (ka...@biplane.com.au) +61-2-64957160 (h) > http://www.biplane.com.au/kauer/ +61-428-957160 (mob) > > GPG fingerprint: DA41 51B1 1481 16E1 F7E2 B2E9 3007 14ED 5736 F687 > Old fingerprint: B386 7819 B227 2961 8301 C5A9 2EBC 754B CD97 0156 -------------------------------------------------------------------- IETF IPv6 working group mailing list ipv6@ietf.org Administrative Requests: https://www.ietf.org/mailman/listinfo/ipv6 --------------------------------------------------------------------
