90 likes | 169 Views
RSVP/Mobility Interactions. Michael Thomas mat@cisco.com. Overview. #include <std.disclaimers> Draft is about 2 years old – many dead brain cells I’m hopelessly behind on mailing list Quick MIP overview Localization Problems with RSVP Cross Realm Policy Questions. MIP Overview.
E N D
RSVP/Mobility Interactions Michael Thomas mat@cisco.com
Overview • #include <std.disclaimers> • Draft is about 2 years old – many dead brain cells • I’m hopelessly behind on mailing list • Quick MIP overview • Localization Problems with RSVP • Cross Realm Policy Questions
MIP Overview • First: MIP != Mobility • I’ll talk about MIP because it’s illustrative, other kinds of mobility need to be considered too • MIP is a protocol which allows a mobile node to wander away from it’s “attachment” point (ie, the thing that subtends its prefix) through the use of L3 tunneling • Home Agent is that attachment point for both forward and reverse tunnels • IPv4 tunnels to a “Foreign Agent” on the remote network • IPv6 colocates FA into the mobile node • IPv6 also allows direct mobile node/correspondent node communication – makes signaling much more complex • Various schemes for localized tunneling for fast handovers
Typical MIP Packet Exchange <- Corr Home -> Home Agent CN <-- Remote Access AGG data AR1 AR2
Real Time == Real Pain • MIP/RSVP should work adequately where loss is tolerable • Speed of light is not our friend • Restoration of reservations after move requires round trip(s) from MN->CN for MN rx • Distance from MN->CN is arbitrary • 100ms is an eternity for voice traffic
Localization • What we really need for RT traffic is to localize signaling during handovers • That is, topology changes need to heal “close” to the mobile node and not touch far nodes not involved • For RSVP this is problematic as a change of CoA requires brand new reservations in both directions • Would be nice to not tie reservation to src address • Interesting question for MIPv6 of whether to use HoA or CoA
More Localization • Worse, even if you could reuse the same reservation, no way to heal the CN->MN reservation without signaling to the CN • MN->CN could be heal with a PATH from MN to a local join point • No such thing as naked RESV’s and the join point is clueless that it should issue a new PATH • Lastly: route flap damping protection is at odds with fast convergence • I believe that RSVP sez to wait 100’s of ms – not good • How does an interior node know that it’s a move rather than a possibly flapping route?
Cross Realm Policy • Everybody’s Problem; Nobody’s Solution • RSVP Identity (2752) is pretty “enterprise” oriented (eg single domain) • Not clear whether policy is meant to be e2e or hop-by-hop • Both have tradeoffs • Worse: it’s not clear that it matters all the time • In many cases, there will be *no* cross realm relationship. • In that case, RSVP might needlessly cease to work for policy reasons • The net effect is that local loop QoS is pooched for no good reason
Conclusion • RSVP has much to recommend it • Much work over many, many years • If I hear one more context-free “RSVP doesn’t scale”, I’m going to smack them in the gob • It seems quite possible that many of these issues can be resolved within RSVP itself • Some issues are more architectural in nature • End to end vs end to middle, etc