Guaranteed QoS Synthesiser (GQS) Bob Briscoe, Peter Hovell BT Research Jan 2005 GQS goals (realised) • v cheap Internet-wide statistical guarantees for inter-domain μflows • cores rarely congest but if/when they do you’re screwed • uses existing protocols, but not their architectures • RSVP (other signalling possible), DSCP, ECN • not Intserv, not Diffserv, not e2e ECN • neither path signalling nor state… • …on core AND border routers • incremental deployment • scales better as more networks join • no-one has to trust anyone else • incentives against inter-provider cheating IP routers Reservation enabled RSVP/ECN gateway ECN only GQS system arrangement Data path processing 1 Reserved flow processing 2 Policing flow entry to G 4 Meter congestion per peer 3 Bulk ECN marking G prioritised over N table of ECN fraction per previous RSVP hop aggregate reservation signalling 1 2 guaranteed 3 guaranteed 3 guaranteed (G) non-guaranteed (N) 3 3 4 1 guaranteed legend heterogeneity connectionoriented (CO) connectionless gateways GQS GQS/core CO core CO/core CO access CO/core CO assume app layer signalling (SIP) initiates out of band various QoS signalling access networks b/w broker ECN ECN ECN MPLS RSVP-TE PSTN ECN ECN ECN data plane functions: ingress GQS explanation easier if we start by assuming we have already admitted a flow reserved flow table filterspec : flowspec packet arrives Y filterspec matches reservation and passes policer set traffic class G and set ECN-capable transport 1 2 3 3 3 3 4 1 N re-mark any spoof G to BE data plane functions: ECN marking probability virtual queue mgmt (VQM) 1 Pg virtual queue q vg+qn CoSg G virtual X output ~ 99% vg priority queuing 1 traffic class? 2 CoSn CoSg qn N qg 1 Pn qn drop (or ECN marking) probability 3 3 3 line rate, X 3 4 1 data plane functions: egress reserved flow table filterspec : prev RSVP hop lookup prev hop G previous hop load prev RSVP hop : ECN fraction maintain moving ave ECN fraction clear traffic class G clear ECN-capable transport 1 2 traffic class? N GQS 3 3 3 3 4 1 Q wot, no inter-domain signalling gateway? A no 1 2 X Q wot, no per session inter-domain charging? Y 3 3 3 3 4 1 A no could by intercepting signalling, but not necessary, so if you did I’d hide it Q can’t X admit calls despite congestion in Y? transp QoS IP S1 QoS IP IP NA A emulate inter-domain policing monthly charge for bulk ECN counter at each border dead simple to meter & account IP IP IP NB QoS IP transp QoS IP R1 ND summary • it works • it’s simple • it’s cheap • it’s robust • it’s secure • it’s deployable • what else do you want? status • two implementations (2000-4) • FreeBSD & Linux • extensive simulation • full design documentation • looking for vendor take-up Guaranteed QoS Synthesiser (GQS) spare slides no time for… • probing when no active flows between gateway pair • robustness during re-routes • ECN mechanisms for incentives • ingress setting of ECT(0) & ECT(1) • VBR • preventing starvation • policy-based admission control inter-class load self-balancing Guaranteed admission threshold Non-Guaranteed TCP #flows op. point