Pathlet Routing. P. Brighten Godfrey, Igor Ganichev, Scott Shenker, and Ion Stoica SIGCOMM (maurizio patrignani)

Transcription

1 Pathlet Routing P. Brighten Godfrey, Igor Ganichev, Scott Shenker, and Ion Stoica SIGCOMM 2009 (maurizio patrignani) Reti di Calcolatori di Nuova Generazione 1/40

2 challanges of interdomain routing scalability required memory is linear in the number of IP prefixes which is increasing at an increasing rate slow data plane for technologic reasons uses SRAM rather than DRAM 2/40

3 challanges of interdomain routing multipath routing useful for end users solves poor reliability or quality simply switching among alternative paths useful for network providers new service that can be sold improves route control technologies 3/40

4 reliability [F. Wang, Z. M. Mao, J. Wang, L. Gao, R. Bush 06] a sequence of UDP burst are sent to a beacon routing to the beacon changes due to alternate annoncements failure injected 4/40

5 path quality is not taken into account when BGP chooses the best path lowest latency path highest bandwidth path path the network picked for you 5/40

6 throughput by using multiple paths simultaneously the throughput of a transmission could be increased multiple paths simultaneously used 6/40

7 multipath BGP? standard BGP does not have a multipath service BGP multipath proposals do not support all BGP s routing policies expose a limited set of alternative paths make difficult to know which paths will be used increase the size of the FIB 7/40

8 Multipath Interdomain ROuting MIRO is a BGP multi-path extension MIRO router r 1 asks a (remote) MIRO router r 2 a collection of alternative paths to a destination d r 2 replies with a collection P of paths to d an IP number to be used for tunnelling r 1 may choose a path p from P tunnel traffic to the IP number of r 2 r 2 will deliver the packets to destination 8/40

9 New Internet Routing Architecture NIRA is a BGP multi-path extension each AS learns all its up-routes publishes all its up-routes at a well-known Name-to-Route Lookup Service (NRLS) each AS X builds valley-free paths to a destination AS Y by choosing an up-route to an upstream AS U asking the NRLS for up-routes from Y to U 9/40

10 new idea: pathlet routing building blocks vnode virtual node pathlet fragment of a path: a sequence of vnodes topology = vnodes + pathlets new routing protocol source routing over pathlets i.e.: source routing over a virtual topology representing routing policies 10/40

11 what do AS announce? AS announce pathlets sequence of virtual nodes pathlets do not refer to prefixes prefixes are associated to vnodes the virtual graph is a flexible way of representing policy constraints 11/40

12 what do senders select a sender concatenates its selection of pathlets into a full end-to-end source route source routing over pathlets provides the biggest possibile set of path alternatives exponentially many alternatives! 12/40

13 a simple example routers vnodes A B C D E a b c d e simple case each router has defined a single vnode routers know vnodes of their neighbors prefixes are associated with vnodes 13/40

14 routers vnodes a simple example A B C D E a b c d e pathlets are sequences of vnodes vnodes can construct one-hop pathlets to their neighbors as A, C, and D have done in the figure above each pathlet has a FID (forwarding ID) the FID identifies the pathlet in the routing table of the first node i.e.: FID=3 identifies path a b in the routing table of a FIDs uniqueness can be locally enforced 14/40

15 a simple example routers vnodes A B C D E a b c d e vnodes can build longer pathlets using shorter ones b constructs b c d e by using c d and d e and assigns FID=2 to it 15/40

16 routers vnodes a simple example A B C D E a b c d e c has to send a packet to a prefix of e c concatenates pathlets 7 and 1 to reach e c specifies 7,1 in the header of the packet to e c pops off label 7 and forwards the packet to d d pops off label 1 and forwards the packet to e e receives a packet with an empty route 16/40

17 routers vnodes a simple example A B C D E a b c d e a has to send a packet to a prefix of e a concatenates pathlets 3 and 2 to reach e a specifies 3,2 in the header of the packet to e a pops off label 3 and forwards the packet to b b pops off label 2, pushes labels 7,1 and forwards the packet to c 17/40

18 representing policies each router defines a number of internal vnodes internal vnodes are connected to vnodes of adjacent routers traffic among ordered pairs of internal vnodes pairs is either allowed or disallowed any directed path in the virtual graph is a feasible pathlet a c b d 18/40

19 example: customer-provider policy provider provider ingress from a provider egress to a provider ingress from a customer customer egress to a customer customer customer-provider policies can be enforced by setting up a suitable number of vnodes and connections 19/40

20 vnodes vnodes implement the AS policies each vnodes has a distinct routing table a globally-unique vnode identifier initially each router is configured with at least one vnode when router X accepts traffic from router Y X exposes an ingress vnode to Y Y exposes an egress vnode to X such vnodes can be different for different neighbors 20/40

21 pathlets a pathlet represents a sequence of vnodes v 1 v 2 v 3 v n if the pathlet is announced by AS X the first vnode v 1 is in X the other vnodes may be in X or in other ASes a pathlet is identified by a forwarding identifier (of FID) f FID f is unique only to the vnode v 1 other vnodes can use FID f for different pathlets FID f means that if a packet is given to v 1 it will reach v n and f will be popped off its source route 21/40

22 forwarding indentifier encoding FID values 0xxx 10xx xxxx 110x xxxx xxxx xxxx 1110 xxxx xxxx xxxx xxxx xxxx xxx xxxx xxxx xxxx xxxx xxxx xxxx bit length any AS can unilaterally pick its own encoding scheme as FIDs are opaque to other ASes 22/40

23 pathlet construction base case: two nodes only v 1 v 2 if P 1,, P n are existing pathlets v 1 can construct v 1 P 1 P n v 1 introduces a new pathlet by 1. assigning a locally unique FID f to the new pathlet 2. injecting in its forwarding plane an association of f with a next-hope rule specifying where to forward the packets an association of f with r 1,,r n, where r i is the FID of P i this is empty if the pathlet is just of two vnodes 23/40

24 packet forwarding each router has a forwarding table for each vnode v i keys are FIDs of pathlets beginning at vi values are next-hop rule push down FIDs example of routing table FID next-hop rule v 2 v 4 v 2 push down FIDs 3,4,5 3 24/40

25 packet forwarding v 1 v 2 v 3 v routing table of v 1 FID next-hop rule v 2 v 5 v 6 push down FIDs 7,1 3 incoming packet outgoing packet (to v 2 ) this is actually a label-swap!!! /40

26 packet forwarding and cheaters if a packet enters a vnode with a null source route it has to be delivered locally if the prefix is not local the packet is discarded if the first FID of a packet source route is not found in the routing table the packet is discarded as it is malformed it is impossible to cheat even if you gain knowledge of the entire virtual network there are no entries in the routing table for FIDs corresponding to not-allowed paths 26/40

27 route selection each router learns a set of pathlets each router represents a pathlet v 1 v 2 v 3 v n as a single edge v 1 v n in a routing graph edge v 1 v n is labeled with the pathlet FID when a packet has to be sent from v i to v j compute the shortest path P from v i to v j in the routing graph similar to link-state routing algorithms uses P to source-route from v i to v j 27/40

28 source-routing overhead each vnode has the map of the entire network! but this is not more than current BGP state only edge routers need to source-route, while core routers simply label-swap edge routers can cache routers to the (presumably small) set of its destinations 28/40

29 route choice the route chosen to source-route to a destination may simply consider the shortest path may take into account external information performance and availability measurements information gained via commercial route selection products information published by the target network to its customers third-party route selection services global internet weathermap 29/40

30 pathlet dissemination independent from the label-swapping and source-routing algorithms seen above trivial algorithm 1 broadcast all pathlets to the entire network similar to link-state algorithms not a policy concern, as policy is enforced at the data-plane too many pathlets transmissions trivial algorithm 2 forward only a subset of the received pathlets routers could not be reached by the information that a pathlets is not available anymore 30/40

31 path-vector pathlet dissemination a pathlet announcement contains pathlet FID sequence of traversed vnodes of the pathlet v 1 v 2 v 3 v n sequence of vnodes traversed by the announcement v a,v b,v c,,v 1 path-vector properties solves the scalability problem is a pathlet fails it will be withdrawn differences with BGP no best path is selected pathlets are not replicated for different prefixes 31/40

32 dissemination tentative rules vnodes v announces enough pathlets to form a tree from v to all destination vnodes reachable from v announces any additional pathlet that is reachable from v up to limit(d) pathlets originating at each AS with d AS-level neighbors limit(d) = 10+d 32/40

33 emulating a local transit policy an AS X implements a local transit (LT) policy when willingness to carry traffic along a route traversing X depends only on the ASes preceding and following X policy can be constructed which never extend beyond X s neighboring vnodes observe that local transit policy only refers to traffic traversing X traffic sent by X can use the fully fledged source-routing selection 33/40

34 neighbor ingress-egress vnodes for each peer Y of AS X, X could introduce two vnodes ingress vnode for Y that receives traffic from Y egress vnode for Y that sends traffic to Y connections among ingress and egress vnodes determine the allowed traffic a quadratic number of connections! in order to simplify the configuration neighbors could be classified into homogeneous sets example: providers, peer-to-peer, customers 34/40

35 effect of local transit policies LT policies allow for an exponential number of paths to a destination source destination 35/40

36 emulating unrestricted source routing unrestricted source routing (USR) disseminates the entire topology globally and lets a source AS to use any path pathlet routing can emulate USR by defining a pathlet for each directed link c a b d 36/40

37 emulating BGP X v w X introduces at least two vnodes v and w v is the ingress vnode for all neighbors of X w is the vnode containing all prefixes of X X creates pathlet v w all other ASes construct a pathlet to w along their mostpreferred available path observe that you have only one path to the destination w 37/40

38 emulating MIRO and NIRA pathlet routing easily emulates MIRO by source routing with the concatenation of two pathlets the prefix pathlet leads to r2 the postfix pathlet leads to the destination d r 1 r 2 d analogously, pathlet routing can emulate NIRA by concatenating up-routes and down-routes 38/40

39 mixed scenario LT BGP LT LT LT mixed BGP-like and local transit policies are allowed 39/40

40 pathlet routing conclusions source routing over a virtual topology formed by pathlets and vnodes highly flexible supports both local policies with small forwarding tables and many paths, and complex BGP policies challenges incentives to provide multiple paths; selecting paths; security;... 40/40