Introduction to Intra-Domain Routing

Transcription

1 Introduction to Intra-Domain Routing Stefano Vissicchio UCL Computer Science COMP3

2 Agenda We delve into network layer s main functionality. Setting Context Routing players. Intra-domain routing problem definition

3 Let s consider a network with several LANs host host DNS host host... DNS.../3 router..3.9 router router /4 3

4 Several LANs can belong to a single organization Domain (e.g., UCL) host host DNS host host... DNS.../3 router..3.9 router router /4 Internet 4

5 Collecting evidence from the real world Traceroute exposes real Internet paths Unix: traceroute <destination> Windows: tracert <destination> It displays all hops on the path between host where launched and <destination> it sends a sequence of carefully constructed packets that expire after hop, hops, each of those packets elicits ICMP error from router that many hops from sender 5

6 A traceroute from UCL to Cambridge traceroute to (8.3..), 64 hops max, 4 byte packets cisco ( ).37 ms.3 ms.36 ms ( ).483 ms.348 ms.487 ms (8.4..).486 ms.34 ms.36 ms (8.4..6).486 ms.474 ms.363 ms 5 ulcc-gsr.lmn.net.uk ( ).485 ms.346 ms.36 ms 6 london-bar.ja.net ( ).485 ms.47 ms.488 ms 7 po-.lond-scr.ja.net ( ).735 ms.7 ms.6 ms 8 po-.cambridge-bar.ja.net ( ) 5.3 ms ms ms 9 route-enet-3.cam.ac.uk ( ) 4.98 ms 4.84 ms 4.86 ms route-cent-3.cam.ac.uk ( ) ms ms 4.86 ms 6

7 A traceroute from UCL to Cambridge traceroute to (8.3..), 64 hops max, 4 byte packets cisco ( ).37 ms.3 ms.36 ms ( ).483 ms.348 ms.487 ms (8.4..).486 ms.34 ms.36 ms (8.4..6).486 ms.474 ms.363 ms 5 ulcc-gsr.lmn.net.uk ( ).485 ms.346 ms.36 ms 6 london-bar.ja.net ( ).485 ms.47 ms.488 ms 7 po-.lond-scr.ja.net ( ).735 ms.7 ms.6 ms 8 po-.cambridge-bar.ja.net ( ) 5.3 ms ms ms 9 route-enet-3.cam.ac.uk ( ) 4.98 ms 4.84 ms 4.86 ms route-cent-3.cam.ac.uk ( ) ms ms 4.86 ms 7

8 The big picture emerging from the traceroute Host Interface (IP address A) Subnet /4 Interface (IP address B) Router Interface (IP address C) Subnet /4 Interface (IP address D) Router 9

9 Routers keep routing tables Three fields Destination: destination IP Outgoing interface: on which to forward packet for the given destination Metric: total cost to reach the destination depends on interfaces metrics (set by net admins) Destination Interface Metric A B

10 Routers destinations are IP prefixes Each host (interface) has unique 3-bit IP address E.g., Must every router in entire Internet know about every other router and host? No; interfaces on same subnet share IP prefix e.g., /4 for IP routing destination is subnet s prefix Not single IP addresses

11 Routers use routing table to forward packets Packet for destination D arrives Router searches D in destination field of routing table using longest-prefix match If more than one route match D, choose the route with the longest prefix Possible outcomes: if entry for D, forward on interface in the entry if no entry, no route known à drop packet 3

12 Routers use routing table to forward packets Packet for destination D arrives Router searches D in destination field of routing table using longest-prefix How to populate match routing tables for any router in a domain? If more than one route match D, choose the route with the longest prefix Possible outcomes: if entry for D, forward on interface in the entry if no entry, no route known à drop packet 4

13 At startup, a router initializes its routing table for directly connected subnets Subnet /4 Host Interface (IP address B) Destination Interface Metric /4 Interface (IP address A) /4 Router Interface (IP address C) Subnet /4 Interface (IP address D) Router 5

14 At startup, a router initializes its routing table for directly connected subnets Subnet /4 Host Interface (IP address B) Destination Interface Metric /4 Interface (IP address A) /4 Router Interface (IP address C) Subnet /4 Packets to other destinations would be dropped Interface (IP address D) Router 6

15 At startup, a router initializes its routing table for directly connected subnets Subnet /4 Host Interface (IP address B) Destination Interface Metric /4 Interface (IP address A) /4 Routing protocols are needed to add information on remote destinations Router Interface (IP address C) Subnet /4 Packets to other destinations would be dropped Interface (IP address D) Router 7

16 Agenda We delve into network layer s main functionality. Setting Context Routing players. Intra-domain routing problem definition 8

17 Consider an arbitrary network Each router has an ID and several interfaces Linked to many other routers and subnets D Each router must choose the next-hop for received packets Packets include info on destinations (IP addresses) But not on intermediate hops?? S? 9

18 Routing protocols build routing tables Routing protocols define information and messages exchanged by routers For routers to accumulate state (routing table entries) to forward packets D Routing protocols also determine how to choose between alternative next-hops S

19 Intra-domain routing goals Provide good performance Typically, choose good paths for any pair of routers Automated path computation Possibly, efficient in time and memory Deal with failures Routes must be coherent with topology

20 Typical solution: shortest path routing View network as weighted graph Routers are vertices, links are edges Link metrics are edge weights 5 5 5

21 Typical solution: shortest path routing Shortest paths problem: What path between two vertices offers minimal sum of edge weights? Classic graph algorithms find single-source shortest paths when the entire graph is known centrally Dijkstra s Algorithm, Bellman-Ford Algorithm Typically, no central knowledge of entire graph Each router only knows its own interfaces addresses 3

22 Challenge: deal with changing networks Network components are not reliable Links may be cut, router or their interfaces may fail Changing networks induce hazards Need for (fast) re-convergence to new, correct paths after a failure Potential for forwarding loops Amplify traffic à congest links TTL will eventually expire, but typically too late Possibility of temporary disconnections (i.e., blackholes) 4

23 Distance Vector Routing Stefano Vissicchio UCL Computer Science COMP3

24 Agenda We are now ready to study Distance Vector routing. Algorithm. Pathologies 3. Optimizations 6

25 Basic Distance Vector Algorithm Distributed Bellman-Ford (DBF) Principle: tell everything you know to your neighbours Periodically, send all your routing table entries (destination and metric fields) to all your neighbouring routers 7

26 Basic Distance Vector Algorithm: DBF (Failures Not Yet Considered) Upon receipt of routing table entry for destination D with metric m on interface i: m += configured metric for interface i rt = lookup(d) in roung table if (rt = not found ) then rt_new = new roung table entry rt_new.dst = D; rt_new.metric = m; rt_new.iface = i add rt_new to roung table else if (m < rt.metric) then rt.metric = m; rt.iface = i 8

27 Distance Vector: Example Consider simple network where all nodes are routers, addresses are simply single letters Initial routing tables when routers first start: A local Dst Metric A A D B E B local C C local D local E local 9

28 Distance Vector: Iteration Routers incorporate received announcements: A local B D A D D local A E B local A C E B E E local D B C C local B E C 3

29 Distance Vector: Iteration Routers incorporate received announcements: A local B local B A D C E E C D A D D local E local A D E B B C C A B E C C local B E A D 3

30 Distance Vector: Iteration Convergence: Routers incorporate received announcements: routing tables no longer changing; A local B local routes reflect up-to-date knowledge of topology B A D C E E C D A D D local E local A D E B B C C A B E C C local B E A D 3

31 Link Failure () A local B local B A D C E E C! A D D B! E C C local B D local E local E A D A E B D B C C A 33

32 Link Failure () A local B local B Inf A Inf D C E E C Inf A D D B E C C local B D local E local E A D A E B D B C C A 34

33 Link Failure () A local B local B Inf A Inf D C E Dst Metric C Inf A B Inf D E C Inf A D E D B E C C local B D local E local E A D A E B D B C C A 35

34 Link Failure () A local B Inf D E Dst Metric C Inf A B Inf D E C Inf A D D local A E B C Problem: D ignores the update for the route to B because the metric is worse than the one B it local already has. A Inf C E D B E E local D B C A C C local B E A D 36

35 DV Algorithm, Revised Upon receipt of routing table entry for destination D with metric m on interface i: m += metric for interface i rt = lookup(d) in rou<ng table if (rt = not found ) then r_new = new rou<ng table entry r_new.d = D; r_new.metric = m; r_new.iface = i add r_new to table else if (i == rt.iface) then rt.metric = m else if (m < rt.metric) then rt.metric = m; r.iface = i 37

36 Link Failure (3) A D Dst Metric A D local D local B Inf D + A E A E E 3 B B Inf C Inf C C 38

37 Link Failure (3) A local B local B Inf A Inf D C E E C Inf A D D B E C C local B D local E local E A D A E B D B Inf C C A 39

38 Link Failure (4) A local B local B Inf D E C Inf A D A Inf Dst Metric C B E A Inf D C E D B E C C local B D local E local E A D A E B D B Inf C C A 4

39 Link Failure (5) B E Dst Metric B E local E local A Inf C + D B D B E C C D 3 A A Inf 4

40 Link Failure (5) A local B local B Inf A Inf D C E E C Inf A D D B E C C local B D local E local E A D A E B D B Inf C C A Inf 4

41 Link Failure (6) A local B local Next B round: all Inf routers A broadcast Inf their tables to neighbours D E C E C Inf D A D D local E local A D E B B Inf C C A Inf B E C C local B E A D 43

42 Link Failure (7) A local B local B Inf A Inf D C E E C 3 A D D B E C C local B D local E local E A D A Inf E B D B C C A 44

43 Link Failure (8) A local B local B 3 A 3 D C E E C 3 A D D B E C C local B D local E local E A D A 3 E B D B C C A 45

44 Agenda We are now ready to study Distance Vector routing. Algorithm. Pathologies Bouncing Counting to Infinity 3. Optimizations 46

45 Bouncing (I) Consider same network, where link (C, E) has metric ; all others have metric Consider all nodes routes to C after convergence Now suppose link (B, C) breaks C A D C 3 C B E C C C local 47

46 Bouncing (II) Suppose A advertises its table first C A D C 3 C Inf B E C C C local 48

47 Bouncing (III) Suppose A advertises its table first and B advertises its table next C A D C 3 C 3 B E C C C local 49

48 Bouncing (IV) Suppose A advertises its table first and B advertises its table next Loop between A and B for destination C! If C now advertises its table, E will ignore cost route! C 4 A D C 3 C 3 B E C 4 C C local 5

49 Bouncing (V) Suppose A and E advertise next (B,D update metrics) C 4 A D C 5 C 5 B E C 4 C C local 5

50 Bouncing (VI) Suppose A and E advertise next (B,D update metrics) and B advertises next (A and E update metrics) C 6 A D C 5 C 5 B E C 6 C C local 5

51 Bouncing (VII) Suppose A and E advertise next (B,D update metrics) and B advertises next (A and E update metrics) And so on and A advertises next (B updates metric) C 6 A D C 5 C 7 B E C 6 C C local 53

52 Bouncing (VIII) Long, painful convergence process, details dependent on message ordering Transient loops Eventually, converged state: C A D C C B E C C C local 54

53 Counting to Infinity (I) Converged after link (A, B) breaks Suppose (D, E) now breaks A local B 3 D E C 3 D local A E A D B local A 3 C E D B E E local D B C C local B E A 3 D B C C A 55

54 Counting to Infinity (II) A local B 3 D E C 3 Dst Metric A B 4 D E 3 C 4 + A D D local A E Inf B Inf C Inf Network partitioned Focus on {A, D} partition Suppose sequence of events: D notices link failure A advertises its routing table Loop for {B, C, E} between A and D! How long will loop persist? D local A E 3 B 4 C 4 56

55 Counting to Infinity (III) A local A local A local A B 5 D E 4 C 5 B 7 D E 6 C 7 B Inf D E Inf C Inf D D local D local D local A E 3 A E 5 A E Inf B 4 B 6 B Inf C 4 C 6 C Inf Each advertisement increments metrics for partitioned destinations by one Loop persists until count reaches infinity! 57

56 Agenda We are now ready to study Distance Vector routing. Algorithm. Pathologies 3. Optimizations Split Horizon Poison Reverse 58

57 Avoiding direct loops Bouncing and counting to infinity cause slow convergence, create loops Consider any link (X, Y), and destination Z with X s next hop toward Z being Y Split Horizon: clearly, Y should never choose X as next hop toward Z Intuition: X should never announce to Y a path with short distance to Z if it s forwarding to Z via Y! 59

58 Split Horizon with Poison Reverse Again, consider link (X, Y), destination Z with X s next hop toward Z being Y More generally: routers should announce different routing tables to different neighbors Split horizon: don t announce route for destination Z on interface used as next hop toward Z! Poison Reverse (optional): X announces to Y its distance to Z is infinity! 6

59 Example: Split Horizon and Poison Reverse A local B 3 D E C 3 Dst Metric A B Inf D Inf E Inf C Inf + A D D local A E Inf B Inf C Inf Same example as counting to infinity: {A, D} partitioned D detects link break, A announces first No loop, immediate convergence after one advertisement! D local A E Inf B Inf C Inf 6

60 Limitations: Split Horizon and Poison Reverse Consider same example, but {B, C, E} partition Link (A, B) already failed, routing has converged Now link (D, E) fails Consider only destination D A D B E D D C D 6

61 Limitations (II): Split Horizon and Poison Reverse E notices failed link, updates local table B E D C D D Inf 63

62 Limitations (III): Split Horizon and Poison Reverse E advertises its new table Suppose advertisement reaches B, but not yet C B E D Inf C D D Inf 64

63 Limitations (IV): Split Horizon and Poison Reverse C advertises its table, with split horizon and poison reverse B E D Inf C + Dst Metric D 3 D D 3 D Inf + Dst D Metric Inf D Inf 65

64 Limitations (V): Split Horizon and Poison Reverse B advertises its routing table, with split horizon and poison reverse For destination D, loop {C à E à B à C}! resolved only by counting to infinity D 3 B E C D + Dst D Metric Inf D D Inf + Dst Metric D 4 D 4 66

65 Summary: Distance Vector routing DV algorithm: periodically dump routing table contents to neighbors Convergence: after topology change, point when routing tables stop changing Pros: simple finds correct routes after topology changes Cons: bouncing, counting to infinity cause loops slow to converge after some topology changes split horizon, poison reverse only partial solutions 67