Routing Ethernet Switches to BGP (almost)

Transcription

1 Routing Ethernet Switches to BGP (almost) John Jannotti March 7, 2013 John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

2 The story so far... You have invented the Local Area Network (LAN) Your computers Are connected to some shared media Identify themselves with a unique address Send short frames to other connected computers. Sadly Your invention doesn t work over long distances. You aren t the only enterprising inventor. You need to connect (your and others ) LANs. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

3 A Running Analogy Your town has a strange postal system (LAN). You put your letter (payload) in a special envelope (frame) with an address (destination) and return (source) address on it into any mailbox (network interface) which projects your letter in the sky (media) for everyone in town to see. (Usually, only the intended recipient bothers to look.) Keep in mind, your postal system has its own addresses and envelopes. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

4 Bridges and extended LANs (intercity courier) A B C Port 1 Bridge Port 2 X Y Z LANs have physical limitations (e.g., 2500m or radio range) Idea: Connect two or more LANs with a bridge Operates on the LAN s notion of address. (e.g Ethernet) No encapsulation required (same envelopes) John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

5 Bridges and extended LANs (intercity courier) A B C Port 1 Bridge Port 2 X Y Z LANs have physical limitations (e.g., 2500m or radio range) Idea: Connect two or more LANs with a bridge Operates on the LAN s notion of address. (e.g Ethernet) No encapsulation required (same envelopes) A bridge connects two (or more) LANs. and copies every frame from input to all outputs. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

6 Bridges and extended LANs (intercity courier) A B C Port 1 Bridge Port 2 X Y Z LANs have physical limitations (e.g., 2500m or radio range) Idea: Connect two or more LANs with a bridge Operates on the LAN s notion of address. (e.g Ethernet) No encapsulation required (same envelopes) A bridge connects two (or more) LANs. and copies every frame from input to all outputs. which isn t very efficient. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

7 Learning bridges (smart, lazy couriers) A B C Port 1 Bridge Port 2 Idea: Don t forward packet if not useful If you know recipient is not on that port X Y Z Switch builds a table as it observes source address. A B C X Y Z Table says when not to forward packet No entry? Copy it for correctness. Packets can be marked Broadcast. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

8 Dealing with loops A B3 B C B5 B2 D B7 K E F B1 G H I B6 B4 J Problem: LANs might be connected in loops. Maybe accidentally or maybe to provide redundancy. Don t want to forward packets forever! John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

9 Spanning tree A B3 B C B5 B2 D B7 K E F B1 G H I B6 B4 J Need to disable ports to break cycles. Like creating a spanning tree in a graph. View switches and networks as nodes, ports as edges But you need a distributed solution. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

10 Spanning tree algorithm Every bridge has a unique ID (Ethernet address) Let bridge with the smallest ID be the root. Each segment has one designated bridge responsible for forwarding packets over it. Bridge closest to root is designated bridge. If tied, smallest ID wins. Overview: Node begins by assuming it is the root. Broadcasts messages, may learn of a better root. Stop forwarding messages if not designated bridge. Eventually designated bridges form a spanning tree. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

11 Limitations of bridging Does not scale well to many (thousands?) of computers. Spanning tree algorithm does not scale. Broadcast does not scale. No way to route around congested links. Doesn t even statically optimize routes. What about other kinds of LANs? (Towns with another kind of envelope.) John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

12 Add a Network Layer Goal: Glue lower-level networks together Network 1 (Ethernet) H1 H2 H3 H7 R3 H8 Network 2 (Ethernet) R1 Network 4 (point-to-point) R2 H4 Network 3 (FDDI) H5 H6 TCP H1 R1 R2 R3 TCP H8 IP IP IP IP IP ETH ETH FDDI FDDI PPP PPP ETH ETH John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

13 IP: Internet Protocol Agree on (globally unique) addresses. Forward based on destination address TTL (time to live) decremented at each hop (avoids loops) Originally was in seconds (no longer) TTL mostly saves from routing loops Following IP header is payload data Typically beginning with TCP or UDP header A new, inner envelope that all towns agree on. Local postal systems stay the same, but new envelope is inside. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

14 Service model Connectionless (datagram-based) Best-effort delivery (unreliable service) packets are lost packets are delivered out of order duplicate copies of a packet are delivered packets can be delayed for a long time No coincidence that this model resembles many LAN models. The End-to-end Argument makes a compelling case that this is the right choice. The end-hosts will need to ensure reliability anyway, so don t waste too much effort along the way. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

15 What is routing? Packet forwarding moving packets between ports Look up destination address in forwarding table Find out-port or out-port, MAC addr pair Routing is process of populating forwarding table Routers exchange messages about destinations they can reach. Goal: Find optimal route for every destination, or maybe good route, or maybe just any route (depending on scale). Intra-domain vs. Inter-domain routing Intra-: All routers under same administrative control Intra-: Scale to 100 networks (e.g., campus like Brown) Inter-: Decentralized, scale to Internet John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

16 Optimality C 4 B 3 9 A 1 D 1 1 E 6 2 F View network as a graph Assign cost to each edge Can be based on latency, b/w, utilization, queue length,... Problem: Find lowest cost path between two nodes Again, must be computed in a distributed way John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

17 Distance Vector Each node maintains a set of triples (Destination, Cost, NextHop) Exchange updates directly connected neighbors periodically (on the order of several seconds to minutes) whenever table changes (called triggered update) Each update is a list of pairs: (Destination, Cost) the sender is the implied NextHop Update local table as routes are received. usually, store all advertisements but forwarding will only be affected by changes to the best choice. Refresh existing routes; delete if they time out John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

18 Example B A F E C G D Destination Cost NextHop A 1 A C 1 C D 2 C E 2 A F 2 A G 3 A B s routing table John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

19 Adapting to failures B A C D E F G F detects that link to G has failed F sets distance to G to infinity and sends update to A A sets distance to G to infinity since it uses F to reach G A receives periodic update from C with 2-hop path to G A sets distance to G to 3 and sends update to F F decides it can reach G in 4 hops via A John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

20 Danger: Loops count to infinity A B C D E F G link from A to E fails A advertises distance of infinity to E B and C advertise a distance of 2 to E B decides it can reach E in 3 hops; advertises this to A A decides it can reach E in 4 hops; advertises this to C C decides that it can reach E in 5 hops... Solution? John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

21 Danger: Loops count to infinity A B C D E F G link from A to E fails A advertises distance of infinity to E B and C advertise a distance of 2 to E B decides it can reach E in 3 hops; advertises this to A A decides it can reach E in 4 hops; advertises this to C C decides that it can reach E in 5 hops... Solution: Make infinity small. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

22 Link State Strategy Broadcast to all nodes (not just neighbors) But only send information about directly connected links (not entire routing table) Less data, but sent everywhere. Link State Packet (LSP) ID of the node that created the LSP Cost of link to each directly connected neighbor Sequence number (SEQNO) Time-to-live (TTL) for this packet LSPs must be flooded (reliably and scalably). John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

23 LSP flooding Store most recent LSP from each node Forward LSP to all nodes but one that sent it Generate new LSP periodically Increment SEQNO Start SEQNO at 0 when reboot If you hear your own packet w. SEQNO= n, set your next SEQNO to n + 1 Decrement TTL of each stored LSP discard when TTL= 0 Contrast this to Ethernet s broadcast mechanism. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

24 Calculating best path Dijkstra s shortest path algorithm Let: N denote set of nodes in the graph l(i, j) denotes non-negative cost (weight) for edge (i, j) s denotes yourself (node computing paths) Initialize variables M {s} (set of nodes incorporated so far) C n l(s, n) (cost of the path from s to n) R n (next hop on path to n) John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

25 Dijkstra s algorithm While N M Let w (N M) be node with lowest Cw M M {w} Foreach n (N M), if C w + l(w, n) < C n then C n c w + l(w, n), R n w Example: D (D, 0, )(C, 2, C)(B, 5, C)(A, 10, C) A 5 B D C John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

26 Distance Vector vs. Link State # of messages DV: O(d) where d is # of neighbors of node LS: O(n d) for n nodes in system Computation DV: Could count all the way to if loop LS: O(n 2 ) Robustness what happens with malfunctioning router? DV: Node can advertise incorrect path cost DV: Costs used by others, errors propagate through net LS: Node can advertise incorrect link cost John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

27 Intradomain routing protocols RIP (routing information protocol) Fairly simple implementation of DV OSPF (open shortest path first) LS-based protocol Adds notion of areas for scalability Area 0 is special backbone area Traffic between two areas must always go through area 0 Only need to know how to route exactly within area Else, just route to appropriate area (Virtual links can allow distant routers to be in area 0) Addresses in an areas must be aggregated somehow. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

28 OSPF areas Area 1 Area 0 Area 3 R9 R8 R7 R1 R3 R4 R2 Area 2 R6 R5 John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

29 Address Aggregation Each router can forward to any destination IP address. Given address, it needs to consult table for next hop. Naïve: Have an entry for each address. Large tables, just for Brown. Internet-wide, there could be 2 32 entries! John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

30 Address Aggregation Each router can forward to any destination IP address. Given address, it needs to consult table for next hop. Naïve: Have an entry for each address. Large tables, just for Brown. Internet-wide, there could be 2 32 entries! Solution: One entry can cover a range of addresses But can t do this if addresses are assigned randomly! Addresses allocation is a big deal. Must take advantage of network structure. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

31 The Internet, 1990 Stanford NSFNET backbone ISU Berkeley BARRNET regional PARC NCAR Westnet regional UNM UNL MidNet regional KU UA Hierarchical structure. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

32 The Internet, today Large corporation Peering point Consumer ISP Backbone service provider Consumer ISP Peering point Large corporation Consumer ISP Small corporation Multiple backbones. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

33 Exploit structure of network Consider IP address to be hierarchical. First idea Class A (8-bit prefix), B (16-bit), C (24-bit) Routers need only know route for each network Better: Classless Inter-Domain Routing (CIDR) Be explicit about the width of the prefix /24 is just like a Class B. But /25 is also legal. Make route choice based on longest-prefix-match. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

34 Route Propagation Know a smarter router hosts know local router local routers know site routers site routers know core router core routers know everything (everything?!) Instead, allow Autonomous Systems (AS) corresponds to an administrative domain examples: University, company, backbone network. Two-level route propagation hierarchy interior gateway protocol (each AS selects its own) exterior gateway protocol (Internet-wide standard) John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

35 Autonomous systems Corresponds to an administrative domain Internet is not a single network ASes reflect organization of the Internet E.g., Brown, large company, etc. Goals: ASes want to choose their own local routing algorithm. ASes want to set policies about non-local routing. Each AS assigned unique 16-bit number John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

36 Typical Policies depend on size, connectivity. Tier-1 AS: Default-less routing Connects to multiple ASes and carries transit traffic. Usually gets paid (except by other Tier-1s) Tier-3 ASs: Stubs, Customers Connects to (at least one) Tier-2 or higher. Carries no transit traffic. Tier-2 AS: Regional, Smaller countries Connects to paying customers. Connects to (at least) one Tier-1 (that it pays). Carries transit traffic (carefully!). Seeks peering agreements to limit costs. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

37 Tier-2 ISPs are the most careful They must advertise routes for their own customers to all. They also must route default traffic to their Tier-1. To save money, they seek peering agreements with other Tier-2s. Peers typically do not pay each other. Intent is to save money on direct connection between customers of each. Use BGP policies To prefer your (cheap!) peering route over other options. To avoid advertising the peering route to your Tier-1. Or to your other peers. (Just customers.) John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

38 Which algorithm should INTER-domain routing use? Constraints: Scaling Autonomy (policy and privacy) Link-state? Requires sharing of complete network informatin Can t express policy Distance Vector? Retains privacy (only advertises next hop) Can t implement policy Can t avoid loops if shortest paths not taken John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35

39 Path Vector Protocol Distance vector algorithm with extra information For each route, store the complete path (ASs) No extra computation, just extra storage Advantages: Can make policy choices based on set of ASs in path. Can easily avoid loops. John Jannotti Routing Ethernet Switches to BGP (almost) March 7, / 35