A packet does not know how to get to its destination

Must rely on the routers to send it in the right direction

So how do the routers do that?

A router can't possibly know where everything in the world is: it is only connected to a handful of neighbour routers

How can a router in England know that to send a packet to Australia it might have to forward it to America first?

Routers exchange information with each other to find the best possible route to a destination

Two types of routing•Autonomous System (AS): Local routing

within an organisation (requiring an interior gateway protocol)

•Global (non-local) routing between organisations (AS’s) (requiring an exterior gateway protocol)

Very different requirements, with exterior protocols mostly driven by local policy and economical constraints

Interior Gateway Protocols Requires static route added by hand,

e.g., the route commandroute add default gw 213.121.147.69adds a default route to a gateway with

IP address 213.121.147.69 Routing tables on most non-routers

are trivial and set up manually by the operating system at boot time

ICMP Redirect

Sometimes routing tables are not perfectly set up

H1 wants to send to H2 but H1's routing table tells it to route via R2

When the packet reaches R2, it sees it should be routed out on the interface it came in on: so R2 knows H1's table need improving

R2 forwards the packet and send an ICMP redirect to H1

H1 gets the redirect and uses it to update its routing table. The route will be marked D or M

Next time H1 will be able to route directly to R1

ICMP redirects are OK for small fixes, but in general routers need something more substantial

Could get administrators to set up the tables

Better to get the routers to do it themselves

Dynamic routing is the passing of routing information between routers

Several protocols are used• Routing Information Protocol (RIP)• Open Shortest Path First (OSPF)• Border Gateway Protocol (BGP)• Many more

Each protocol is suited to a certain purpose, no single protocol fits all

Internet is managed as a collection of Autonomous Systems (AS), each administered by a single entity, e.g., a University or company

Each AS chooses a routing protocol to direct packets within the AS: these are interior gateway protocols (IGP), e.g., RIP or OSPF

Between ASs run exterior gateway protocols (EGP), e.g., BGP

Typically a router runs a program whose sole purpose is to manage routing information and update the routing table. This program may understand several routing protocols•routed talks RIP•gated talks RIP, OSPF and BGP

Distance-Vector and Link-State Routers can collect all kinds of

information about who is connected to whom and in what manner: this must be condensed down to something that can be put in a routing table

Two main algorithms are used:•distance-vector protocols• link-state protocols

Distance-vector gathers collections (vectors) of hop counts (distances) from its neighbouring routers to selected destinations. From this it computes its own vector of distances. RIP is an example

Link-state gathers maps of connectivity from all the routers in the AS (or some subset) and uses this to compute its own map. OSPF is an example

Distance-Vector vs Link-State Distance-vector is simple, but has

problems Link-state is more complex, but has

advantages

When a router starts it broadcasts a RIP request on all its interfaces with address family 0 and metric 16: this is a “send me all your routes” request

A router receiving such a request replies with its vector of distances

Otherwise a request is for a route to a particular address or addresses

A reply is a metric for a route, or 16 to indicate infinity or “no route”

Given a reply we can update our routing table: our metric is the received metric plus 1 for the hop to the router that replied

If a new route with a smaller metric than an existing route we can update our table to use the new route: RIP always deems shorter paths to be better

RIP also sends a chunk of the routing table every 30 seconds to its neighbours: routes are timed out if they are not confirmed for 3 minutes

A timed-out route is set to metric 16, but not deleted for 60 seconds: this ensures the invalidation is propagated

R3 knows a route to H with metric 1 R3 sends a RIP message to R2, R2

learns a route to H via R3 with metric 2

R2 sends a message to R1, R1 learns a route to H via R2 with metric 3

Note that R2 sends a message to R3, too, so R3 learns there is a route to H via R2 with metric 3

But R3 can ignore this as it already knows a better route

Updating of Vector Distance tables

RIP has problems, in particular the long time (several minutes) it takes to readjust after a route is added or removed somewhere: the slow convergence, or count to infinity problem

The count-to-infinity problem:

A is down

Then A comes up. The good news spreads relatively quickly.

The count-to-infinity problem

A is up

Bad news spread slowly.

Another problem is that RIP is not suitable for use between ASs as it is ignorant of subnetting and possibly reveals too much about the internal structure of the AS

As RIP uses broadcast a router only gets information from its immediate neighbours and this makes getting a global view of the network difficult

The metric limit of 15 is not big enough for larger networks: making it bigger only makes the count to infinity worse

The only criterion for choice of route is the metric: this is much too simplistic. Other considerations including network speed, bandwidth and cost should be taken into account

The Open Shortest Path First protocol is mainly based on the Dijkstra's shortest path first Algorithm

This is a method for finding best paths through a network, where “best” means “shortest” or “lowest cost” or whatever you want to measure

It is used in several routing protocols

We want to find the shortest/cheapest path between V and Z, where costs are as given

We make a table of paths, containing a predecessor to a node and a cost to get to that node. Also mark each node determined or undetermined

1.Initialise all costs to ∞, predecessors to blank and all nodes undetermined

2.Initialise the cost of the starting node V to 0

3.While there are any undetermined nodes:a) pick an undetermined node of lowest cost; call

it Cb) mark C determinedc) for each undetermined neighbour N of C

if cost to C + cost N to C < cost to N we have found a shorter path to N via C; make the cost of N be cost to C + cost N to C and the predecessor of N to be C

Determined nodes are those for which we definitely know the best route; undetermined are those for which we might yet find a better route

Start. Pick cheapest undetermined node: V

Make it determined and consider its neighbours W, X and Z

All have infinite cost, so make their predecessors V and adjust their costs

V W X Y Z

0 2 4 ∞ 7

- V V - V

det undet undet undet undet

Pick cheapest undetermined node: W Make it determined, and consider its

undetermined neighbours X and Y Cost to X via W is 2+1<4, so revalue

X; cost to Y via W is 2+1<∞, so revalue Y

V W X Y Z

0 2 3 3 7

- V W W V

det det undet undet undet

Pick cheapest undetermined node: X (or Y)

Make it determined and consider its undetermined neighbour Z

Cost to Z via X is 3+2<7 so revalue Z

V W X Y Z

0 2 3 3 5

- V W W X

det det det undet undet

Pick cheapest undetermined node: Y Make it determined and consider its

undetermined neighbour Z Cost to Z via Y is 3+3>5 so don't

revalue Z

V W X Y Z

0 2 3 3 5

- V W W X

det det det det undet

Pick cheapest undetermined node: Z Make it determined. It has no

undetermined neighbours so we are done

V W X Y Z

0 2 3 3 5

- V W W X

det det det det det

Final costs:

So cheapest path from V to Z is of cost 5

Also have computed cheapest paths from V to every other node

V W X Y Z

0 2 3 3 5

- V W W X

Final costs:

And have computed the paths: to get from V to Z read the table in reverse. Start with Z and read successive predecessors: Z X W V

So path is V W X Z

V W X Y Z

0 2 3 3 5

- V W W X

TANENBAUMComputer Networks 139

Software responsible for deciding which output line an incoming packet should be transmitted on

Datagram• Routing decision at every node

VC• Routing decision at the setup phase

TANENBAUMComputer Networks 140

The Optimality Principle Shortest Path Routing Flooding Distance Vector Routing Link State Routing Hierarchical Routing Broadcast Routing Multicast Routing Routing for Mobile Hosts Routing in Ad Hoc Networks

TANENBAUMComputer Networks 141

Correctness Simplicity Robustness

• What happens if the topology changes? Stability

• Must converge to an equilibrium Fairness Optimality

• Minimize mean packet delay• Maximize total network throughput

TANENBAUMComputer Networks 142

Conflict between fairness and optimality

TANENBAUMComputer Networks 143

Nonadaptive algorithms• Route is computed in advance and stored in

the router• This is called static routing

Adaptive algorithms• Base their decision on current

measurement of traffic and topology

TANENBAUMComputer Networks 144

Treat routers as nodes Assign weights to links

• E.g. Fix all weight to 1 to minimize the number of

hops crossed Use distance in kilometers as weight to minimize

delay Use Dijkstra’s algorithm to determine

the shortest path

TANENBAUMComputer Networks 145

TANENBAUMComputer Networks 146

Initialize• T = {s} set of nodes incorporated (s source)• L(n)=w(i,j) for ns initial path costs to neighbors

Get next node• Add the neighbor x not in T that has least cost

L(x) to s.• Add the edge from this neighbor to old T that

contributes to the path Update least cost paths

• For all nT, update costs by L(n)=min[L(n),L(x)+w(x,n)]

TANENBAUMComputer Networks 147

Dijkstra's algorithm to compute the shortest path through a graph.

5-8 top

TANENBAUMComputer Networks 148

Dijkstra's algorithm to compute the shortest path through a graph.

5-8 bottom

TANENBAUMComputer Networks 149

Every incoming packet is sent out on every outgoing line except the one it arrived on

Generates vast number of duplicate packets

Solutions• Put a hop counter to every packet• Avoid flooding the same packet second time

Needs packet sequence numbers• Use selective flooding

Do not put the packet to every outgoing line, instead put the packet to certain outgoing lines

Used in military applications and measurements

TANENBAUMComputer Networks 150

Takes load into account Needs

• Topology• Load on the lines• Capacity of the lines

Given the load and capacity, delay of the line can be computed via queueing theory• T = 1 / (C - ) where

1/ is mean packet size in bits C is line capacity in bps mean flow in packets per second

TANENBAUMComputer Networks 151

Distance vector routing• Did not take line bandwidth into account• Took too long to converge

Link state routing: Each router• Discover its neighbors, learn their network

address.• Measure the delay or cost to each of its

neighbors.• Construct a packet telling all it has just learned.• Send this packet to all other routers.• Compute the shortest path to every other

router.

TANENBAUMComputer Networks 152

When a rooter is booted, it sends HELLO packet to all of its neighbors

Neighbors reply by sending their globally unique names

Modeling LANs need extra mechanisms• In the next slide an artificial node N is

added to represent the LAN From A to C, the path is represented as ANC

TANENBAUMComputer Networks 153

(a) Nine routers and a LAN. (b) A graph model of (a).

TANENBAUMComputer Networks 154

Each router should know an estimate of delay to each of its neighbors

To estimate the delay, a router can send an echo packet to its neighbor and measure the time when it receives it back

While measuring time, it is possible to include or exclude the time the echo packet spends on the queues which corresponds to consider load on the line

Including load may cause oscillations in the algorithm

TANENBAUMComputer Networks 155

A subnet in which the East and West parts are connected by two lines.

TANENBAUMComputer Networks 156

Packets contain• Identity of the sender• Sequence number• Age• List of neighbors and delays

Packets canbe issued• Periodically or• When a significant event occurs

TANENBAUMComputer Networks 157

(a) A subnet. (b) The link state packets for this subnet.

TANENBAUMComputer Networks 158

Flooding is used Packet sequence numbers are used to

discard duplicates or old packets by holding packets for a short while before flooding them

32-bit sequence numbers are used to prevent wrap-arounds

For each packet, in addition to sequence number, age is used and decremented every second to solve• Router crash problem• Corrupted sequence number problem

TANENBAUMComputer Networks 159

The packet buffer for router B in the previous slide (Fig. 5-13).

TANENBAUMComputer Networks 160

When a router accumulates a full set of link state packets, it can construct the entire network graph

It then runs Dijkstra’s algorithm to construct shortest paths to all possible destinations

Even for a large subnet, the memory requirements are reasonable

Malfunctioning routers may cause problems Protocols using link state routing

• OSPF and IS-IS