routing in communication networks routing: network layer protocol that guides information units to...

66
Routing in Communication Networks Routing: Network layer protocol that guides information units to correct destinations. A complex collection of decision rules that work somewhat independently. - Determination of allowable path set for each origin - destination pair - Broadcast of routing related information to all network nodes - Determination of particular path per arrived call.

Upload: virgil-shields

Post on 27-Dec-2015

230 views

Category:

Documents


3 download

TRANSCRIPT

Routing in Communication Networks

Routing: Network layer protocol that guides information units to correct destinations. A complex collection of decision rules that work somewhat independently.

- Determination of allowable path set for each origin - destination pair

- Broadcast of routing related information to all network nodes

- Determination of particular path per arrived call.

Routing Types

Virtual circuit routing

Datagram routing

Characteristics

Performance indices

Throughput

Delay

Under light load: throughput = offered load

Under heavy load: throughput = offered load - rejected load

Routing interacts with flow control to maintain the performance

Traffic experiences delay depending on routing

Flow control rejects a portion of load depending on delay

Routing Types Revisited

Virtual circuit vs. datagram

Centralized vs. distributed

Static vs. adaptive

Hierarchical vs. non hierarchical

Broadcasting

Send information from a node to all nodes in networks (e.g., announcing equipment and link failures)

Send information from a node to some selected subset of nodes

Examples of broadcasting method: Flooding

A node send a packet to all of its neighbors.

Each neighbor sends the received packet to its neighbors and so on.

Loop back should be eliminated.

Multiple transmissions must be eliminated.

Not efficient.

Example: Spanning Tree

Connected sub-graph of network including all nodes

No cycles

Packets transmitted over spanning tree

More efficient than flooding

Packets transmitted N-1 times (N = number of nodes)

Routing Examples Shortest path routing

Each link has a link length (> 0).Path length is the sum of link lengths over the path.Shortest path routing routes each packet over a minimum length path (i.e.,a greedy algorithm).Minimum hop routing: If every link has a unit link cost.

Optimal routingRoutes are determined for all OD pairs simultaneously for the good of entire network.Can eliminate potential oscillations in the shortest path routing.

Hot potato routingMinimize buffer requirements.Arrived packets are routed on an out-going link from the list of links.Packets may get out of sequence.Packets may experience long delay.

Cut-through routing

Minimizes delay.

Packets are allowed to go out on an out-going link as soon as initial few bits arrive.

Error checking and ARQ cannot be done.

Link reservation is needed while transmitting a packet.

Speed mismatch among links can create overflows/congestions.

ARPANET routingAdaptive and distributed datagram routingShortest path routing based on link delay as link lengthShortest paths are calculated using Bellman-Ford algorithm.Link lengths are updated every 10 seconds.

TYMNET routingCentralized and adaptive virtual circuit routingSupervisor determines the routing for all nodes.Shortest lath routing based on link load as link length

SNA routingNetwork manager chooses virtual paths for each OD pair

Circuit switched routingLike virtual circuit routingTypically implemented with hierarchy

Interconnected Network Routing

Involves gateways, routers, and bridges between networks

Non hierarchical routing

Control becomes complex.

Much information needs be stored at each node.

Hierarchical routing

long distance routing

Routing within sub-network

Lower information storage at potentially lower performance

Graphical Model for Routing

Use of Graph Theoretic Model Undirected graph G = (N, A)N = finite non empty set of nodesA = collection of pairs of distinct nodes from N called arc.

N = {1,2,3,4}A = {(1,2), (2,3), (4,1), (2,4)}

Note:No loopNo multiple arcs between the same pair of nodesA node is also called a vertex.An arc is also called an edge, link, or branch.

1 2

34

Definition:

A walk is a sequence of nodes (n1, n2, n3,…,nl) such that (n1, n2) (n2, n3),…(nl-1, nl)

are each an arc in G.

Definition:

A path is a walk with no repeated node.

Definition:

A cycle is a path with no repeated node expect for n1 = nl.

1, 4, 2, 1, 4, 1 walk1, 4, 2, 3 path1, 2, 4, 1 cycle1

2

4

3

Definition

A graph is connected if from each node i, there is a path to any other nodes j.

1

2

3

5

4

unconnected graph

Lemma: Let G = (N,A) is a graph. S N and S N. Then G is connected if for any such S there is at least one arc (i, j) with i S and j S.

Definition: g= (n, a) is a sub-graph of G=(N,A) if g is a graph and n N and a A.

Definition: A tree is a connected graph with no cycle.

Definition: A spanning tree of a graph G is a sub-graph of G that is a tree and that includes all nodes of G.

Si

Sc

j

G

Spanning Tree

Tree

1

2

3

4

5

7

6

Spanning Tree

2

1 5

4

3

6

7

1

23

4

5

6

7

G = (N, A)

Using the lemma, we can find an algorithm for generating a spanning tree of a connected graph G = (N,A).

2. toGo }.{'' and

)},{('' update ,' and ' with ),( If .3

.3 tocontinue Else

. of treespanning a is }' ,'{' stop. then ,' If .2

}.' ,'{'{},' },{' :any Start with .1

jNN

jiAANNjNiAji

GANGNN

ANGAnNn

1

2

4

31

2

11

22

3 3

4

Proof (that the algorithm above gives a spanning tree.):

tree.spanning a have weThus

. stops algorithm The tree.a have westepeach at Therefore,

.hypothesis inductive by the cycle no is thereWithout

) .on incident arc oneonly is (There cycle. ain

becannot node because cycle, a havecannot Now connected. is

subgraph updated theSo . nodeevery path to a has Similarly,

. path to a has nodeeach

,for arc and ' node addingAfter

lemma. theprove toinduction use will weI.e.,

update. after the treea remainsit that prove We

.3 step of updatean before treea is

)','(' suppose Now tree.a is }) {},({'Initially

N'N

j

jj)(i,

jG'

N'kj

jN'k

N')(iAj)(i,NNj

ANGnG

Note:

In each step, we add one node and one arc. Since we start with one node and no arc, the algorithm must end with |N| nodes and |N |-1 arcs. We can use this algorithm to show that if G = (N, A) is connected then |A| |N|-1 and G is a tree if |A| = |N|-1.

Minimum Weight Spanning Tree (MST)

For G = (N,A), let each arc (i, j) has some weight wij. We want to find a spanning tree with the minimum sum of arc weights.

Definition

Any sub-tree of an MST is called a fragment.

Definition

An arc having one node in a fragment and the other one outside the fragment is called an outgoing arc.

Fact

Given a fragment F, let =(i, j) be a minimum weight outgoing arc form F where i F and j F. The F extended by the arc and node j is fragment.

F

MST4

5

22

j

4

6

3

2

2 3

1

1

i

Kruskal’s Algorithm

Start with each node being a fragment. Order the arc weights in an ascending order w1, w2, w3, …,w|A|.. At each iteration, add an arc that has the minimum weight out of all arcs that are outgoing from all of the fragments.

1 6

58

49

23

7

Kruskal’s algorithm can be made distributed if MST is unique. (Each fragment adds the minimum weight outgoing arc and includes the fragment on the other side of that arc.)

If MST is not unique, Kruskal’s algorithm cannot be made distributed.

Prim-Dijkstra Algorithm

Start with a single node (which is a fragment).

.2 toGo )}.,{(' and }{'' :follows as Update

. from arc outgoing weight minimum a be ),(Let .3

3. togo Else stop. ,' If .2

.arbitrary for }) {},({)','( .1

jiAAjNNF

Fji

NN

nnANF

1 6

58

49

2 3 7

The resulting MST may not be unique.It all arc weights are distinct, the MST is

unique.

Directed Graph (Digraph)

For a graph G = (N,A), let N be a finite non empty set of nodes and let A be a collection of ordered pairs of distinct nodes from N. (There is no loop as before.)

Definition Given a digraph G = (N, A), a graph G´=(N´, A´) is (undirected) associated

graph if N´= N and (i, j) A´ if (i,j) A or (j,i) A.Definition

Walk, path, cycle in a digraph are those of the associated graph.Definition

(n1, n2, n3, …, nl ) is a directed walk in G if (ni, ni+1) is a directed arc in A for i=1, 2, 3,…, l-1.

1

2

3

4

Definition

G= (N, A) is strongly connected if for each pair i N and j N, there is a directed path from i to j.

Fact

G is connected if the associated graph is connected.

Definition

For a given digraph G = (N, A), assign to each arc (i, j), some number dij as the length of the arc. The length of a path is the sum of the lengths of the arcs that comprise the path.

Problem Definition

Shortest Path Problem is a problem that finds, for any two given nodes i and j, a path with the shortest path length.

Bellman – Ford Algorithm

At each step, calculate the shortest path using h or less arcs.

At the end of the first step, we have the shortest paths using one arcs.

At the end of the second step, we have shortest paths using two or less arcs.

And so on.

Example:

Example

1

1

1

4

4

4

2

2

2

23 5

8

1

1

43

5

42

112 )(D

)(14D

)(15D

413 )(D

011 )(D

021 )(D

122 )(D

924 )(D

625 )(D

223 )(D

. all 0 and 1 allfor

,definitionby Also .1 allfor d Hence arcs. less

or using node to1 node fromlength path shortest thebe Let

node. source theis 1 node Assume .),( if Let

)(1

)0(1

)1(

)(

hDi

DiD

hi D

Ajid

h

iii

hi

ij

G = (N, A)

031 )(D

132 )(D 93

4 )(D

233 )(D 43

5 )(D

142 )(D 84

4 )(D

041 )(D

243 )(D 44

5 )(D

. allfor )(min:Use

.1 allfor

. allfor 0

)()1(

)0(

)(1

jidDD

iD

hD

jih

jAj

hi

i

h

A path can contain at most |N|-1 arcs. Thus is the shortest path length from 1 to i. If for all i, is the shortest path length and we can stop the algorithm.

At step h, the algorithm finds the shortest ( h) walk from 1 to i subject to constraint that node 1 is not repeated in the walk. Other nodes can potentially be repeated, however. If there is a negative length cycle, the algorithm will not converge. (We can have negative length arcs, but cannot have negative length cycles.)

To get the shortest path, solve Bellman – Ford equations:

1 allfor

idDD jijAj

i )(min

)( 1NiD

)()( hi

hi DD 1 )(h

iD

Dijkstra’s Algorithm

All arc lengths must be positive

Dijkstra’s algorithm is less computationally intensive than Bellman – Ford

1. step toGo

such that for Update2. Step

2. togo Else stop. If Set

such that Find 1. Step

1for and 0,1 Initialize 0. Step 11

],min[

}.{

.min

.},{

ijijj

j

jPj

i

jj

dDDD

PjD

NPiPP

DDPi

jdDDP

Example

Initial :

],min[

],min[

],min[

}..{:

,,},{.,

25255

24244

23233

543211

9

2

21 thereforeand 21

411 means This 0Set

dDDD

dDDD

dDDD

Pist

DDDDPdDD jj

1

2

1

.min jPj

i DD

1

1

1

4

4

4

2

2

2

23 5

8

Find shortest path from node 1.At each step, we need to use

4

9

321 thereforeand 32

35355

34344

],min[

],min[

}.,,{:

dDDD

dDDD

Pind

8

5321 thereforeand 53

54544

],min[

}.,,,{:

dDDD

Pird

1

11

2

3

1

3 52

1

2

1

Stop. .54321 thereforeand 54 },,,,{: Pith

1

1

32

5

4

4

1

2

Floyd-Warshall

Find shortest path between al pairs of nodes at the same time.

Arc lengths can be negative.

No negative length cycles allowed.

Iterate on the number of intermediate nodes in a path:

path. on the nodes

teintermedia as used becan 1,2,3...n nodesonly that

constraint with j toi fromlength path shortest where

all ],min[

1,...,2,1,0For

. such that , allfor

)(

)()1(

)()1(

)()1(

)0(

nij

njn

nni

nij

nij

ijij

D

jiDDDD

Nn

jijidD

Distributed Bellman - Ford

Finds the shortest path to a given node (say node 1) from all other nodes.

Bellman-Ford algorithm is given by

Where N(i) is set of all neighbors of i.

Note:

An alternative (perhaps better known) version is to find shortest path fro node 1 to all other nodes. (1) can be implemented in distributed manner.

1. All nodes will execute (1) simultaneously.

2. Exchange the results of the computation with their neighbors.

3. Execute (1) again with index h incremented by 1.

Initial condition:

0

1 1

11

1

)(

)(

)(

)( )()(min

h

hjij

iNj

hi

D

iDdD

0

1 allfor 0

1

0

)(

)(

D

iDi

Algorithm will end in N-1 steps.

It is not easy to synchronize all the nodes.

If the network changes (e.g., due to failures), re-synchronization is needed.

Distributed version of Bellman-Ford. Execute indefinitely the following :

Where is the latest estimate received from node are

the latest arc length. More formal notation: At each time t, node i has

)(min)(

jijiNj

i DdD

jD sdiNj ij ')( and

).( allfor has nodeEach

)(1with and allfor 0)(

allfor 0)(

iteration. Ford-Bellman toaccording nodeat computedlast which was

node of distanceshotest theof estimate the:)(

. with atedcommunicatlast which was)(

node of distanceshortest theof estimate the:)(

01

01

iNjdi

iNi tttD

tttD

i

itD

iiNj

tD

ij

i

i

ij

Main Components of Distributed Bellman-Ford

changed. is estimate no and idle is Node 3.

y.accordingl updates and

nodeat updated which was value the

neighbors moreor one from receives Node 2.

unchanged. allfor leaves and

)]([min

by updates Node .1

)(

(t)D

jDN(i) j

i

N(i)j (t)D

tDd(t)D

(t)Di

ij

j

ij

ijij

iNji

i

Main Assumptions for Proper Working of Distributed Bellman-Ford Algorithm

Nodes never stop updating the own estimates and receiving messages from all their neighbors.

All initial node estimates and are non negative.

All estimates communicated to node before initial time and received

after are non negative.

Old distance information is eventually purged.

With the assumptions above:

Proposition : There is a time such that for all

Di is the correct shortest distance.

)(tDi 0(t)D i

j

0t

0t

mt ....,3,2,1for )( , NiDtDtt iim

Routing in ARPANET

Original version (1969)

Length of each link depends on the number of packets waiting in the link queue.

Nodes exchange information on the shortest distance every 625 msec.

Bellman-Ford algorithm is used for shortest paths.

Prone to oscillation.

Modified version (1979)

Length of each link is proportional to the average delay.

Link length is updated every 10 seconds.

Nodes exchange information on the length of their outgoing links every 60 seconds.

Shortest path is calculated by Dijkstra’s algorithm.

Prone to oscillation.

Optimal Routing

We want to quantify the traffic congestion in a network in terms of the arrival processes.

One common performance measure is the sum of packet delays.

Formal representation of network performance:

queue. M/M/1an isit that assumption on the based link of queue

system in the packets ofnumber average thegives formula This

.link on delay n propagatio and processing

asunit same in the link ofcapacity

of Example

function.cost increasinglly monotonica a bemust

link on rate arrival traffici.e., ,link on flow where

)(,

(i,j)

(i,j)d

.F(i,j)C

Fd-FC

F)(FD

:D

D

(i,j)(i,j)F

FD

ij

ijij

ijijijij

ijijij

ij

ij

ij

jiijij

Notation

Pp:(i,j)pij

ij

wp

Ppwp

p

w

p

w

xF

F(i,j)

WwPpx

W wrx

PW,pw:{x

Wwr

px

Ww

P

W

link. thisg traversinflowspath of sum theis, i.e., ,link of flow Total

and allfor 0

allfor

satisfymust } Clearly

pair ODfor rate arrival ficinput traf

path on flow fractional

.pair OD of nodes

ndestinatio and nodeorigin theconnecting paths directed all ofset

pairs. (OD)n destinatio -origin ofset

Optimal Routing Problem

Minimize

),( ),(:ji Pjippij xD

vector.flowpath thebe Let

321

321

321

:as numbered are paths the

pair, ODeach for and pairs, OD ofset thebe Let

W wallfor

allfor 0 : Subject to

3

2

1

111

23112

1

321

m

m

w

n

mm-m-m-w

w

w

m

Ppwp

wp

x

x

x

x

x

),...n,n,n(nP

),...n,n,n(nP

),...n,,(P

},...,w,w,w{wW

rx

W,wPp x

.path oflength )(

then ,link oflength let the weIf

.)(

:Note

. where

0and

0

, and allfor if and problem routing theof

sconstraint thesatisfiesit if vector flowpath optimalan is Then

'

link containing

paths allpath on

links all

link containing

paths all),(

px

xD D(i,j)

xDx

xD

xDD(x)

x rx

)x(xx

)D(x

WwPp

x

p

ij

(i,j)

p

p(i,j)

ijp

(i,j)

pji

ij

pwPp

p

Ww Pppp

p

w

w

w

Optimization Lemma

Let f be differentiable convex function of the n- dimensional vector

x={x1, x2 , … xn). Let. X be a convex set of vectors. Then x* X is an optimal solution

of the problem:

Minimize : f(x)

Subject to : x X

.0* ),(

as written also becan condition The

. where

allfor 0)( iff1

xxxf

x

f(x)

x

)f(x

Xxxxx

)f(x

xxii

n

iii

i

Proof

).g( ofconvexity thecontradit (2) and (1) equations The

(2) 0)()(

)0(

Also

(1) ~1 0 Then,

~)Consider

~such that ~Let

minimum.not is but ,0)()(

suppose Conversely

.0)()()()0(

[0,1].over minimum is 0 Therefore,

[0,1]. where))-(()let ,every For

solution. optimalan is Suppose

1

1

10

ii

n

i i

ii

n

i i

ii

n

i i

xxx

xfg

)xf()g()f(x)g(

))-xxα(f(xg(α

).f(x)xf(Xx

xxxx

xf

xxx

xfgg

)f(x)g(

xxxfg(αXx

x

0)()0(

)0()(

0. and 0Let

)()(

)()(

)()(

)()(

)()()(

)()(

)()()()( Then,

1,0,)()()()-(1

. 0 allfor (0) )( then,

,0)0(

and abledifferenti andconvex is )( If

122

21

1212

121

112

1211

12

112112

12121

1

ggg

ggg

ggg

t

gtggg

gggg

ggg

gg

gg

Optimality Condition

(2). allfor 0)()(

to(1)simplify can We

. allfor and allfor 0, satisfying allfor

(1) 0)()(

AND , allfor

, allfor and allfor 0

if vector flowpath optimalan is

Wwxxx

xD

PpW wrxxx

xxx

xD

Wwrx

WwPpx

xx

pp

xPp p

ww Pp

ppp

Wwpp

xPp p

w Pp

p

wp

p

w

w

w

w

Proof

true.also is Converse (2). useThen

.or pathsother allfor and , and 0Let

.'path other some and 0for which path Consider

(Proof)

.' allFor

)()( iff 0 allFor

further. (2)simplify can We

. and ' allfor let each For

""'

'

'

p'pp"xxx, x-xx

pxp

Pp

x

xD

x

xDW, xw

Piww xxW,w

pppppp

p

w

xpxpp

wii

Summery of Optimal Condition

A set of path flows is optimal iff path flow is positive only on the paths with a minimum first derivative length. Also at the optimum path flow vector, the path along which the traffic rw is split must have equal length.

Example:

r

x1

r

x2

C2

C1

ies.possibilit twoare There

.(obvious) havemust We.Let

.0 ,0 , as wellas

satisfied bemust (1)condition path shortest theoptimum, At the

. and )( where

)()()(

2121

2121

21

2211

xxCC

xxrxx

CCrxC

CxD

xDxDxD

ii

iii

(2), and (1) From

(2) . also and

(1) )()(

Therefore,

.)()(

and )()(

havemust we true,be toFor this .0 ,0 2. Case

1

)(

)()(

havemust we true,be toFor this .0 , .1 Case

21

21122

21

21211

21

222

22

11

1

2

2

1

1

21

21

2112

21

1

21

21

CC

CCCrCx

CC

CCCrCx

rxx

xC

C

xC

C

x

xD

x

xD

x

xD

x

xD

xx

CCCrCrC

C

x

xD

x

xD

xrx

xxxx

xx

211 CCC 21 CC

2x

1x

21 and xx

r

Example

r r

1x

2x

3x

possible.not is which 022 0 , ngSubstituti

.2,1for

havemust we true,be toFor this

0 ,0 1. Case

ies.possibilit threeare There

. havemust we,)( ofsymmetry theFrom

. allfor 0 and ,

2)(

213

3

3213

21

321

323

22

21

r,xxrx

ix

D(x)

x

D(x)

rxxxx

xxxD

ixrxxx

xxxxxD

xix

i

3

1 ,

3

2

3

222

22

,2

ngSubstituti

3

havemust we true,be toFor this2

, .3 Case

.222

2 2

,0 ngSubstituti

.2,1for

havemust we true,be toFor this2

,0 2. Case

213

321

21

213

213

3

213

rxx

rx

rr

xr

xx

x

D(x)

x

D(x)

x

D(x)

rxxx

rrr

xxx

ix

D(x)

x

D(x)

rxxx

xxx

xix

1 2 3 4

1

2

3x

21 xx

r

321 ,, xxx

Feasible Direction Methods

In optimal routing, there is a positive flow only on those paths with minimum first derivative length (MFDL).

A path flow vector x is strictly sub optimal if there is a positive flow on a non NFDL path.

Given a sub optimal path flow vector, we can improve the routing by shifting flows from non MFDL paths to MFDL paths.

A path flow vector x={xp} that satisfies the constrains

Suppose we have a non MFDL path flow vector x which is feasible.

vectorflowpath fesible a called is allfor 0

allfor

wp

wPp

p

PW, pwx

Wwrxw

.0,for feasible also bemust vector new The

.by change We

xx

xxx p

Wwx

Wwrxxxx

Wwxx

Δxx

xD

xWwPpxWwx

pPp

wpPp

p

Ppp

Ww Ppp

p

pwpPp

p

w

w

w

w

w

allfor 0

allfor )( feasiable. also is

. allfor r feasiable. is

:1 of Proof

flow. ofon conservati thecalled is 1condition The

0)(

.2

.0for which , allfor 0 and allfor 0 .1

have,must werouting, theimprove toas wellas vector feasiable a be tofor thisorder In

w

).( of gradients thecalled is )(

direction.descant a is say we true,is 2condition theIf

negative.non be for 0 havemust we0, if Moreover,

2

1

xDxDx

D(x)

x

D(x)

x

xxxxx ppppp

x

Constraint Set

Feasible descent directions at x

Surface of equal cost

x

)(xD

Repeat.

feasible. is and )(such constant afor Let

direction.descent feasible a be and feasible, be Let

Algorithm Iterative

MFDLnon oneleast at for 0

MFDLnon arepath that allfor 0

flow. ofon conservati thesatisfiesIt

:such that choosingby thisachievecan We.function theofpoint minimum a find want to We

xxxDx)D(xxxx

xx

x

x

xD(x)

p

p

Frank – Wolfe Method

path.shortest path toshortest non from shifted is Flow

. ~ where onefor except 0,each For

:Note

Repeat. 3.

., allfor ))~(~ ~Set

)).~(~(min))~(~(such that valuea be Let 2.

path. NFDL thealong routed is input all if vector flowpath thebe Let 1.

.~at evaluated are sDerivative

pair. ODeach for path length derivativefirst minimum a find want toWe

vector.flowpath feasible a is x~ Assume

]1,0[

ppwp

wpppp

w

p

rxPpxWw

WwPpxxxx

xxxDxxxD

rx

xx

D(x)

descent. ofdirection optimalan is ~ resulting The

satisfied. is (1)such that )(

Minimize

:becomes problem The

negative.most is change of rate initial the, alongsuch that

(1) allfor 0

allfor 0

i.e.,,direction flowpath feasible a utilize willWe

,

xxx

xx

xD

x

WwPpxx

Wwx

x

pWw Pp p

wpp

Ppp

w

w

Example

0.for impossible is which 55.01.02

1 , ,

2

-1 ngSubstituti

.3

havemust we true,be toFor this .2

-1 , 2. Case

O.K. is which 0.552

1

2

1 0, ngSubstituti

.2,1for havemust we true,be toFor this .2

1 ,0 .1 Case

es.possibliti twoare There . havemust we, )( ofsymmetry theFrom

55.01.02

1)( :Given

. 3,2,1for 0 and 1

321

21213

213

3213

21

323

22

21

321

xxx

x

D(x)

x

D(x)

x

D(x)xxx

xxx

ix

D(x)

x

D(x)xxx

xxxD

xxxxxD

ixxxx

xxx

xix

i

r=1 r

1x

2x

3x

(0,1,0).or )0,0,1( is flowpath shortest

. if ispath shortest

. if ispath shortest

feasible. is vector theif 0.55 and 0.55 havecannot We

.55.0 because becannot path shortest Obviously

.55.01.0 , ,

Then, vector.feasible a be Let Algorithm. Wolfe-Frank apply the want toWe

212

211

21

11

33

33

22

11

x

xxx

xxx

x

D(x)x

x

D(x)x

x

D(x)x

xx

D(x)x

x

D(x)x

x

D(x)

x

21

3

312

1

3

2

1

3

2

1

3

2

1

21

3

2

321

3

2

1

3

2

1

3

2

1

if

)1(

)(

)1(

1

if

)1(

)1(

)(1

:equations Update

xx

x

xxx

x

x

x

x

x

x

x

x

x

x

xx

x

x

xxx

x

x

x

x

x

x

x

x

x

Topological Design

Design a data network at minimum cost while maintaining the performance requirements.

Assumptions

Locations of terminal are known.

Traffic level between each OD pair is known.

Selection

Topology of a communication sub network meeting the traffic demand.

Topology of local access network

Objectives

Satisfy the average delay per packet constraint for a given routing scheme.

Satisfy the reliability constraint with respect to node and link failures.

Minimize cost.

T T

TT N

N N

N

C

C

Access Network Subnet

Subnet Design

Given

Location of nodes

Traffic demand between nodes

Selections

Capacity of links

Flow of links

Simplifications

No reliability consideration

Linear cost

.cost Optimal . 1

11

,constraint theFrom y.numericall for now solvemust We

0Set

.multiplier Lagrange is where1

Let .1

i.e., ,constraint theofboundary at the be willsolution The

),(),(

2

),(),(),(

),(),(),(

*

jiijij

nm

mnmnijijijij

ji

ijij

ji

ijij

ji ijij

ij

ij

ijijij

ij

ji ijij

ij

jiijij

ji ijij

ij

ij

CPr

FP

r

FP

TFC

r

FP

Tr

FP

FC

F

rT

rP

FFC

dC

dL

FC

F

rCPLT

FC

F

r

C

.1

delay average given the , Minimize

) network, theinto arrival total: ),,(link per cost :( s'Known :1 Problem

),(),(

ji ijij

ij

jiijij

wijij

TFC

F

rCP

rrrjiPF

solve toproblemDifficult -

exist. solutions localMany -

convex.globally not isfunction This

0Set

unknown. are s' :2 Problem

)

(i,jijij

ij

ij

CPdF

d

F

Local Access Design

Problem 1

Sub network is available.

Design a network that connects a set of terminals to

. techniquegprogramminlinear aby solved becan problem This

. and 1Given

Minimize

otherwise 0

or concentrat toconnected is source if 1 -

or concentrat toconnected becan that nodes ofnumber maximum : -

or concentrat to source connecting ofcost : -

orsconcentratknown ally geographic -

sourcesknown ally geographic -

11

n

1i

m

1j

m

ijij

m

jij

ijij

ij

j

ij

j kx i x

xa

jix

jk

jia

m

n

problem.er much tough is This

. techniqueialcombinator aby solved becan problem This

. and 1Given

Minimize

none is thereif 0

siteat or conecntrat a is thereif 1 -

siteat or concentrat a locate cost to : -

ors.concentrayfor locations potential -

fixed.not orsconcentrat of locations -

sourcesknown ally geographic -

2 Problem

11

1 11

jykxix

ybxa

jy

jb

m

n

jj

n

jij

m

jij

n

i

m

jjj

m

jijij

j

j