Survivable Network Design

David Tipper
Associate Professor
Department of Information Science and
Telecommunications
University of Pittsburgh
Telcom 2110 Slides 15

Survivable Network Design

• Spare Capacity Allocation (SCA) Problem:

– given working paths and network (or virtual network) topology
– provision spare capacity and find backup routes for fault tolerance
– Goal: minimum spare capacity or cost

• Survivable Mesh Networks

– Consider preplanned protection in mesh networks
• STM - DCS, ATM - VP, WDM, MPLS, etc..

– determine routing/capacity allocation for normal demand

– find location and amount of spare capacity for failure scenarios
– spare capacity required depends on restoration/survivability technique

1
 Classification of
Survivability Techniques

• Path-based (Global) versus Link-based (Local)

• Failure Dependent vs. Failure Independent
• Protection versus Restoration
• Dedicated-Backup versus Shared- Backup Capacity
• Ring versus Mesh topology
• Dual and multi-homing
• P cycle
• Etc.

Failure Dependent vs. Failure Independent

• Failure Dependent – the backup path depends on which

device fails – need a set of paths one for each failure case
• Failure Independent – backup path link and node disjoint
with working path - one backup path per working path
• Example:
7
9
13
Failure Dependent backup path
for link 2-3 failure
1 2 3

Working path 6
12
10 5
Failure Dependent backup 4

path for link 1-2 failure

11 8

Failure Independent backup path

2
 SCA Problem
• SCA for Failure Independent Shared Backup Path
Restoration
• Notation
r = 1,2,…, D set of demands (source-destination pairs)
p = 1,2,…, Pr set of possible paths for demand pair r
l = 1,2,…, L set of network links
• Input parameters (constants)
αr offered traffic load of demand pair r
cl unit cost of capacity on link l
δl r,p = 1 if l belongs to path p realizing demand r
= 0, otherwise
f set of link failure scenarios
• variables
x r,p flow of demand r on path p
sl spare capacity on link l

SCA Path-flow model

Find sl and x r,p , which

Total spare capacity
minimize ∑ cl ⋅ sl
l∈L Single backup path for each flow

s.t. ∑x r, p
= 1, ∀r ∈ D
p∈P r
Enough spare capacity on each link

∑α ∑ δ
r∈D f
r
l
r, p
⋅ x r , p ≤ sl , ∀l ∈ L − { f }, ∀f , f ∈ L
p∈P r

3
 Matrix Based Formulation of SCA

• Matrix Based formulation of Optimization model for FID

shared backup path restoration*
• Consider path incident matrices P and Q for working and
backup paths where each matrix has
number of rows = number of flows in the network
number of columns = number of links in the network
– row i in the matrix P corresponds to the set of links used by flow i
– where pij = 1 if flow i uses link j it is 0 otherwise
– similary row i in the matrix Q corresponds to the set of backup path links used
by flow i where qij = 1 if flow i uses link j it is 0 otherwise
• Relate to spare provision matrix G, and spare capacity reservation s
– G = QT P, element Gij gives required spare capacity on link i when link
j fails
– s = max(G), or s ≥ G , spare capacity reservations are the maximum
spare capacity for any single link failure
• * Y.Liu, D.Tipper, and P. Siripongwutikorn, “Approximating Optimal Spare Capacity
Allocation by Successive Survivable Routing,'' ACM/IEEE Transactions on Networking, Vol.
13., No. 1, pp. 198-211, Feb., 2005

Example
From working and
Link i 1 2 3 4 5 6 7
backup paths, G= QT P Backup path link
incident matrix
s G QT
From G, 1 2 0 2 1 1 1 0 1 0 0 1 1 0 1 1 0 0 0
2 2 2 0 2 1 1 0 0 1 1 0 0 0 1 1 0 0 0
s=maxG 3 1 0 1 0 0 0 1 1 0 0 1 0 0 0 0 0 1 0
4 1 1 1 1 0 0 1 0 1 0 0 1 1 0 0 0 1 0
5 2 1 1 1 0 0 0 2 0 1 1 0 0 0 0 0 0 1
6 1 0 0 1 0 1 0 1 0 0 0 0 1 0 0 1 0 1
Working path 7 2 1 0 2 0 2 0 0 0 1 0 0 0 1 0 1 0 0
Backup path 11
Flows 1 2 3 4 5 6 7 8 9 10
src dst
1 0 0 0 0 0 0 1 a b
1 0 1 0 0 0 0 2 a c
b 3 c P 0 1 0 0 0 0 1 3 a d
1 0 1 0 0 0 0 0 4 a e
Working path link 0 0 1 0 0 0 0 5 b c
a 4 6 5 incident matrix 0 0 1 0 1 0 0 6 b d An example:
2
0 0 0 1 0 0 0 7 b e when link 2 fails,
0 0 0 0 1 0 0 8 c d
0 0 0 0 0 1 0 9 c e how much capacity is
e 7 d
0 0 0 0 0 0 1 10 d e need on link 1 ?

4
 Matrix Based SCA for Link Failures
min S = eT s Total spare capacity
Q,s
s.t. s≥G Enough spare capacity on each link
G = QT M P Calculation of spare provision matrix
P+Q≤1 Link-disjointed backup paths
Q BT = D (mod 2) Flow conservation of backup
Q is a binary matrix Integer programming
Decision variable: Q, s
Given: M – traffic demand matrix
P – working path link incidence matrix
B and D – node-link & flow-node incidence matrices

Another way to find G

G = Σr Gr, where Gr = qrT pr , pr and qr are

vectors for working and backup paths of flow r

Links G

Failures
G1
Q
G2
…

Flows
P GR-1
GR

5
 Find spare capacity s

Link i 1 2 3 4 5 6 7

From QT and w i =sum P(*,i ) 2 2 3 1 2 1

Total working
2 13
Spare path and link
incident matrix
QT

P, get G 1
s =max(G)
2 0 2
G=QT P
1 1 1 0 1 0 0 1 1 0 1 1 0 0 0
2 2 2 0 2 1 1 0 0 1 1 0 0 0 1 1 0 0 0
3 1 0 1 0 0 0 1 1 0 0 1 0 0 0 0 0 1 0

From G, get s 4
5
1
2
1
1
1
1
1 0
1 0
0
0
1
0
0
2
1
0
0
1
0
1
1
0
1
0
0
0
0
0
0
0
1
0
0
1
6 1 0 0 1 0 1 0 1 0 0 0 0 1 0 0 1 0 1
7 2 1 0 2 0 2 0 0 0 1 0 0 0 1 0 1 0 0
b 3 c Total spare 11
Flows 1 2 3 4 5 6 7 8 9 10
1 src dst
1 0 0 0 0 0 0 1 a b
a 4 6 5 1 0 1 0 0 0 0 2 a c r=2
0 1 0 0 0 0 1 3 a d G2
2 0 1 0 0 0 0 0 4 a e 0 0 0 0 0 0 0
Working path and link 0 0 1 0 0 0 0 5 b c 1 0 1 0 0 0 0
e 7 d incident matrix 0 0 1 0 1 0 0 6 b d 0 0 0 0 0 0 0
P 0 0 0 1 0 0 0 7 b e 0 0 0 0 0 0 0
0 0 0 0 1 0 0 8 c d 1 0 1 0 0 0 0

For r =2, find 0

0
0
0
0
0
0
0
0
0
1
0
0 9
1 10
c
d
e
e
0
1
0
0
0
1
0
0
0
0
0
0
0
0

G2=qrT pr

Approximation algorithm

• Decomposition
– multi-commodity flow Æ multiple single flows
• Using shortest path algorithm for each flow to
– route link-disjointed backup paths
– using spare provision matrix G to calculate
link cost – incremental spare reservation vr ;
• Flows successively update their backup paths
Æ termed: successive survivable routing (SSR)

6
 Link cost and local objective

• Goal: Each flow seeks a new backup path with

minimal additional reservation
• Additional reservation as link cost:
– Let G+ = (e-pr)T pr , and s+= max (G++G)
• (e - pr) assumes that a backup path uses all possible links
• s+ is a temporary spare capacity reservation vector
– Additional spare reservation vr = s+ - s
• vr tells how much additional spare capacity needed if a link is
used on a new backup path
• Run shortest path with link cost vr

Example of link cost

Link i 1 2 3 4 5 6 7
Total working Spare path and link
w i =sum P(*,i ) 2 2 3 1 2 1 2 13 incident matrix Assume backup path are
QT using all possible links
s=max(G) G=QT.P qr+ = (e-pr)
1 2 0 2 1 1 1 0 1 0 0 1 1 0 1 1 0 0 0 0
2 2 2 0 2 1 1 0 0 1 1 0 0 0 1 1 0 0 0 1
3 1 0 1 0 0 0 1 1 0 0 1 0 0 0 0 0 1 0 1 Find the contribution
4 1 1 1 1 0 0 1 0 1 0 0 1 1 0 0 0 1 0 1
1
G+ = (e-pr)T pr
5 2 1 1 1 0 0 0 2 0 1 1 0 0 0 0 0 0 1
6 1 0 0 1 0 1 0 1 0 0 0 0 1 0 0 1 0 1 1
7 2 1 0 2 0 2 0 0 0 1 0 0 0 1 0 1 0 0 1 Find vr=s+ – s,
Total spare 11 where s+ = max(G+G+)
Flows 1 2 3 4 5 6 7 8 9 10 11
src dst m
1 0 0 0 0 0 0 1 a b 1 Find qr, using shortest path
1 0 1 0 0 0 0 2 a c 1 r = 11 routing with link metric vr
0 1 0 0 0 0 1 3 a d 1 G+ v11 q11
0 1 0 0 0 0 0 4 a e 1 0 0 0 0 0 0 0 ∞ 0 ∞ b 0 3 c
Working path and link 0 0 1 0 0 0 0 5 b c 1 1 0 0 0 0 0 0 1 1
1
incident matrix 0 0 1 0 1 0 0 6 b d 1 1 0 0 0 0 0 0 0 1 1 0 0
P 0 0 0 1 0 0 0 7 b e 1 1 0 0 0 0 0 0 1 0 a 4 6 5
0 0 0 0 1 0 0 8 c d 1 1 0 0 0 0 0 0 0 0 1
0 0 0 0 0 1 0 9 c e 1 1 0 0 0 0 0 0 0 1 2
0 0 0 0 0 0 1 10 d e 1 1 0 0 0 0 0 0 0 0 0
0 0 0 1 0 0 0 12 b e 1 e 7 d
1 0 0 0 0 0 0 11 a b 1

7
 SSR flowchart of flow r

• On source node of flow r:

1. Given pr and dr
– pr , qr : working and backup
path vectors
2. Periodically update G
– dr : destination node
– G, s: spare provision matrix
3. Calculate vr
and spare reservation vector
– vr : incremental spare
4. Update qr using vr
reservations as link cost
5. Update s, and G • Stop after no backup path
update on the network

Complexity

• Polynomial running time

– shortest path algorithm for each flow, O(N2)
– Limited backup path update iterations for each
flow
• Polynomial space complexity
– Advertised information in O(L2)
– No per-flow based information

8
 Numerical comparison

• Compare different algorithms and bounds

– RAFT: Resource aggregation fault tolerance
– SPI: Sharing with partial information
– SR: Survivable routing (SSR without iteration)
– SSR : Successive survivable routing
– SA: Simulated annealing
– BB: Branch and bound on a path-flow model – optimal
– LP: Linear programming lower bound
• Metrics:
– % Redundancy = spare capacity/working capacity,
– execution time

Experiment networks

1 2 3 4
1 2 13 9 7
1 10 9 7
13
17 4 2 3
4 5 2 12 11 9 6
4 2 3
6 12
3 6 12
10 4 5
8
7 8 3 10 4 5
16
6 11 8
4 7 14
9 10 11 8
5
15

5 6 7 8 1 2 4 9 14

13 17 26
8 46
3 13
8
2 3 4 5 6 7 8
6 5
1
2
3 12 16
1 21 20 47
18 45
17 16 11 16 7 38
5 13 20
9
22 21
6 25
5 23 15 48
10 8
14 15 17 22
9 1 24
14 18
39
15 18 19 23 10 49 50
19 19 27
3 17 9
2 12 6 24 44
7 12 11
4 25 11 20 23 28 40
4 16
43
18 7 10 26

15 14 13 12 11 10 22 29 41
32 42
36
30
21 33
34
35
31 37

Network node degree ranges from 2.31 to 4.4

Consider balanced mesh load case

9
 Redundancy versus Time
on Network 3
75
• SSR, SR, have 64 70 Fast response
Network 3
Worse solutions,
random cases with fast
65
different flow orders
60 RAFT

Redundancy (%)
• Range of solutions Near optimal
55 SPI
solutions, fast
• Time is the sum of
50
time to compute all Better solutions,
45
slow, not scalable
64 cases
40 SR SSR SA
LP BB
35
Infeasible
30
-2 0 2 4
10 10 10 10
Time (second)

Typical SSR results

100
LP BB SA SSR SR SPI RAFT
90

80

70

60
Redundancy

50

40

30

20

10

0
1 2 3 4 5 6 7 8
Networks

10
 Multilayer Networks
• Backbone networks have multiple technology layers
• Converging toward IP/MPLS/WDM
• Typically have survivability at each layer
• Multiple Layers present several survivability challenges
• Coordination of recovery actions at different layers
– Which layer is responsible for fault recovery?
• Spare Capacity Allocation (SCA)
– How to prevent over allocation, when each layer provides spare resources?
• Failure Propagation
– Lower layer failure can affect multiple higher layer links!
3

1 MPLS connections
5 WDM Physical Path

2 3

4 5

Model Arbitrary Failures

• To model failure propagation need to model multiple link failures
• F failure matrix, fik =1 Î link k fails in the ith failure scenario
• U flow failure incidence matrix - how flows affect by ith failure scenario
U = P ☼ FT , ☼ binary multiply, capture logical relations
• T flow tabu-link matrix of links avoid in determining backup
T=U☼F
⎡1 0 ⎤
⎢0 0 ⎥
⎡1000000 ⎤
Scenario 1 b 3 c
⎢0100001 ⎥ ⎡1000000 ⎤⎢ ⎥ ⎡10 ⎤
⎢ 0100001 ⎥ ⎢ 0 0 ⎥ ⎢ 01 ⎥
P=⎢ ⎥
1
U = ⎢ ⎥ ⎢0 0 ⎥ = ⎢ ⎥
⎥⎢ ⎥ ⎢ 01
a 4 6 5
⎢0010100 ⎥ ⎢ 0010100 ⎥
2
⎢ ⎥ ⎢ ⎥ ⎢0 1 ⎥ ⎢ ⎥
⎣ 0100010 ⎦ ⎢ 0 0 ⎥ ⎣ 00 ⎦
e 7 d ⎣0100010 ⎦ ⎢ ⎥
⎢0 1 ⎥
Scenario 2
⎣ ⎦

⎡1000000 ⎤ ⎡10 ⎤ ⎡1000000 ⎤

F=⎢ ⎥ ⎢01 ⎥ 1000000 ⎢ ⎥
⎣0000101 ⎦ ⎡ ⎤ ⎢0000101 ⎥
T = ⎢ ⎥⎢ =
⎢01 ⎥ ⎣0000101 ⎥⎦ ⎢0000101 ⎥
⎢ ⎥ ⎢ ⎥
⎣00 ⎦ ⎣0000000 ⎦

11
 SCA model for arbitrary
failures
min S = eT s Total spare capacity
Q,s
s.t. s≥G Enough spare capacity on each link
G=Q MU T
Calculation of spare provision matrix
T+Q≤1 Failure-disjointed backup paths
Q BT = D (mod 2)
Flow conservation of backup
Q is a binary matrix
U = P ☼ FT Integer programming
T=U☼F Path failure incident matrix
Decision variable: Q, s
Path tabu-link incident matrix
Given: M – traffic demand matrix
P – working path link incidence matrix
B and D – node-link & flow-node incidence matrices
Solve SCA model using Branch and Bound algorithm – NP hard

Heuristic Solution Algorithm

• Successive survivable routing algorithm*
– Decompose multi-commodity flow Æ multiple single flows
– Goal: Each flow seeks a backup path with minimal additional spare
capacity
– Using shortest path algorithm for each flow to
• route link-disjointed backup paths
• using spare provision matrix G to calculate
link cost – incremental spare reservation vr ;
• Flows successively update their backup paths
Æ termed: successive survivable routing (SSR)
• Randomly order flows for successively updating.
• Fast computation find near optimal solution

• *Apparatus and Method for Spare Capacity Allocation, Y. Liu and D. Tipper
, U.S. Patent 6,744,727 B2, June 1, 2004
• Presented in part (IEEE Infocom 2001, IEEE Trans. On Networking Feb.,2005)

12
 Why Optical Layer Protection?
• Backbone networks have multiple technology layers
• Converging toward IP/MPLS/WDM
• Typically have survivability at each layer
• Optical layer provides lightpath services to its client layers
(e.g., SONET, IP, ATM)
• Protection mechanisms exist in the client layers, so why need
protection in optical layer?
– IP and ATM networks don’t have extensive protection functions as
SONET
– Capacity efficient due to protection capacity sharing across multiple
pairs of client layer equipments
– Significant savings in equipment cost
– Handle fiber cuts more efficiently than the client layers
– Provide an additional degree of resilience (e.g., protect against
multiple failures)
– Can use mesh-based protection schemes that require significantly less
protection capacity than ring-based schemes

Limitations of Optical Layer

Protection

• Can’t handle all failures: client equipment failures

must be dealt with by the client layer
• Protect traffic in units of lightpaths: can’t protect
only part of the traffic within a lightpath
• Need pay careful attention to interworking of
protection schemes between different layers

13
 Where are Optical Networks Used?

• Access networks:
– Fat pipes for transporting traffic between multiple end users and
POPs

• Metro networks:
– Fat pipes for interconnecting multiple access networks, and
providing access to backbone networks
• Long haul backbone networks:
– Fat pipes for transporting aggregates of traffic in the backbone
• Grid networks (e-science):
Traffic
– Fat pipes for transferring large files Aggregation
• Storage networks:
– Fat pipes for transferring large files

Optical Network Evolution

Interconnected rings
Evolution of the OXC
OXC and mesh topologies

Optical Network OXC

OADM
Architecture OADM OXC
OADM OADM

OXC
WDM rings OADM OADM OADM
OADM
OADM OADM

OADM OADM MSPP

WDM linear add/drop

OADM OADM

WDM point-to-point
SONET/SDH ADM

1995 2000 2005

14
 Optical Layer Protection Schemes
• Optical channel (OCh) layer (or path layer) protection
schemes
– Restore one lightpath at a time
– Need demultiplex all wavelengths
• Optical multiplex section (OMS) layer (or line layer)
protection schemes
– Restore the entire group of lightpaths on a link
– Require less equipment
• OLTs and OADMs can provide both OCh and OMS layer
protection in linear or ring configurations
• OXCs can provide OCh layer protection in linear, ring, and
mesh configurations
• Backbone networks: use unprotected WDM point-to-point
systems and rely on OXCs to perform the protection functions
• Metropolitan networks: use WDM line terminals and OADMs
to perform protection functions

Optical Layer Protection Schemes

• OMS layer protection schemes

– 1+1
– 1:1
– OMS-Dedicated Protection Ring (DPR)
– OMS-SPRing
• OCh layer protection schemes
– 1+1
– OCh-SPRing
– OCh-Mesh

15
 1:1 OMS Protection

• Shared protection in point-to-point links

• The composite WDM signal is transmitted over
only the working fiber
– Use a switch at the transmitter, instead of a splitter
• If the working fiber is cut, both ends switch over
to the protection fiber
– Need an APS protocol
• Support low priority traffic on the protection fiber
• Allow N working systems to share a single
protection system

OMS-DPRing
• Dedicated protection ring
• Two fibers operate in opposite directions
• Each node transmits on both directions of the ring
– Different nodes must transmit at different wavelengths
• Normal operation: the ring functions as a bus, with
one pair of amplifiers turned off and all the others
turned on
• When a link fails: an amplifier pair next to the
failed link are turned off and the ones that were
originally inactive are turned on
• Equivalent to a Sonet USHR

16
 OMS-SPRing
• Shared protection ring
• Four fibers, analogous to a SONET BLSR/4 (BSHR)
• The two protection fibers do not have attached WDM
equipment
• Use either span switch or ring switch
• Two-fiber version
– Dedicate half the wavelengths on each fiber for protection
purposes
– Make the protection wavelengths on one fiber correspond
to the working wavelengths on the other fiberÎsignals can
be rerouted w/o requiring wavelength conversion

1+1 OCh Dedicated Protection

• Works in point-to-point, ring (OCh-DPRing), and

mesh configurations
• Two lightpaths on disjoint routes are setup for each
client connection
• The client signal is split at the input, the destination
selects the better of the two lightpaths
• No signaling requiredÎfast restoration
• No protection bandwidth sharingÎbandwidth
inefficient

17
 OCh-SPRing

• Shared protection ring

• Similar to SONET BLSR/4, but operate at the
optical channel layer
• Working lightpaths are set up on the shortest path
along the ring
• When a working lightpath fails, it’s restored using
span switch or ring switch
• Non-overlapping lightpaths in the ring can share a
single wavelength around the ring for
protectionÎmore efficient than OCh-DPRing

OCh-Mesh Protection

• Backbone networks are meshed

• For mesh networks, OCh-mesh protection
schemes are more bandwidth-efficient than
rings
– Efficiency improvements range from 20% to
60%

18
 Path-Based v.s Link-Based
Protection

• Path-based: the connection is rerouted end

to end on an alternate path
– Need notify the source node upon a failure
• Link-based: the connection is rerouted on
an alternate path around the failed link
– Need not notify the source node upon a failure
– Enable faster restoration than path-based
schemes

Offline v.s Online Protection

• Offline protection
– Protection path and wavelengths are reserved at the time
of connection setup
• In path-based scheme, a link-disjoint protection path is reserved
• In link-based scheme, protection paths are reserved around each
link of the working lightpath
– Fast and guaranteed restoration
• Online protection
– Search for protection paths using the spare capacity in
the network upon a failure
– Capacity efficient
– Slow and no guarantee of restoration

19
 Dedicated v.s Shared Protection

• An offline scheme can use either dedicated

protection or shared protection
• Dedicated protection: each working lightpath is
assigned its own dedicated protection bandwidth
• Shared protection: if two working lightpaths are
link-disjoint, they can share protection bandwidth
– More capacity efficient than dedicated protection
– Protection lightpaths are set up after a failure occurs

Classification of OCh-Mesh
Protection Schemes

OCh-mesh protection schemes

offline online

Dedicated Shared Link-based Path-based

Link-based Path-based Link-based Path-based

20
 Internetworking between Layers

• Need coordination between protection

mechanisms in different layers
• Bottom-up sequential approach: start at the layer
where the failure occurs, let the layer try to restore
service first, then let the higher layer try
– Option 1: have the restoration in the lower layer happen
so quickly that the upper layer doesn’t detect the failure
– Option 2: impose an additional hold-off time in the
higher layer before it attempts restoration

Challenges in All Optical Networks

• Dense wavelength division multiplexing (DWDM) -- Provide an enormous
bandwidth by allowing simultaneous transmission of data on multiple
wavelength in one fiber
– Concern: a large gap between the capacity of one wavelength channel and the
bandwidth requirement for a connection request
– Solution: traffic grooming

– Concern: a fiber cut or link failure can result in a loss of a large volume of data.
– Solution: mesh restoration

• Wavelength continuity constraint -- A signal must stay on the same

wavelength from its source to its destination. Wavelength conversion is too
expensive.
– Concern: channel network resource utilization is reduced.
– Solution: Wavelength retuning

21
 The Traffic Grooming Problem

• Number of wavelengths per fiber = 4 -100+

• Per wavelength capacity = 2.5 Gbps to10 Gbps
• Bandwidth requirements of most applications << 2.5
Gbps

∴ Group several sessions on the same wavelength channel in

order to better utilize the available bandwidth

Traffic Grooming: The intelligent allocation of traffic

demands onto an available set of wavelengths in a way that
reduces the overall cost of the network.

The Traffic Grooming Problem:

CapEx

• Dominant cost factor: Electronic layer

multiplexing; number of electronic layer
Line Terminating Equipment (LTs):
– SONET/SDH ADMs 3-4 times as
expensive as OXC
– IP/MPLS router ports ports

• Solution: Assign the traffic such that

minimum number of LTs is used

22
 Traffic Grooming Problems

• Network design problem: dimensioning and

network provisioning
– Reduce capital and operational expenditure
– Maximize revenue
NP-Complete Problem
• Solution types:
– Exact solutions (based on ILP or MILP)
– Heuristic and approximate solutions
– Bounds

Traffic Grooming for Ring Networks: Heuristics

Paper Topology Traffic Result

Simmons et al ‘98 BLSR Uniform all-to-all Heuristic
Chiu and Modiano ’00 BLSR Uniform all-to-all Heuristic
Cho et al. ‘01 Hubbed, and Uniform all-to-all SA
single hop
Li et al ‘00 BLSR, single Arbitrary 10/9 approximation
hub
Wan et al. ‘00 BLSR Arbitrary Heuristic
Zhang and Qiao ‘00 UPSR/BLSR Arbitrary Heuristic
Xu et al. UPSR Arbitrary GA
Wang et al. ‘01 BLSR Arbitrary SA
Mustafa & Kamal ’03 UPSR/BLSR Arbitrary Heuristic

23