Performing Gateway Load Balancing in MANETs PhD Dissertation February 10 th 2012 Vinh Pham

1

Performing Gateway Load Balancing in MANETs

PhD DissertationFebruary 10th 2012

Vinh Pham

Vinh Pham, PhD Dissertation 2012

2

Outline• Introduction

– Motivation– MANET– Challenges– Main focus of the Thesis– Thesis overview

• Contributions– Part I: Mobility– Part II: Load balancing for Intradomain– Part III: Load balancing for Interdomain

• Concluding remarks

3

Motivation

Today- Mainly low bandwidth, single-hop, voice-only

Tomorrow- Based on MANET- Broadband multihop- Voice, data, video- Situational awareness, e.g. position sharing- Access critical information: victim info, maps,

construction drawings

Improving the information flow in emergency and rescue operations

4

MANET – What is it?Mobile Ad Hoc Network

Transmission range

• Infrastructureless• Dynamic: mobility, join/leave the network• Rapid deployment• Cost effective

AB

C

5

Challenges• Medium access: Contention-based random access

– Interference and collisions• Distributed routing

– Inconsistency and overhead• Link quality

– Varying, affected by topography, weather etc.• Node mobility

– Link breaks, packet loss, loss of connectivity• Network Capacity

– Low capacity, limited scalability

Need to overcome these challenges to achievehigher network performance!

6

Main Focus of the Thesis

• Mobility– Rerouting time due to mobility and link break

• Load balancing– Intradomain traffic– Interdomain traffic

7

Overview of the work

Part I

Part II

Part III

8

Contributions Part I

9

MobilityPaper A“Rerouting Time and Queueing in Proactive Ad Hoc Networks” V. Pham, E. Larsen, K. Øvsthus, P. Engelstad and Ø. Kure, In proceedings of the Performance, Computing, and Communications Conference 2007 (IPCCC 2007), New Orleans, USA, April 11-13, 2007, pp. 160-169.

10

Motivation• Discovery: Rerouting due to mobility exceeds the

expected 4-6 seconds.

A C

B

11

The contributions

• Analysing the rerouting time

• Proposed solution: Adaptive Retry Limit

12

Last successfull data transmission from A directly to C

Garbage packets are discarded from A’s queueLink is broken. A’s queue being filled up.A transmits data to CNew route to C. Last Hello from C received at A

Analysing a Link Break

A B

C

Queue Accumulation Phase

t0 t1 t2 t3

Queue Reduction Phase

Last Hello from C

I II III

t

td te

13

Solution – Adaptive Retry Limit

123

Packet 1 is transmitted 7 times and discarded

Assumption:• Retry limit = 7• All packets to the same destination

8 79 23789

Packet 2 is transmitted 6 times and discarded

789

Packet 7 is transmitted 1 time and discarded

9 8

In Out

Node A’s Interface Queue

14

Results

Gain

15

Conclusion - Part I

• Rerouting time is affected by:– Packet size and rate– MAC layer queue size– MAC layer retries

• Adapting the MAC layer retries reduces the rerouting time.

16

Contributions Part II

17

Load Balancing for Intradomain Traffic

Paper B“Routing of Internal MANET Traffic over External Networks” V. Pham, E. Larsen, K. Øvsthus, Ø. Kure and P. Engelstad, Mobile Information Systems Journal, iiWAS/MoMM special issue, Volume 5, Number 3, 2009

18

Motivation• Transit routing → Load balancing for intradomain traffic• Alleviate traffic load in the MANET• Improve performance of intradomain traffic:

– Throughput– E2E delay– Packet delivery probability

650 kbps

300 kbps

19

Contributions

• Analysis and simulation of transit routing

• Proposed solution: Cost metric algorithm for transit routing

20

Solution: The Cost Metric AlgorithmCost for ad hoc path

Ci = k

Cost for wired path, no interferenceCii = max(m,n)+c

Cost for wired path, with interferenceCii = sum(m,n)+c

c = constant between 0 and 1

A B

AP_a AP_b

nm

k

l

21

EvaluationA B C

Transit routing enabled

Transit routing not enabled

12

3

45

6

78

910 11

12

22

Results

Average enhancement = 40.3 %

1 2 3 4 5 6 7 8 9 10 11 120

100

200

300

400

500

600

700

Throughput Comparison w/ background traffic

SP Transit

Source Destination Pairs

Thro

ughp

ut (k

pbs)

23

Conclusion - Part II

• Enabling transit routing can be beneficial• Reduce load in MANET• Increase E2E throughput• Reduce E2E delay• Increase probability for packet delivery

24

Contributions Part III

25

Load Balancing for Interdomain Traffic

Paper C-E“Performance Analysis of Gateway Load Balancing in Ad Hoc Networks with Random Topologies” V. Pham, E. Larsen, K. Øvsthus, P. Engelstad and Ø. Kure and, Proceedings of The 7th ACM International Symposium on Mobility Management and Wireless Access (Mobiwac09), Tenerife, Canary Islands October 26-30, 2009

“Gateway Load Balancing in Future Tactical Networks”, V. Pham, E. Larsen, K. Øvsthus, Ø. Kure and P. Engelstad, IEEE Military Communications Conference 2010 (MILCOM 2010), San Jose, CA, USA, October 31 - November 3, 2010

“A Radio Load Based Gateway Load Balancing Scheme with Admission Control” V. Pham, E. Larsen, Q. Le-Trung, P. Engelstad and Ø. Kure, Proceedings of the International Symposium on Wireless and Pervasive Computing (ISWPC 2011), Hong Kong, China, February 23-25, 2011

26

Motivation

GW0 GW1Bottleneck

Global Internet

Address gateway load balancing → Load balancing for interdomain traffic

Focus on outbound traffic! However the result is anticipated to be applicable for inbound traffic as well!

27

Contributions

• Analysing the nature of gateway load balancing (LB)– Explore factors that impact the performance of LB

• Proposed solutions:– RLLB (Radio Load Based Load Balancing)– RLAC (Radio Load Based Load Balancing with Admission Control)

28

Analysis

29

Factors that affect LB• Shortcoming in previous work:

– Simple and small topologies– Grid or constructed topologies– Few topologies used in evaluation

• Needs many random topologies to statistically explore factors affecting the benefit of LB

• ~200 random topologies and thousands of simulations

• Factors that affect performance of LB:– Offered load – Gateway distance– Asymmetry level – Level of spatial reuse/sensing range– Network shape and size

• However, these factors alone cannot explain why the benefit of LB is high for certain topologies while it is poor for others. The layout of the topology is a crucial factor!

30

The congested area represents an obstacle preventing efficient rerouting of traffic!The efficiency of LB depends on where nodes are located relative to the congested centre area

Congestion Map

The centre of the area is most congested, not the area around the GWs!

# transmissions

31

Load Balancing Solution

32

Overview RLAC ArchitectureRadio Load Based Load Balancing Scheme with Admission Control

33

Radio Load Metric

• RL metric provides a measure for the condition in the network

• Pros: The RL info is FREE! RL does not depend on active probing such as in ETX, Packet Pair, RTT

(1 ) busynew previous

window

TL L

T

time (sec)Twindow

Tbusy

34

Gateway Selection

GW0 GW1

?Gateway selection algorithm needs to

consider:1. The load L at the GWs2. The distance h to the GWs3. The bottleneck capacity B to the GWs

L0 L1

h0

B0 B1

h1

Condition at the GWsThe properties of the path towards a specific GW

35

Synchronized ReroutingGW0 GW1

1

2

3

4

5

6

Problem: A group of nodes may simultaneously and repeatedly reroute their traffic back and forth from one GW to another.

Solution: Use a randomized rather than a deterministic gateway selection approach!

36

Randomized Gateway Selection

Draw a random number R ∈ [0,1], and let the outcome decide

Deterministic: If GWn is least congested then select it as default GW

Randomized:

01

1

00 11,10,0

5.0PP

PP

hcLbBaP

1.

2.

37

Randomized Gateway Selection Example

GW0 GW1

1

2

3

4

5

6

0 1P0=1/3 P1=2/3

R4 R5R3

38

Admission Control (AC)• AC prevents traffic load from reaching a critical high

level• AC is enabled/disabled using: L, B and h • When the network load is high:

– Preempt nodes farther away while giving priority to nodes closer to the GWs– Cost is higher to send packets for nodes farther away– Reasonable to enable AC on nodes that have very low packet delivery ratio

• Disadvantage: Unfairness

39

Results

40

Static Topologies (1)

TS1

TS2

41

Static Topologies (2)

42

Mobile Topologies

3 4 5 6 7 8 9 10

-1

0

1

2

3

4

5 1m/s 2m/s3m/s 4m/s5m/s 10m/s1 m/s no_AC

Offered load per node(pkts/s)

Thro

ughp

ut E

nhan

cem

ent (

%)

43

Performance of LB in Mobile Topologies The performance of LB is lower in mobile topologies compared to static topologies because:

• More challenging to perform LB in mobile topologies:– Link breaks, shorter link life time and lower effective capacity– Cannot capture the changes in the topology fast enough due to the

inherent delay in the routing protocol.– The solution in paper A can be used to improve performance

• Condition for evaluation not same. LB is only feasible/ beneficial for a certain amount of the simulation time:– Low asymmetry level or the load is evenly distributed– Partitioning

44

Conclusion - Part IIIPerformance of LB is affected by:

– Asymmetry level, offered load, sensing range,…– Layout of topology

Static asymmetric topologies– LB can improve performance

Mobile topologies– More challenging– Low mobility: moderate enhancement– High mobility: low/no enhancement

45

Concluding Remarks• We are not there yet:

– Challenges need to be solved before we can realize a communication system of tomorrow

• Challenges addressed in this thesis:– Mobility – rerouting time– Load balancing – Intradomain and interdomain traffic

• Increased understanding has been provided through this work

• Proposed solutions that increase the network performance

46

Thank You!