Multi-Variate Analysis of Mobility Models for Network Protocol Performance Evaluation Carey Williamson Nayden Markatchev {carey,nayden}@cpsc.ucalgary.ca.

Multi-Variate Analysis of Mobility Models for Network Protocol

Performance Evaluation

Carey Williamson

Nayden Markatchev

{carey,nayden}@cpsc.ucalgary.ca

University of Calgary

Preamble and Motivation

• Consider mobile host movement in an arbitrary internetwork

• Can disconnect from one network at any time, move to another location, and reconnect, while maintaining same identity

• See IETF Mobile IP

C

B

A

Example: Three different “home networks”, each withtheir own (stationary) router or base station (A, B, C).

Small circles and triangles represent mobile hosts.Triangles belong to multicast group G, while circles do not.

C

B

A

Observation 1: Mobile hosts can move anywhere anytime.

C

B

A

C

B

A

C

B

A

C

B

A

Mobile Host (MH) registers with Foreign Agent (FA) at thevisited network, and with its Home Agent (HA) as wellto enable packet forwarding (via tunneling).

C

B

A

Packet fromCH to MH

C

B

A

Packet fromCH to MH

Packet fromHA to FA

Basics of IETF Mobile IP packet forwarding

C

B

A

Observation 2: Similar rules apply for mobile hoststhat are members of multicast groups.

C

B

A

C

B

A

C

B

A

Packet fromMS to G

C

B

A

Packet fromMS to MH

Packet fromHA to FA

Can be done using unicast “bidirectional tunneling”.

C

B

A

Observation 3: This can be inefficient if multiple groupmembers are away at the same location.

C

B

A

C

B

A

C

B

A

Packet fromMS to G

C

B

A

Packet fromMS to MH

Packet fromHA to FA

Packet fromMS to MH2

Packet fromHA to FA

C

B

A

Packet fromMS to G

Packet fromHA to FA

More efficient solution is to tunnel the multicast itself.

C

B

A

Observation 4: Inefficiency still exists if multiple HA’shave group members away at the same location.

C

B

A

C

B

A

C

B

A

C

B

A

Packet fromMS to G

C

B

A

Packet fromMS to G

Packet fromHA to FA

Packet from

MS to G

Packet from

HA to FA

This is called the “tunnel convergence problem”.

B

C

B

A

Packet fromMS to G

Packet fromHA to FA

The solution in the MoM (Mobile Multicast) protocol is toselect a Designated Multicast Service Provider (DMSP)to forward multicast packets to G at a certain network.

Multicast group

DMSP (HA)

Mobile Host

Observation 5: The generalcase can be very messy!The performance of MoM(or any other protocol)depends on group size andon MOBILITY PATTERNS.

Multi-Variate Analysis of Mobility Models for Network Protocol

Performance Evaluation

Carey Williamson

Nayden Markatchev

{carey,nayden}@cpsc.ucalgary.ca

University of Calgary

Motivation

• The performance of a mobility support protocol is highly sensitive to user mobility patterns.

• Very little is known about mobile user behaviors in operational networks.

• Most simulation studies evaluating protocol performance use simple models of user mobility. (e.g., “random walk”)

Overview of this Research

• Proposes a more general suite of mobility models

• Models are classified along two orthogonal axes: degree of correlation (I/C) and degree of skewness (U/N):– Independent Uniform (IU)

– Independent Non-Uniform (IN)

– Correlated Uniform (CU)

– Correlated Non-Uniform (CN)

• Uses the MoM protocol as a case study for the models.

• Impacts of mobility model parameters assessed using the Analysis of Variance (ANOVA) statistical technique.

Background and Related Work

• Mobile Computing and Mobile IP. – IETF Mobile IP protocol

• Mobile Host (MH)

• Foreign Agent (FA)

• Home Agent (HA)

• The model works but multicast support is inefficient. (tunnel convergence problem)

Therefore…

Background and Related Work(2)

• New protocols, such as the MoM (Mobile Multicast) protocol, are proposed to deal with this issue.

• MoM uses the Home Agent for delivery of multicast datagrams to mobile users, and achieves scalability through a Designated Multicast Service Provider (DMSP) for each multicast group on a foreign network.

Basic Mobility Model in MoM

New Mobility Models

• To broaden the range of mobility patterns considered, we introduce two new model parameters

• Correlation

– The tendency for certain hosts to move in patterns that are related either geographically (i.e., location) or temporally (i.e., time).

• Skewness

– Some destinations are more popular than others.

• The combination of those two factors leads to four different mobility models: CU, CN, IU, IN.

Mobility Model Parameters

• Homing Probability - HOMING_PROB (0.5)• Mean Residency Time (60 time units) and Mean

Travel Time (6 time units).• Skewness

– Degree of skewness – k >= 0.

• Correlation (i.e., follow the leader)– FRACTION_FOLLOWERS (% of mobile hosts)

– FOLLOW_PROBABILITY (per-move by a follower)

Model Validation

Experimental Parameters

Experimental Design

• Simulations are used to assess the performance impacts of multicast group size, network size, number of mobile hosts, and host mobility model.

• Simulations run for 26,000 time units, of which the first 6,000 time units are for warm up.

• Only one multicast group is simulated.

Performance Metrics

• DMSP forwarding overhead per HA.

• Number of DMSP handoffs.

• The average number of foreign networks visited by mobile multicast group members (per HA).

MoM Performance

MoM Performance (zoom)

Line A - Average number of group members away.Line B - Average number of different foreign networks at

which the away group members reside.Line C- DMSP forwarding overhead.

Impact of Mobility Model on Number of Foreign LANs Visited

Analysis of Variance (ANOVA)

• ANOVA is a statistical technique to analyze multi-variate data and figure out which factor is “most important”.

• The method separates the total variation of the performance index into components associated with possible source of variation.

• Tabular analysis: row effect vs. column effect.

• F-test values determine the level of factors influence.

• Multiple independent replications of experiments are used to identify the interaction effects between different factors.

DMSP Overhead per HA(3 replications)

10 LANS, 10 Hosts per LAN

Multicast group size = 100

Note: lower is better.CN is best case.IU is worst case.

ANOVA Results:DMSP overhead per HA

• Correlation factor - 67.0%

• Skewness factor - 28.5%

• Interaction - 2.25%

• Error - 2.22%

DMSP Handoffs(3 replications)

ANOVA Results: DMSP Handoffs

Correlation factor - SSA/SST = 349,515/399,980 = 87.4%Skewness factor - SSB/SST = 6.2%Interaction - SS(A+B)/SST = 0.4%Error - SSE/SST = 6.0%

The P value indicates the statistical significance of each value.

Average Foreign LANs Visited (per HA) (3 replications)

ANOVA Results:Foreign LANs Visited (per HA)

Number of Hosts per LAN - 58.0%

Number of LANs - 31.3%

Interaction - 10.7%

Error - 0.003%

Effect of Correlation Parameterson LANs Visited (3 replications)

ANOVA Results:Impact of Correlation Parameters• FRACTION_FOLLOWERS accounts for

34.2% of the total variation.

• FOLLOW_PROBABILITY accounts for 35.9% of the total variation.

• Interaction effects account for 29.2%.

• Errors contribute 0.7%.

Effect of Skewness Parameterson LANs Visited (3 replications)

ANOVA Results:Effect of Skewness

• Correlation factor contributes 57.6% of the total variation.

• Skewness contributes 33.9% of total variation.

• The interaction effect accounts for 8.0%.

• The effect of errors is 0.6%.

Summary and Conclusions

• The proposed suite of models (IU, IN, CU, CN) represents a broad set of possible behaviors for mobile users.

• The choice of mobility model can have a significant effect on protocol performance.

• The degree of correlation between mobile hosts has a greater impact than the degree of skewness.

• For the MoM protocol, the Independent Uniform (IU) model is actually the worst case stress test.

Future Work

• Extending the correlation models to include dynamic multicast group membership.

• Applying our mobility models to routing in ad hoc wireless networks

• Applying our mobility models to the evaluation of the rekeying protocols for secure multicast groups.