Weighted Way Mobility Model and its Impact on Ad Hoc
Networks
Kashyap Merchant*, Wei-jen Hsu**, Haw-wei Shu**, Chih-hsin Hsu**, and Ahmed Helmy**
*Computer Science Department
**Electrical Engineering Department
University of Southern California, Los Angeles CA 90089-2562
{kkmercha, weijenhs, hshu, chihhsih, helmy}@usc.edu
The effects of different mobility models on protocol
performance in ad hoc networks have been studied extensively in
some recent works [1][2]. The most widely used mobility model
is perhaps the Random Way Point (RWP) model, due to its
simplicity and capability to synthesis scenarios with varying
degrees of mobility. It has been reported that although RWP
model is widely used, it fails to capture the normal human
mobility characteristics [2].
There are studies on mobility models that better fit human
behavior [1][2]. However, these works do not address an
important issue: The destination is not pure randomly chosen in
real life. We investigate this issue in this work and propose a new
model called the Weighted Way Point (WWP) model. The major
differences of WWP model and the popular RWP model are:
(a) A mobile node no longer randomly chooses its destination: In
daily life, it is very unlikely that a person chooses a random
location as his/her destination. We model such behavior by
assigning different “weights” to different locations. Such a nonuniform probability of choosing destination results in nonuniform node distribution. (b) The “weights” of choosing next
destination location type depends on both current location type
and time: To set up WWP model for a campus, we categorize
locations on campus into 4 types: classrooms, libraries, dining
areas, and other areas. We collect the data during two time slots:
9am-1pm and 1pm-5pm. (c) The pause time distribution at each
type of location is different and is treated as a property of that
location. (d) Our study focuses on a campus environment, but we
also model the mobile nodes leaving the campus with certain
probabilities and coming back later to include a more general
behavior. (e) The probability and duration of using wireless
networks are also treated as properties of location. Our work is
divided into two parts 1) modeling the parameters of the WWP
model based on the USC campus survey results and 2) Comparing
the characteristics of the RWP and WWP models using a
simulation study.
We collected 268 survey responses randomly on USC
campus. From the survey we captured statistics about the
following parameters: (a) The pause time at each type of locations,
(b) The probability of going to each type of location, given the
current location and time section of the day (a time-varying
transition probability matrix defined over two time intervals of
morning- 9 am to 1 pm and afternoon 1 pm to 5 pm). (c) Wireless
usage behavior, about the probability and duration a respondent
uses wireless networks at different types of locations. Unlike
previous studies on base station usage measurements, ours is a
novel approach to capture real mobility patterns and wireless
usage data. Although base station measurements also provide
some information about wireless usage, we believe that by giving
out surveys we have an unbiased sampling across the general
population, instead of looking at wireless users only. Also, the
collected traces from base stations are ‘usage’ patterns and not
‘mobility’ patterns. Hence, we feel our data complements base
station measurements.
The survey results showed some very interesting results such
as a) The duration of using wireless network (flow duration)
follows a heavy tailed distribution for library locations b) The
distribution of pause time at classroom is like a bell-shapped
normal distribution with the peak around the 60-120 minutes
interval (USC average class duration is 90 minutes) c) People
tend to go to from a library to a dinning area in the morning
interval (lunch time) and tend to go from library to classroom in
the afternoon interval (afternoon classes). d) Also most transition
are of the type ‘offcampus-class-offcampus’ or ‘offcampuslibrary-offcampus’ which we believe reflects the general student
movement on-campus. We use the simulation environment as
shown in Fig. 1. which we refer to as a “virtual campus”.
Fig. 1 Virtual Campus
This virtual campus topology is adopted from a small
part of the actual USC campus. In this scenario we define
noncontiguous locations: 3 classrooms, 2 libraries, and 2
cafeterias.
Fig 2. MN density vs time
We use simulations to show the characteristics of WWP
model, in comparison to RWP model. First, WWP model shows
“uneven spatial distribution” of MNs. The MNs tend to cluster
within the locations as seen in Fig 2. for classroom1, library1 and
cafe1. However the node density is quite low for other area and
off campus locations. Second, although for a given transition
probability matrix there should be some theoretical steady state
MN distribution, the matrix is time-dependent and changes from
time to time throughout the day, hence MN distribution in
simulation area never reaches a steady state. This suggests
stationarity is not necessarily a realistic requirement for mobile
models. Third, we use move-stop ratio as metric to describe
“mobileness” of a mobility model and found that the WWP model
based on real survey data has a lower move-stop ratio than the
RWP model with common parameter settings.
We further show the impact of WWP model on wireless
communication. We consider both last-hop wireless networks
(802.11 WLANs) and ad hoc networks. Assuming MN only uses
wireless networks when it stops within a location, we show that as
the number of MNs increases in the system, the WWP model has
about twice the number of flows as compared to the RWP model.
Also the congestion ratio (ratio of the number of congested flows
divided by the number of total flows) of the WWP model is
double that of the RWP model.
Next we investigate the effect of varying the distance
between the locations for the 2 MNs current location relationships:
same and different. We consider two configurations (Near and Far)
of 3 locations each. Near configuration has a distance of 50 m
while Far has a distance of 150 m between the edges of the
locations as shown in Fig 4.
(a) Near location scenario
(b)Far location scnario
Fig. 4. Near and Far location arrangements
From Fig 5 we can see that increasing the distance between
the locations for the different location (diff_location) relationship
severely reduces the average success ratio (from 55.14% to 4.48%)
of establishing a path from a MN to the other. This is a very
significant reduction and shows that it is hard to find a sufficient
amount of usable nodes to form an ad hoc path between MNs in
different locations.
Fig 3. Number of flows generated vs. ratio of congested flows
Another interesting result shown in Fig 3. reveals how even
when both models have the same number of flows, the WWP
model always has a higher congestion ratio than the RWP model.
This is because in the WWP model locations are chosen as MN
destinations with non-uniform weights. If a location is more
popular than others, it attracts more MNs hence a greater
proportion of the flows are initiated at the location. Thus some
locations have more flows and these flows are likely to be
congested. Whereas in the RWP model the flows are more evenly
distributed among the locations hence the congestion ratio is not
as high given the same number of total flows.
For ad hoc networks, we test the successfulness of route
discovery under 2 different MN location relationships. If WWP
model is used, we show that a) for MNs in the same location it is
relatively easy to discover a path between them. b) If the MNs are
present in different locations the success reduces to about a third
of the value of the first relationship. The reason for this is that due
to the preference of MNs to choose locations as destinations, the
number of nodes present in outside areas is very small. Hence for
MNs present in different locations it is difficult to form an ad hoc
path between each other.
Fig 5. Comparison of success rate of Near and Far location
arrangements
We come to the conclusion that ways of choosing
destinations in mobility model has non-negligible impact on
wireless network performance. In our on-going work we aim to
correlate survey data and base station measurements at USC to
improve upon our model. In the future we plan to study traces
from various campuses to establish a procedure to set up
parameters of WWP model for a more general campus mobility
model.
REFERENCES
[1] F. Bai, N. Sadagopan, and A. Helmy, "The IMPORTANT
Framework for Analyzing the Impact of Mobility on
Performance of Routing for Ad Hoc Networks", AdHoc
Networks Journal, Vol. 1, Issue 4, pp. 383-403, Nov 2003.
[2] A. Jardosh, E. M. Belding-Royer, K. C. Almeroth, and S. Suri,
“Towards Realistic Mobility Models for Mobile Ad hoc
Networks”, ACM MobiCom, pp.217-229, September 2003.