Project Objective and motivation

1
Characterizing User Behavior and Network
Performance in a Public Wireless LAN

• Anand Balachandran
• Geoffrey M. Voelker
• P. Venkat Rangan
• - UC San Diego
• Paramvir Bahl (Microsoft Research)
• ACM SIGMETRICS, June 2002

Presentation by Karthik Nandakumar 25th March
2002
2
Overview

• Goals of this study
• Related work
• Methodology
• User Behavior
• Network Performance
• Conclusions

3
High-Level Goals

• To understand wireless user behavior in terms of
  user distribution across Access Points, user
  session times, application mix, etc.
• To characterize wireless network performance in
  terms of total throughput, peak offered load,
  packet error rates, etc.
• To characterize wireless users in terms of a
  parameterized model for use with analytic and
  simulation studies involving wireless LAN traffic
• To apply the workload analysis results to issues
  in wireless network deployment such as capacity
  planning and potential network optimizations like
  load balancing

4
Related Work

• Tang and Baker, Stanford focused mainly on user
  mobility
• Eckardt and Steenkiste, CMU focus was on the
  error model
• and signal characteristics of the RF environment
  in the
• presence of obstacles
• Noble Satyanarayanan, CMU and Nguyen Katz ,
  UC
• Berkeley Trace Modulation recreates the
  observed
• end-to-end characteristics of a real wireless
  network in a
• controlled and repeated manner

5
Trace-Based Mobile Network Emulation

• Trace collection
• Reduce observations to a list of parameters of a
  simple,
• time-varying network model
• Network performance reproduced in a controlled
  manner
• Applications run unmodified in the emulated
  network
• end-to-end performance mirrors the original
  network
• Creates synthetic networking environment rather
  than a
• synthetic workload
• Develops a good model for network behavior, but
  does not
• characterize user activity

6
Methodology Network Environment
7
Data Trace Collection

• Trace collected over 3 days at the ACM SIGCOMM01
• conference held at UC San Diego August 2001
• Trace consists of two parts
• SNMP Data at the APs aggregate packet level
  statistics of
• all traffic at the APs
• Tcpdump Trace network level headers of packets
  passing
• through the switch

8
User behavior Characterization

• User Distribution Across Access Points
• User Session Duration
• User Data Rates
• User Application Popularity
• User Mobility

9
User Distribution Across Access Points

• User arrivals closely follows the conference
  schedule
• User arrivals are correlated in time (at each AP)

10
User Distribution Across Access Points

• Users are evenly distributed across all APs
• User arrivals are correlated in space (between
  APs)

11
User Arrivals Model

• User arrivals can be modeled as a
  Markov-Modulated
• Poisson Process (MMPP)
• Underlying Markov chain has ON and OFF states
• Mean duration of OFF state 6 minutes
• More or less constant arrival rate during the ON
  state
• Mean inter-arrival time during the ON state 38
  seconds
• MMPP model is well-suited when arrivals are
  correlated in
• time and space. But this model cannot be
  generalized

12
User Session Duration
CDF of User Session Time

• 60 of the sessions last less than 10 minutes
• 90 of the sessions last less than one hour

13
User Session Duration Model

• PDF of session time follows a General Pareto
  Distribution
• ( Shape parameter 0.78, Scale parameter
  30.76)
• 88 of the sessions are active for longer than
  70 of the session time. Only 4 are inactive for
  more than half of their session length
• Relatively idle sessions are those with longer
  session lengths
• Implication DHCP servers must be configured to
  have short lease times on IP addresses (say 10
  minutes) overcomes problem of limited DHCP
  addresses (by quickly recycling addresses)

14
User Data Rates

• Minimum, Average and Peak bandwidths of each user
  session
• are widely spread. Per session average bandwidth
  ranges from
• 15 kbps to 590 kbps
• Session classification
• Lower 25th percentile of the sessions light
  sessions
• 25th to 90th percentile medium sessions
• Top 10 of the sessions heavy sessions

15
User Application Mix

• Web browsing (HTTP) and SSH are the most popular
• applications ( 64 of bytes and 58 of flows)

16
User Mobility

• Users were mobile when expected i.e. at the
  beginning and
• end of conference sessions
• Majority of the users (gt80) were seen at more
  than one AP
• Most users switch to a different AP each time
  they exit and
• re-enter the conference hall
• User Mobility is not an important aspect in this
  setting

17
Summary of User Behavior

• Users are evenly distributed across all APs and
  user arrivals
• are correlated in time and space
• Most users have short session times 60 of the
  user sessions
• last less than 10 minutes
• User sessions can be broadly classified into
  light, medium and
• heavy based on average data rates
• Web SSH traffic accounts for 64 of the total
  bytes transferred
• Users are mobile between APs but only between
  sessions

18
Network Performance Offered Load

• Peak Throughput seen at the AP is 3.2 Mbps

19
Network Performance Offered Load (2)

• Observations
• Load distribution is highly uneven across APs
  though the
• number of users in each AP is roughly the same
• APs do not reach peak offered load when the
  number of
• associated users is maximum
• Inferences
• Offered load is more sensitive to individual user
  data rates
• rather than just the number of users
• Load balancing algorithms must also consider
  individual user
• bandwidth requirements in addition to the number
  of users

20
Network Performance Packet Errors

• Error rates are low, but not insignificant, and
  similar for all APs
• Error rates are bursty over time, can be quite
  high for significant
• periods of time
• Channel oscillates between good and bad states

21
Summary of Network Performance

• Offered load on the network directly correlates
  with the
• conference schedule
• Bandwidth distribution across APs is highly
  uneven and does
• not correlate to the number of users at an AP
• Even with just 4 APs for 195 users, the network
  is
• over-provisioned. None of the APs reach their
  maximum
• capacity even at peak loads
• Wireless channel characteristics are similar
  across all APs
• variation is more time-dependent than
  location-dependent

22
Conclusions

• User arrivals better modeled as tightly
  correlated in time and space (MMPP)
• DHCP can provide short-term leases on IP
  addresses (short session durations)
• As low as four APs suffice to handle traffic of
  195 users capacity planning will be even easier
  with higher capacity networks (802.11a)
• AP throughput
• Not well correlated with number of users
• Load balancing algorithms should incorporate
  individual user data rates

23
Conclusions (2)

• Generalization
• Applicable in other similar public wireless
  LANs like classrooms, airport gates, etc.
• Not Applicable in places like university
  campuses, malls, enterprise, etc.

24

• Thank You!