Transcript ppt

Measurement and Modeling of
User Transitioning among
Networks
Sookhyun Yang, Jim Kurose, Simon Heimlicher, and
Arun Venkataramani
University of Massachusetts Amherst
[email protected]
This research is supported by US NSF awards CNS-1040781 and CNS-1345300.
UNIVERSITY OF MASSACHUSETTS, AMHERST • School of Computer Science
Outline





Introduction
Measurement Methodology
Measurement Analysis and Findings
Empirical Investigation of Model
Conclusion
2
Mobility is the key driver of networking

Historic shift from PC’s to mobile/embedded devices
~1B server/PC’s
~1B smartphones
INTERNET
(2011)
~2B server/PC’s
~1B Internet
-connected PC’s
~5B cell phones
~10B mobiles
INTERNET
(2020)
[1] Cisco Visual Networking Index: Global Mobile Data Traffic Forecast Update, 2014–2019
[2] Pew Research Center, The Internet of Things Will Thrive by 2025, 2014
3
Mobility in, and among, among networks

Physical mobility among access points
to
Internet
VLR
Mobile
Switching
Center
mobile
user
visited
network
Cellular network mobility (e.g., [3])
Wi-Fi network mobility (e.g., [4])
Device mobility within the same type of a network
[3] U. Paul et al, Understanding traffic dynamics in cellular data networks, INFOCOM 2011
[4] M. Kim et al, Extracting a mobility model from real user traces, INFOCOM 2006
4
Mobility in, and among, among networks

Virtual mobility among access networks


Move among edge and provider networks
Persistently keep his/her ID (name) across networks
Cable network

Enterprise network
via VPN
Cellular network
For instance, a stationary user with multi-homing,
multiple devices
5
Our contribution

Quantitative understanding of virtual mobility





Sequence of associated networks
Network residence time
Degree of multi-homing
Network transition rate
Gives insights and implications on locationindependent architectures

e.g., Mobile IP, MobilityFirst [5], XIA [6]
[5] A. Venkataramani, J. Kurose, D. Raychaudhuri, K. Nagaraja, M. Mao, and S. Banerjee.
Mobilityfirst: A mobility-centric and trustworthy internet architecture. ACM CCR, 2014
[6] D. Han et al. XIA: Efficient support for evolvable internetworking. USENIX NSDI, 2012
6
Outline





Introduction
Measurement Methodology
Measurement Analysis and Findings
Empirical Investigation of Model
Conclusion
7
How to get traces of virtual mobility?
Question: What is the most feasible way to capture
such user’s virtual mobility?
Large population of users!
Difficult to install SW on all their devices!
Far too many servers
and application servers
to be monitored!
8
Can we log virtual mobility via mail server?

User frequently accesses his/her mailboxes




mail periodically pushed (e.g., every 5mins) to user
Same user ID is used across multiple networks
and sessions.
Mail server logs allow us to identify the
network address where a user is resident.
IMAP mail access server logs


Contain sign-in logs with user ID, IP address,
and timestamp
Informal lower-bound of the actual amount of
network-transitioning performed.
9

CS-only users




IMAP servers for UMass
School of CS
81 users, one year
405 IP prefixes, 387 ASes
UMass-wide users



Servers for all UMass
students (primarily),
faculty, and staff
7,137 users, 4 months
9,016 IP prefixes, 1,777
ASes
Fraction of Sign-in logs
IMAP mail access logs
ASes in decreasing order of the
fraction of Sign-in logs
(e.g., Comcast cable, Verizon, Five colleges
network incl. UMass, AT&T Wireless, Sprint
Wireless)
10
How to reconstruct a user’s session?

Given a series of IMAP sign-in logs,


Time window
At least one log for a time window indicates that a
user is connected for the entire time window
Alice made Verizon connections
Alice made Comcast connections
t1
∆t
t2
∆t
t3
∆t
t4
∆t
t5
∆t
t6 time
Alice has been connected to Comcast from t1 to t3.
Alice has been connected to Verizon from t2 to t3
contemporaneously .
11
Appropriate size of a time window?

Time window dilemma in session identification [7]


# of sessions as a function of time window sizes
Number of sessions
(X106)

Small window overestimates
Large window underestimates
Knee (elbow) at 15mins!
[7] J. Padhye and J. F. Kurose. Continuous-media courseware server: A study of client
interactions. IEEE Internet Computing, 1999
12
Outline





Introduction
Measurement Methodology
Measurement Analysis and Findings
Empirical Investigation of Model
Conclusion
13
Mobility among networks
How frequently does a user switch a network in 15mins?
40%
70%
UMass-wide users
CS users
Daily number of a user’s mobility
among ASes
Approx. 70% of CS users (or 40% of UMass-wide users)
moves among networks at least once a day.
14
Network residence time (over all users)
Comcast cable
Verizon online
Charter communications
Hughes network
HOUSE
WORK
Sprint Wireless
Five colleges
(incl. UMass)
AT&T Wireless
Verizon Wireless
WORK
MOBILE
MOBILE
HOUSE

80-to-90% from three categories only with “8” ASes
out of 400
15
An individual user’s network residence time?
75% of users spent more than 90% of their time
in their top three networks.
Fraction of a user’s
top three networks (ASNs) residence time (%)

Overall, users spent more than 60% of their time in their top
three networks.
16
Contemporaneous connections
(picture of my advisor’s
house)
In the traces, a series of sign-in logs produced
from “multiple” networks in 15mins implies
“contemporaneous connectivity”
17
User’s contemporaneous connections
UMass-wide users
UMass-wide
users
Contemporaneous users
CS users
80% of CS users
50% of UMass-wide users
Fraction of a user’s contemporaneous time to
connection time (%)

Most contemporaneous users spent up to 20% of their
connection time in multiple networks.
18
Outline





Introduction
Measurement Methodology
Measurement Analysis and Findings
Empirical Investigation of Model
Conclusion
19
User virtual mobility model


Characterizes the transition rate at which a user
moves among networks
Predicts signaling overhead to the name and location
translation service


e.g., a home agent, GNS in MobilityFirst
User model via a discrete-time Markov-chain


: # of networks newly attached at time t, w.r.t. time t-1
User’s network transition
: # of networks connected at t
Signaling
overhead
at time t
Attachment
signaling
Detachment signaling
19
(Xt, Yt)-series data properties

Investigate stationary, memoryless properties

Time series plot on a daily value of Yt (all users)
Model estimation
(phase 1)
Model validation
(phase 2)

KPSS test: data stationarity

Autocorrelation function (ACF): daily/weekly periodicity
21
CS users signaling overhead
How well does the model predict signaling overhead?
model (phase 1)
observed (phase 2)
Q-Q plot
Signaling overhead
over all users
Visually a good fit
Statistically a good fit
22
UMass-wide signaling overhead
Heavy user cluster of
721users
EM
clustering
Signaling overhead
No fit!
But a mixture of Gaussian
distributions.
Signaling overhead
Visually a better fit
These results suggest proper clustering can improve the
model’s signaling overhead predictability.
23
Conclusions

We performed a measurement study of user
virtual mobility and discussed insights and
implications from the measurements.




Users spend most of their time in a few networks.
Large number of users are contemporaneously
connected to more than one networks.
We show the predictability of overall signaling overhead
using an individual user model.
More generally, we believe that this paper is an
important step in deepening the understanding
of managing virtual mobility at global scale.
24
End
Questions or comments
welcome!
UNIVERSITY OF MASSACHUSETTS, AMHERST • School of Computer Science