Aruna Balasubramanian

1
Enhancing Interactive Web Applications in Hybrid
Networks

• Aruna Balasubramanian
• Brian Neil Levine
• Arun Venkataramani
• University of Massachusetts Amherst

Supported in part by NSF awards CNS-0133055,
CNS-0519881 and CNS-0721779 and in part by ARO
award W91NF-07-1-0281
2
Motivation

• Higher demand for access to network applications
  from mobile users

Technologies
3G
Application support
Opportunistic WiFi
Cost
3
Research question

• How can we enable a broader class of interactive
  applications using opportunistic WiFi access?

1. Mobile-to-Infrastructure contacts
2. Mobile-to-Mobile contacts
Internet
4
Application design space
Delay Tolerant Email, bulk transfer
1s
Interactive Web? Web browsing, Web search
Connection duration with AP
60s
VoIP
3600s
1s
60s
3600s
Disconnection duration with access point (AP)
5
Outline

• How can opportunistic mobile-to-Infrastructure
  (m2i) contacts be used to support web search?
• Can performance be improved by simultaneously
  leveraging mobile-to-mobile (m2m) contacts?
• Evaluate the performance of web search using a
  mobile testbed.

6
Web search process
ltyour favorite search enginegt
Retrieving.
Challenges (1) Intermittent
connectivity (2) Short connection
opportunity
Retrieving.
7
Thedu Adapting Web search for intermittently
connected networks

• Use aggressive prefetching web pages to convert
  interactive process to transactional
• Prioritize prefetched web pages under bandwidth
  constraints

8
1. Aggressive prefetching
Thedu Proxy
9
1. Aggressive prefetching
Google, Yahoo, Live , Ask, .
Queries from mobile
Store query
Interface
Thedu Client
Snippets
Prefetch

• Prefetching has been used to improve availability
  Coda91, Chandra01
• Our challenge Allocating bandwidth resource to
  prefetched web pages

Web pages returned to mobile
Thedu proxy
10
2. Prioritizing useful web pages
ltyour favorite search enginegt

• How many web pages to prefetch?
• What is the usefulness of a web page for a query
  in comparison to a web page of another query?

11
Prioritize web pages from different queries

• Search engine often rank web page for a single
  query using relevance scores
• Scores not comparable across queries

q2
q1
Thedus query normalization techniqueusing
Kullback Leibler divergence
Are all web pages returned for a query
useful?
12
Homepage versus Non-homepage queries
Mobicom 2008 lturlgt www.sigmobile.org/mobicom/2008
lt/urlgt ltsnippetgt MobiCom 2008, the 14th Annual
International Conference on Mobile,
lt/snippetgt lttitlegt ACM Mobicom 2008 lt/titlegt
Oil prices summer lturlgt www.eia.doe.gov/steo
lt/urlgt ltsnippetgt Residential natural gas prices
over the same period are projected to
,lt/snippetgt lttitlegt EIA - Short-Term Energy
Outlook lt/titlegt
13
Thedus query-type classifier

• Thedu Aggressive prefetching
• Prefetch top 20 URLs for each query
• Identify home page queries and return only 1
  relevant web page in expectation
• Prioritize the remaining web pages across queries
  by normalizing relevance scores

Thedus query-type classifier accuracy 88
14
Outline

• How can intermittent mobile-to-Internet (m2i)
  connectivity be used to support web search?
• Can performance be improved by simultaneously
  leveraging mobile-to-mobile (m2m) contacts?
• Evaluate the performance of web search using a
  mobile testbed.

15
Leveraging opportunistic m2m contacts
When useful? When meeting opportunities are
skewed
Why useful? Because of decreasing marginal
utility of web pages
Internet
16
Utility based routing IR meets networking

• Goal Maximize relevant responses delivered
  within deadline

Utility of routing X or Ys web page P(web page
is relevant) x P(W can deliver to destination
within deadline)
17
m2m routing between mobile nodes

• Deliver web pages destined to peer
• Route other web pages using similar utility-based
  routing
• Exploit query locality by caching popular web
  pages

18
Outline

• Can intermittent mobile-to-Internet (m2i)
  connectivity be used to support web search?
• Can performance be improved by simultaneously
  leveraging mobile-to-mobile (m2m) contacts?
• Evaluate the performance of web search using a
  mobile testbed.

19
Evaluation goals

• Does Thedu improve performance of web search
  using opportunistic WiFi access?
• Is there a benefit for leveraging m2m contacts?

Evaluation based on deployment of Thedu on
DieselNet testbed and trace-driven simulations
20
Evaluation tools Networking and IR
DieselNet Testbed at UMass
21
Deployment results Per day average
Relevant Web pages delivered
22
Delay in receiving relevant web page
23
Effect of m2i meeting frequency Per day average
Trace-driven Nov, 2007
Deployment March, 2007
24
Is leveraging m2m contacts useful?

• Trace-driven simulation on DieselNet shows
  leveraging m2m contacts provide little benefit
• Ratio of m2i versus m2m contacts is 201
• We developed an analytical model to quantify the
  benefits of m2m routing
• Model suggests that m2m routing is useful only if
  AP density is low

25
Leveraging m2m contacts Sparse AP

• Leveraging m2m contacts
• provides throughput benefit, but not delay
• does not improve performance of interactive
  applications, even in sparse AP scenarios such as
  developing regions

26
Conclusions

• Aggressive prefetching and prioritization
  significantly improves interactive web search
  performance
• Leveraging m2m contacts only provides throughput
  benefits and only when AP density is low
• provides little benefit for interactive
  applications
• Deployment of Thedu on DieselNet shows 4 times
  improvement in number of relevant web pages

trace.cs.umass.edu
27
DieselNet Both m2i and m2m contacts
m2m meeting
m2i meeting