Mobile Search and Advertisement Cache Architecture Dimitrios Lymberopoulos, Emmanouil Koukoumidis, Jie Liu, Doug Burger (MSR Redmond) Varun Kansal, Chen Xia, Kenny Chien,

Download Report

Transcript Mobile Search and Advertisement Cache Architecture Dimitrios Lymberopoulos, Emmanouil Koukoumidis, Jie Liu, Doug Burger (MSR Redmond) Varun Kansal, Chen Xia, Kenny Chien,

Mobile Search and Advertisement Cache Architecture
Dimitrios Lymberopoulos, Emmanouil Koukoumidis, Jie Liu, Doug Burger (MSR Redmond)
Varun Kansal, Chen Xia, Kenny Chien, Bin Wu, Fang Wang, Melissa Dunn (MAXPLAT)
Mobile Search Experience
Push Search/Ads to the Phone
Performance Bottleneck
Sensing
Search/Ad Request
m.bing.com
3G
Community
+
Personal
Data
Data
Phone
3G link is slow
o Average end-to-end delay: 4-8 seconds
o 3G connection time: 10 - 30 seconds
Ads
Data
Context
+
Personalization
Leverage mobile trends
o Increasing flash density at lower cost
o Increasing processing power
Goal: Faster Mobile Search Experience
…
indexes
Subjective Cache
results
Data
Web
Cloud
Challenges
o What information do we cache?
o How do we cache it?
o How do we manage it over time?
SONGO (Search ON the GO): A Data-driven Architecture
Aggregate Volume (%)
Percentage of Unique Users
results
Number of query-link pairs
Ranking
user clicks
query
cache hit
Community
Phone
cache miss
m.bing.com
60% of mobile query volume 50% of the users repeat a
hits 6K queries and 4K links! query at least 70% of the time!
Up-to-date
Community Cache
SONGO cache
…
0.links
Rank links based
on their quality
results
…
15.links 31.links
Link Storage
(Flash ~ 1-2MB)
cache hit
Personalization
SONGO cache
Probability of a new query-link pair
Fetch links
Results Page
Construction
periodic updates
Mobile Search Log Analysis - 100M Queries
query
user clicks
Query
Hash
Link
#1
Link
#2
Flags
…
…
…
…
95431A49
…
(08761A49, Q1) (98BA4311, Q2)
…
…
0xFFFFFFFF
…
1E51C65A
…
(18EE1A49, Q3) (08761A49, Q4)
…
…
0xFFFFFFFF
51BD34A3
…
(08761A49, Q5) (98BA4311, Q6)
…
…
0xFFFFFFFF
Adjust Ranking
Scores
…
…
Hash Table
(RAM ~ 200KB)
Experimental Prototypes
On average 66% of the queries a
user submits hit the cache
10
180
9
160
Energy per query (J)
Response time (sec)
8
7
6
5
4
3
2
140
120
100
80
60
40
1
20
0
0
$
UMTS
EDGE
802.11g
16x faster!
$
UMTS
EDGE
802.11g
23x more
energy efficient!
Beyond Web Search
Real-time:
o search results
o business lookups
o ad delivery
Opportunities:
o Faster user experience
o Monetization of autosuggest
o Fastest mobile ad delivery engine
o Personalized ranking for search/ads
o Privacy: profile on the phone