A Comparison of Alternative Client/Server Architectures

Download Report

Transcript A Comparison of Alternative Client/Server Architectures

Gary M. Weiss and Jeffrey Lockhart
Fordham University, New York, NY
UbiMI Workshop @ UBICOMP
Sept. 8 2012
1

Mobile sensors becoming ubiquitous
 Especially via smartphones

Various architectures are possible ranging
from “smart client” to “dumb client”
 Each architecture has pros and cons

Worthwhile to enumerate and compare
alternative architectures
UbiMI Workshop @ UBICOMP
Sept. 8 2012
2
1. Sensor Collection
2. Data Processing and Transformation
3. Decision Analysis/Model Application
4. Data and Knowledge Reporting
Learning/model generation
Only step 1 is required
UbiMI Workshop @ UBICOMP
Sept. 8 2012
3


Main focus of WISDM lab
Monitors smartphone accelerometer and
uses the data to perform activity recognition
 Activities: walk, jog, stairs, sit, stand, lie down

Results available via the Web
UbiMI Workshop @ UBICOMP
Sept. 8 2012
4

Sensor Collection:
 Actitracker client collects raw accelerometer data for
3 axes 20 times per second and transmits to server

Data Processing and Transformation
 Every 10 sec. server aggregates raw samples into a
single example described by several dozen features

Decision Analysis/Model Application
 Server applies predictive model to examples; activity
classified and saved to database

Data and Knowledge Reporting
 User queries server DB any time via web interface
UbiMI Workshop @ UBICOMP
Sept. 8 2012
5
Client Configurations
Responsibility
CC-1
Dumb
CC-2
CC-3
CC-4
Smart
1
Sensor Collection




2
Data Transformation



3
Model Application


4
Reporting

Model Generation
UbiMI Workshop @ UBICOMP
?
Sept. 8 2012
?
6

Mobile devices have CPU power to build models
 Only makes sense to build a model on the client device if
will apply it on the client
 Thus model construction on device only for CC-3 or CC-4
 In CC-1 and CC-2 either model hardcoded into client or
downloaded from server

Data mining not always required
 Can be done dynamically (on client or server) or statically
 Our research shows dynamically generated personal
models outperform general (impersonal) models1
1 Gary M. Weiss and Jeffrey W. Lockhart. The Impact of Personalization on Smartphone-Based Activity Recognition,
Papers from the AAAI-12 Workshop on Activity Context Representation: Techniques and Languages, AAAI Technical
Report WS-12-05, Toronto, Canada, 98-104.
UbiMI Workshop @ UBICOMP
Sept. 8 2012
7

Resource usage
 battery, CPU, memory, transmission bandwidth

Scalability
 Support for many mobile devices

Access to data
 Researchers and others may want raw data
 Transformed data loses information
▪ With raw data can alter features for data mining and
regenerate results
UbiMI Workshop @ UBICOMP
Sept. 8 2012
8

Privacy/Security
 Users will want to keep data secure and/or private

User Interface
 Users want aesthetics (screen size) & accessibility

Crowdsourcing
 Some applications will require a central server in
order to aggregate data from multiple users/devices
▪ Navigation software that tracks traffic
UbiMI Workshop @ UBICOMP
Sept. 8 2012
9

Resource Usage
 Unclear. Resource usage minimized except heaviest
use of transmission bandwidth (power drain)

Scalability
 Poor since maximizes server work
 Actitracker’s server can handle 942 simult. users

Access to Data
 Best since all raw data can be preserved on server
▪ But Actitracker requires 791 MB/month per user.
UbiMI Workshop @ UBICOMP
Sept. 8 2012
10

Privacy/Security:
 Poor: The more data sent the greater the risk

User Interface:
 Good: data and results on server and can be viewed
over Internet

Crowdsourcing
 Best: All data available on server
UbiMI Workshop @ UBICOMP
Sept. 8 2012
11

Similar to CC-1 except:
 Less data to transmit so bandwidth/energy savings
▪ For Actitracker 95% reduction in data
▪ But more processing which takes up CPU and power
 More scalable (less server work)
 Less access to data (raw data not available)
 Slight improvement in privacy/security (no raw data)
 Minimal impact on user interface (results still on server)
 Crowdsourcing only on aggregated data
UbiMI Workshop @ UBICOMP
Sept. 8 2012
12

Resource usage:
 more processing on the client (more CPU and power);
but only need to transmit results



Much more scalable: server only collects results
Access to data: only results available
Much improved security/privacy
 results may not be nearly as sensitive


Can still view results via web-based interface
Can only crowdsource on results
UbiMI Workshop @ UBICOMP
Sept. 8 2012
13

About same as CC-3
 not sending results saves little power




Perfectly scalable: no server
No access to data
Good security/privacy: nothing leaves device
Can only view results on the device
 Not accessible from other places and small screens

Cannot even crowdsource results
UbiMI Workshop @ UBICOMP
Sept. 8 2012
14

Resource usage:
Scalability:
Access to data:
Security/Privacy:
User Interface:
Centralized Data:

One approach: support multiple architectures





unclear
smart client best
dumb client best
smart client best
smart client worst
dumb client best
 approach taken by our research group
UbiMI Workshop @ UBICOMP
Sept. 8 2012
15


Go to wisdmproject.com
Actitracker should be ready for beta in 1 month
 Actitracker.com

Papers available from:
 http://www.cis.fordham.edu/wisdm/publications.php

My contact info:
 [email protected]
UbiMI Workshop @ UBICOMP
Sept. 8 2012
16