Recent trends indicate that multimedia traffic is on the rise.
Streaming media is becoming more commonplace, as news and sports sites
add video content to their sites and as Internet radio grows in
popularity. Telepresence, distance learning, telecommuting, and remote
collaboration applications also continue to gain popularity. In spite
of this, we still do not completely understand the underlying
characteristics of this traffic, nor its effects on other network
traffic. My contributions in this area are (1) a measurement and
assessment architecture and methodology to determine the
"quality" of a received media stream, and (2) a source
characterization of H.323 traffic, which was a precursor to later work
on telepresence applications.
While much current and prior work has been done in the general area
of wide-area measurement and assessment architectures, little has been
done specifically for measurement of streaming media. Existing
solutions fail to adequately assess a client's perception of the
received media stream, by either removing the client's interactions
with the server or by concentrating solely on network-level statistics
(packet loss, bandwidth, etc.) without mapping these measures to
user-perceived quality. Methods of subjective media quality
measurement, such as the use of mean opinion scores, are not scalable
enough to be used in large networks. A viable solution to this problem must be both scalable, to be used among large numbers of media player clients, and accurate, in that whatever (objective) measurement it uses must correlate well with any "quality score" the user would assign for the same stream under the same conditions.
Carleton College
At Carleton, I am continuing the work I started while at HP Labs (see below). My current work focuses on the following problems and issues:
- Prediction of user-assigned quality scores from objective measurements. Based on a series of user tests in May 2006 and October 2005, we have collected both objective and subjective measurements of stream quality. Results so far demonstrate that there is a strong correlation between certain objective metrics and user-assigned quality scores; our goal is to exploit this information to predict user-assigned quality scores without having to poll users directly.
- Architecture feasibility. Could a system such as the one we're proposing be viable in a large-scale environment such as the Internet? We are currently studying the data storage, processing, and runtime requirements of such a measurement system.
Undergraduate research students who are or have worked with me on this project include: Sami Benzaid (current), Erica Bolan, Ben Kazez, Brandy McMenamy (current), Jamie Olson, Keith Purrington '05, Anna Sallstrom, Ben Sowell, Amrit Tuladhar, Wain Yee.
HP Labs
At HP Labs, in collaboration with Ed Perry, I designed a client-side
media quality assessment architecture that maps objective metrics, such
as packet loss or client buffering patterns, to subjective ideas of
received media quality, as seen by the user. The assessment
architecture accomplishes this automatically, without requiring user
input. The architecture is a scalable system consisting of a client-side
assessment tool, which I developed, that reports playback metrics
either in real-time or off-line to assessment points for analysis and
dissemination to interested parties. Preliminary tests on a prototype
implementation of this system show that it can be used to identify
metrics that are present and/or future indicators of degraded media
quality at the receiver, as experienced by the user.
Links of interest
In this joint project with Ikhlaq Sidhu and Jerry Mahler at 3Com, I
measured and characterized videoconferencing traffic on a LAN using
Microsoft's NetMeeting software and coauthored a technical report
on the results. I determined the individual load placed on the network
due to audio traffic and used this to deduce the load due to video and
"collaboration" (document sharing) traffic. I calculated the
statistics and derived distributions of packet sizes, packet
interarrival times, packet arrival rates, and byte arrival rates for
each of the traffic types and for the three phases of the session:
set-up, data transfer, and tear-down. Additionally, I determined the
mix of UDP and TCP packets for each phase of the session. These results
formed the basis for later telepresence work at 3Com.