|
Video Access Pattern Analysis |
The characterization of video data stored on the web informs us about the properties of video files but not how these files are accessed. Knowledge of such access patterns is useful in designing a range of applications including video proxy caches (MiddleMan), video servers and video file systems. Hence, we acquired and inspected log file records from an ongoing video on the web experiment at Lulea University, Sweden. This experiment is unique because video material is distributed over a high bandwidth network. Thus, users can make access decisions without the network being a major factor. Given that our study of videos on the web indicates that currently content creators, faced with large Internet latencies, deliberately throttle video sizes and durations in an attempt to reduce download wait times, the Luleå analysis provides insights into the potential behavior of both users and content creators in the Internet of the future where bandwidth is more plentiful. Sample findings included:
Inter-arrival times: median interarrival time of about 400 seconds indicate that requests for videos are nowhere near as frequent as those for HTML documents.
Video browsing patterns: users like to view the initial part of videos in order to determine if they are interested or not. If they like it, they continue watching. Otherwise, they stop.
Temporal Locality: accesses to videos also exhibit strong temporal locality. If a video has been accessed recently, chances are that it’ll be accessed again soon.
Our data analysis utilized scripts written using Tcl and Microsoft Excel spreadsheets..
Analysis overview paper in the works.
Raw trace data in text.
Data on the movies accessed in text and as an Excel spreadsheet.
This research is supported by DARPA (contract N00014-95-1-0799), Intel, Xerox, Microsoft, and Kodak
