This thesis describes two methods collectively called Time Series Generation (TSG) that can be used to generate time series inputs modeling packet loss to test IP-based streaming video software. The TSG methods create packet loss models that recreate the mean, variance, and autocorrelation signatures of an actual trace. The synthetic packet loss traces can have their inherent statistics altered, thus allowing for thorough testing of video software in ways that could not be done on actual networks. The two methods comprising TSG, which are individually called the primary and secondary method, use the principle of iterated uniformity to create a time series that attempts to match mean, variance, and autocorrelation. The two methods differ in their approach to generating autocorrelation. This leads to trade-offs between the two. The TSG methods are embodied in a software program called TSGen. An evaluation of TSGen is conducted, including a comparison with the well-known Autoregressive-To-Anything Generation algorithm (ARTAGEN) method and tool. The details of capturing packets and parsing video frame counts from packet streams are explained and demonstrated. Sixteen video stream traces were collected from a variety of sources and used to evaluate TSGen. Synthetic traces are generated for the sixteen original traces and both their summary statistics and autocorrelation signatures are compared against the originals. One of the sixteen traces is also compared against a synthetic trace generated using the ARTAGEN tool. Twelve out of the sixteen synthetic traces when compared to the actual traces had Least Square Error (LSE) values under 0.1, three were under 0.4, and the remaining one was under 1.1. Nine synthetic traces had their percent error differences between the mean and variance of the synthetic and actual traces below 5%, one was below 7%, four were under 18%, and the two remaining were at 41%. TSGen is able to effectively model autocorrelation, mean, and variance. Additional intangible benefits of TSG include adjustable run time for the matching process, with longer run time equating to better accuracy, and a simple theoretical model that was easily implemented.
| Identifer | oai:union.ndltd.org:USF/oai:scholarcommons.usf.edu:etd-2472 |
| Date | 14 November 2003 |
| Creators | Shahbazian, John N |
| Publisher | Scholar Commons |
| Source Sets | University of South Flordia |
| Detected Language | English |
| Type | text |
| Format | application/pdf |
| Source | Graduate Theses and Dissertations |
| Rights | default |
Page generated in 0.0021 seconds