Researchers: Hong Li, Ph.D. student & Prof. Lorne Mason
Description:
Problem: The project is to synthesize delay traces that are
statistically similar to the real delay traces given only the sample means of the delay traces. Given measured network delay traces, that were collected
by sending UDP probing packets periodically among globally located monitors, we were able to analyze the delay
traces to come up with a model to synthesize delay traces.
Approach: In order to synthesize a statistically similar trace to a real network
delay trace given the mean delay, we constructed a parametric model for the network delays.
The end-to-end network delay is divided into three parts: the constant propagation delay, the negligible transmission delay and the
random queuing delay.
Then we fitted various distributions to real measurements on various links at various times, and we found that shifted
gamma distribution was a good approximation. The Weibull distribution also gave good fit. Hence the gamma
distribution can model the queuing delay and the total network delay is modeled by a shifted gamma distribution.
 Fig: Distribution fitting to queuing delays
To reconstruct a synthetic delay trace which is statistically similar to the original one from its sample mean,
the parameters of the shifted gamma distribution is estimated. Then we could reconstruct the statistically similar network delay trace
with the same mean network delay as the real measured network delay trace.
Simulation Results:
The real and synthetic end-to-end delays are shown as follows.
Fig: Real end-to-end delay measurements
Fig: Synthetic end-to-end delays
Publications:
H. Li and L. Mason, "Synthesis of network delays for voice packets in Service Overlay Networks", in Proc. IEEE/ACM Qshine, Aug. 2007, Vancouver, BC, Canada.[pdf]
|