Inferring congestion from delay and loss characteristics using parameters of the three-parameter Weibull distribution

by Ramaisa, Motlalepula.

The increase in the number of services and traffic on the Internet amplifies the need for transport protocols to efficiently utilize network resources. One of the main problems that arise due to increased traffic is congestion. Due to the Internet’s vast size, its rapid growth and heterogeneous nature, the dynamics of the Internet are constantly changing, which makes classical congestion control mechanisms inefficient. This research focused on the end-to-end delay and packet loss characteristics, with regard to their correlation with network congestion. The first part of the work studied the end-to-end internet dynamics. Delay and loss data was gathered from 22 local and overseas sites. This data was analyzed to determine how congestion affects the behavior of these characteristics. Results showed that none of the metrics, including the round-trip time (RTT), loss distance and loss length, could be used in isolation to accurately detect congestion. Other factors which can affect these behaviors, such as distance, were investigated as well. The delay and packet loss characteristics were modeled on the three-parameter Weibull distribution. It was found that for delay, the shape parameter, ?, is less than 2.6 when there is no congestion. A value of ? greater than 3.7 was typically found for a highly congested network path. For moderately congested network path, the value was normally between 2.6 and 3.7. For loss, ? is always less than 2.6, but the scaling parameter, ?, decreases with increasing congestion for loss distance. For loss length, ? increases with increasing congestion. This suggests that ? captures the fact that loss events get bursty with increasing congestion. It was observed that the burst sizes are usually small. A simulation framework for modeling delay and loss as measures for congestion was implemented using a Markov chain and validated. This dissertation suggests that analysis of the Weibull parameter values computed for both packet delay and loss, and the bursty nature of loss, can offer some useful indication regarding the congestion state of the network path. Electrical, Electronic and Computer Engineering 2