Packet-pair bandwidth probing in wired-cum-wireless network paths was tested and analyzed in a C++ simulation environment using link models verified alongside Opnet results. Some major differences were noted between these results and those of pure wired scenarios investigated in earlier work. Attempts were made to use a dynamic Gaussian-mix algorithm to identify data clusters within the bandwidth distribution.

A receiver-based end-to-end scheme for packet-pair end-to-end probing is explored. Dispersion data collected by an agent process at the receiver end of the monitored path is reported back to the controlling manager process at the transmitter via the return path allowing the inter-packet gaps to be dynamically adjusted for optimum measurement. We concentrate particularly on the "independence signature" monitored through the "long" inter-packet gaps (between successive packet pairs) and compared with the corresponding feature in the "short" inter-packet data. A system of dynamic closed-loop monitoring is devised which correctly identifies the narrow-link bottleneck.

As an alternative to pdf approximation methods to identify modes within the dispersion data of a packet-pair network probing experiment, the classical Hough transform is used to identify straight lines within the output vs. input dispersion plot. These lines can then be mapped to the independence, rate, and distribution signatures of the different links within that path.

The TOPP probing algorithm gauges the bandwidth of a network path in terms of the dispersion between packets transmitted in closely-spaced pairs. Its application to wireless networks has been frustrated by the large latency associated with the pre-transmission contention phase. This paper examines the mechanisms involved, and proposes an extended TOPP algorithm to provide consistent estimates of link and available bandwidth in the presence of a large latency.

A brief simulation study of real-time packet dispersion mode-tracking using the Gaussian-mix model (originally devised for real-time background classification in moving pictures) and an adaptation of the kernel-density estimator is presented. The simulated environment consisted of two FIFO store-and-forward nodes where the probe packets interact with Poisson and Pareto-generated cross-traffic with a range of packet sizes. The two models produced broadly similar results, able to track node activity under the dynamically changing conditions associated with the Pareto cross-traffic. The Gaussian model sometimes replaced the primary mode with a double peak, which disappeared when some of the modelpsilas parameters were changed.