QoS Passive Measurement by Frequency Stability Analysis of Internet Flows

Hiroaki Mihara, Harumoto Fukuda, Satoshi Ono
Affiliation: NTT Information Sharing Platform Laboratories

As the Internet evolves into a global communication infrastructure, there is a growing need for guaranteed quality of service (QoS) and QoS management. To meet this need, several architectures, such as Diffserv and Intserv, have been proposed. In the meantime, Internet service providers will offer customers service-level agreements. This requires a measuring method that can observe the end-to-end qualities of millions of actual flows in the Internet. There are two types of methods for measuring QoS of internet flows. Two-point measurement can precisely measure end-to-end characteristics such as delay, jitter, loss rate, and throughput provided that the clocks at the two measuring points are synchronized, but it is difficult to implement. For example, it is almost impossible to set up measuring points at both the sender and receiver of each flow when there are millions of flows. One-point measurement, on the other hand, is a scalable and easy-to-implement measuring method that can measure millions of flows at one point; however, it cannot measure QoS precisely because of a lack of information, such as the departure and arrival times of a packet. Naturally, one-point measurement cannot directly measure the one-way delay or its variance. Moreover, it has not been proposed as a way of measuring the quality differences of flows having different priorities, such as end-to-end delay and its variance.

We propose a novel method of one-point QoS passive measurement by frequency stability analysis of periodically behaving flows like TCP and constant-bit-rate UDP flows. The key idea is that we consider the changing quality of the end-to-end path as a transfer function from the sender to the receiver. This can be evaluated by the frequency stability analysis of flows provided that the input, or the pattern in which the packets are sent from the sender, are periodical. CBR-UDP flows obviously have a periodical input while TCP flows acquire periodic behavior from TCP self-clocking, so we assume the input is periodic. To evaluate the method, we carried out an experiment on a Diffserv-like network, where weighted fair queuing (WFQ) and priority queuing (PQ) were used to forward packets with different priorities.

In the Diffserv-like experimental network, TCP and CBR-UDP flows having different priorities were measured and the priority differences at the PQ router were found to be the differences in the one-way delay variance obtained by two-point measurement and the differences in frequency stability obtained by one-point measurement. The frequency stability of both TCP and CBR-UDP flows showed a positive correlation to their one-way delay variance.

These results indicate that the priorities of a TCP flow and a CBR-UDP flow in a Diffserv-like network can be evaluated by frequency stability analysis of time series data measured at one point.