The packet pair technique estimates the capacity of a path (bottleneck bandwidth) from the dispersion (sp acing) experienced by two back-to-back packets. It has also been claimed that the dispersion of longer p acket bursts ('packet trains') can measure the available bandwidth of a path. This paper examines such p acket pair and packet train dispersion techniques. We first demonstrate that the dispersion of packet pa irs in loaded paths follows a multimodal distribution, and discuss the queueing effects that cause multip le local modes. The path capacity is not necessarily the global mode, and so it cannot be estimated usin g statistical procedures for the most common bandwidth range. The effect of the probing packet size is a lso investigated, showing that the conventional wisdom of using maximum sized packet pairs is not optimal . On the contrary, if the probing packets of different packet pairs are of variable size, the sub-capaci ty local modes become wider and weaker.

We then focus on the dispersion of long packet trains. Increasing the length of the packet train redu ces the measurement variance, but the estimates coverage to a value, referred to as Asymptotic Dispersion Rate (ADR), that is lower than the capacity, and that is not related with the available bandwidth. We s how the effect of cross traffic on the dispersion of long packet trains, and derive analytic expressions for the ADR in certain path configurations.

Putting all pieces together, we present a capacity estimation methodology that has been implemented in a measurement tool called pathrate. We present capacity measurements of several Internet paths u sing pathrate, and evaluate its accuracy and its robustness to cross traffic effects.