Passive capacity estimation: comparison of existing tools

Measuring paths capacity in the Internet is a problem that has received a significant attention from the research community. Three options exist to estimate capacities: the active method, where measurement packets are sent in the network; the embedded method, where the application traffic is altered so as to give it a desired pattern; and the passive method that aims at extracting paths capacity from traffic traces. The active method is the most accurate, but also the most intrusive. The embedded method is useful for some specific applications, e.g., content distribution. The passive method is the only option in a number of scenarios, like the case of an ISP that wants to troubleshoot the performance of its customers. In this paper, we introduce PPrate, a new passive capacity estimation tool, and compare its performance with the two other existing passive tools devised so far, namely Nettimer and MultiQ. We first compare those three passive tools to one state-of-the-art active tool, Pathrate, using Planetlab. In this specific environment, the three passive tools tend to behave in a similar way, though MultiQ is more prone to generate abnormally high estimates. We next investigate the performance of the passive tools in a more realistic and challenging environment, namely an ADSL platform. In this scenario, working with the TCP ack streams should uncover the ADSL modem downlink capacity, while the TCP data streams should uncover the servers capacity in client/server applications like FTP and HTTP. Our study reveals that Nettimer is unable to work properly on ack streams and tends to underestimate path capacities when working on data streams. In contrast, MultiQ works on both ack and data streams but tends to overestimate path capacities. Overall, PPrate offers a good compromise in most situations.