Inline bandwidth measurements: Implementation difficulties and their solutions

We have proposed the concept of inline network measurement, which involves the concept of "plugging" an active bandwidth measurement into an active TCP connection. Mechanisms using this method have the advantage of requiring no extra traffic for measuring available bandwidth, whereas other active measurement tools cannot fundamentally avoid adding probing traffic onto the network. However, when the inline network measurement algorithms are implemented in general-purpose computers, some problems arise, such as the clock resolution of the kernel system, Interrupt Coalescence (IC) deployed in network interface cards, and the behavior of TCP receiver. In the present paper, we explain these difficulties and describe our current solutions. Furthermore, we implement the measurement algorithms and the solutions against those problems in a FreeBSD 4.10 kernel system, and present some results on experimental networks. The experimentally obtained results are used to verify the solutions and to confirm the effectiveness of our concept, inline network measurement, on actual networks. We also compare the performance of the packet interval-based approaches and packet-burst interval-based approaches, and demonstrate that using packet-burst for the measurement in high-speed networks is quite effective.