Attila Pasztor and Darryl Veitch
EMULab at Melbourne University
Active measurement, where probe traffic is generated and injected into a network, is becoming increasingly important due its great flexibility, intrinsically end-to-end nature, and freedom from the need to access core network switching elements. Current active measurement programs however are restricted to connectivity, delay and loss statistics on coarse timescales, seconds or more commonly, minutes.
The primary goal of this project is to create a highly accurate active probing system to enable very fine time scale measurements, capable of supporting detailed modelling of end-to-end traffic characteristics, as well as the developpment of tools to measure network quantities such as bottleneck bandwidth reliably.
In existing systems timing problems resulting from inaccurate time-stamping, process scheduling, and lack of synchronisation (in the case of one-way measurements), combine to produce errors from several milliseconds up to seconds. In many cases this constitutes an appreciable fraction of the quantities to be measured. Each of these sources of error is addressed in the present system. The monitoring components at both sender and receiver are based on GPS synchronised DAG3.2e measurement cards with timestamping accurate to 100 ns. The transmission component is based on RealTime Linux transmitting over a 100Mbs Ethernet, and is capable of delivering packets at specified absolute times with an accuracy limited by that of the packet transmission times of the Ethernet card in the 600MHz machine. For a UDP packet of 40bytes containing only a sequence number as payload, this is of the order of 10us. The sender is capable of taking a specification of an arbitrary probe stream, given as a list of packet types, sizes and target departure times, and of sending them consistently with this accuracy with almost no disruption due to kernel scheduling.
A second goal of the project is to investigate the limitations of simpler systems. The variants considered are with or without GPS, using RT Linux timestamping instead of the DAG, and using well designed sender programs in Linux and FreeBSD rather than RT Linux. By benchmarking against the full system, it can be determined which combinations provide adequate performance for different tasks, allowing cheaper and therefore more numerous measurement nodes. The usefulness and impact of NTP is also examined in detail.
The performance of the system is illustrated using data from experiments performed in a small testbed in the laboratory. Measurements using the system have begun between the EMUlab in Melbourne and the WAND group at the University of Waikato. Early results are presented illustrating the potential of the system, where queueing on the route is `visible to the eye'.