Experimental testing of the performance of packet dropping schemes

To mitigate bufferbloat in the Internet, which is connected with frequent overfilling of large buffers in routers, IETF recommends application of active queue management methods for queues of packets. An important subclass of such methods exploits the dropping function. Namely, an arriving packet can be dropped randomly, with probability depending on the queue length upon its arrival. Several types of the dropping function have been proposed so far, including a linear one, doubly linear, exponential, quadratic, cubic, mixed and others. In most cases, however, the performance of the queue with a particular form of the dropping function was tested using a simulator, instead of real traffic and devices. Moreover, traffic scenarios were different in different studies. The purpose of this paper is to fill this gap. We present results of laboratory experiments with real traffic and devices. Firstly, all popular dropping function types were implemented in a device equipped with the DPDK technology for fast packet processing. Then, the performance of each dropping function was tested using a hardware traffic generator and a set of different traffic scenarios. In every run, several detailed performance characteristics were obtained. This allowed us to compare different aspects of the performance of different functions and single out their strong and weak points.