Ferenc Fejes Gergő Gombos Sándor Laki Szilveszter Nádas
Active queue management (AQM) techniques have evolved in the recent years, after defining the bufferbloat problem. In parallel novel congestion control (CC) algorithms have been developed to achieve better data transport performance, often assuming simple tail dropping buffers. On the other hand, AQM algorithms usually assume legacy CC (Cubic). Though all of the novel AQM and CC algorithms improve the performance under these assumptions, their co-existence has not or only partially been tested so far. Similarly, router buffer sizing was studied mainly in the 2000s, also assuming traditional CC and tail dropping buffers. In this paper we show that combining the different AQMs and CCs evolved in the past few years results in poor fairness, because assumptions used during development do not hold in this heterogeneous case. We also show that while a non-traditional AQM, using core-stateless resource sharing control, has the potential to harmonize conflicting CCs, it suffers from deployment issues. We argue that when dimensioning router buffers in networks where connections with different CCs coexist, the right choice of AQM is more important than the size of the buffer itself.