Revisiting design choices in queue disciplines: The PIE case

Abstract

Bloated buffers in the Internet add significant queuing delays and have a direct impact on the user perceived latency. There has been an active interest in addressing the problem of rising queue delays by designing easy-to-deploy and efficient Active Queue Management (AQM) algorithms for bottleneck devices. The real deployment of AQM algorithms is a complex task because the efficiency of every algorithm depends on appropriate setting of its parameters. Hence, the design of AQM algorithms is usually entrusted on simulation environments where it is relatively straightforward to evaluate the algorithms with different parameter configurations. Unfortunately, several factors that affect the efficiency of AQM algorithms in real deployment do not manifest during simulations, and therefore, lead to inefficient design of the AQM algorithm. In this paper, we revisit the design considerations of Proportional Integral controller Enhanced (PIE), an algorithm widely considered for network deployment, and extensively evaluate its performance using a Linux based testbed. Our experimental study reveals some performance anomalies in certain circumstances and we prove that they can be attributed to a specific design choice of PIE, namely the use of the estimated departure rate to compute the expected queuing delay. Therefore, we designed an alternative approach based on packet timestamps, implemented it in the Linux kernel and proved its effectiveness through an experimental campaign. © 2020