An adaptive prediction based approach for congestion estimation in active queue management (APACE)

Abstract

Abstract-I. INTRODUCTION Active Queue Management (AQM) policies attempt to estimate the congestion at a node and signal the incipient congestion by dropping packet(s) before the buffer is full. The main aim of the RED [1] scheme was of providing "congestion avoidance" by dropping packets in anticipation of congestion. The performance of the RED algorithm depends significantly upon the setting of each of its parameters, which appears to be a difficult problem. In [5], Hollot et al. have studied the problem of tuning RED parameters from a control theoretic view point. The aim was to improve the throughput by controlling oscillations in the instantaneous queue. Feng et al. [6] proposed a mechanism for adaptively varying one of the RED parameters, max p , with the aim of reducing the packet loss rates across congested links. Floyd et al. in [7] discuss the algorithmic modifications to the self-configuring RED algorithm [6] for tuning max p adaptively. Despite all these studies, doubts still persist about the useful deployment of RED [8], [9].To address some of the problems of RED, recently there have been some proposals in active queue management. BRED [10] and FRED [11] aim to improve the fairness of RED by maintaining per-active-flow state information. AVQ [3] tries to decouple congestion measure from the performance measure. Predictive AQM (PAQM) [4] tries to exploit traffic predictability in the calculation of the packet dropping probability.In this paper, we propose and analyze a new AQM strategy called APACE and compare its performance with exist-

Target tracking based network Active Queue Management

“…As a common testbed for AQM algorithms [1,5,6] we consider a network with a dumbbell topology where each source node at one side of a bottleneck link is paired with a sink node at the other side of the bottleneck link. The bottleneck link has a capacity of 1Mbps while the other links in the network have a capacity of 3Mbps so that this simulation is identical to the APACE simulation in [6].…”

“…In particular, NLMS was used to adapt the weights on previous measurements of the queue length in order to predict the queue value at a future time. Simulation results in [6] showed that the APACE method was better able to control the instantaneous queue than RED [1], SRED [2] and AVQ [8].…”

“…In [6], the NLMS algorithm [7] from adaptive ltering was used in the APACE AQM method. In particular, NLMS was used to adapt the weights on previous measurements of the queue length in order to predict the queue value at a future time.…”

A novel adaptive traffic prediction AQM algorithm

“…Therefore, prediction method should accurately capture the traffic characteristics. Second, there are many prediction methods, including neural network [4,24], wavelet [25], adaptive filtering [4,17], time series analysis [26, 27], etc. As neural network needs a mass of training samples, it increases time and space complexity.…”

“…Upon receiving this information, the host decreases its sending rate, so congestion can be avoided or mitigated. Current AQM algorithms can be mainly cataloged into following types: queue-size-based [3,4], traffic-rate-based [5][6][7], the combination of queue-size-based and rate-based [8], control-theory-based [9][10][11][12], etc. In fact, each AQM algorithm includes two parts: one part estimates the incipient congestion and the other part makes dropping decision based on the estimation.…”