RED behavior with different packet sizes

Abstract

We consider the adaptation of random early detection (RED) 1/ IntroductionThe random early detection (RED) algorithm is becoming a de-facto standard for congestion avoidance in the Internet and other packet switched networks. RED is an active queue management algorithm that aims at increasing the overall network throughput while maintaining low delays. The main idea behind RED is to prevent from packets being dropped because of buffer overflow by dropping them randomly when the average queue size is above a certain threshold. When packets are dropped from tail -the default drop strategy in packet switched networks-losses are arbitrarily distributed among different competing flows. By randomly dropping packets RED aims at fairly distributing losses in proportion to the amount of bandwidth used by each flow. Another goal of RED is to avoid global synchronization caused by dropping packets in bursts, as it is the case for tail-drop gateways. Adaptive sources (e.g. TCP) reduce their sending rate as a reaction to packet drops, which are considered as an implicit indication of congestion. A more efficient way, referred to as early congestion notification (ECN), to inform sources of the congestion is to mark packets rather than dropping them [9]. This explicit way aims at preventing the extra delay incurred by packet retransmission.As a consequence of the incremental deployment of RED, several algorithms based on RED have been and are still being proposed to improve its performance (e.g. [3,6]). The authors of [2] proposed RED with In and Out (RIO) as an extension to RED to discriminate low priority packets (Out) in times of congestion. It is expected that a variant of RIO will be used in differentiated services (DiffServ) networks as a means to provide different forwarding treatments for different packet priorities.RFC2309 [1] states that RED should be used as the default mechanism for managing queues in routers unless there are good reasons to use another mechanism. To this end, strong recommendations for testing, standardization and widespread deployment of active queue management in routers, to improve the performance of today's Internet are made. &RUUHVSRQGLQJ DXWKRU HPDLO RPDUHOORXPL#DOFDWHOEH 7HO )D[

A novel congestion control algorithm for multimedia stream

“…A specifically network load conditions will cause queue oscillation, lower throughput and delay jitter exacerbated [4]. A variation of RED algorithm is given in [5], which achieve better results and fairness by considering the size of the packet when discarding packets. In [6], the average length of the queue is replaced with the average delay in RED dropping probability calculation.…”

TCP smart framing: a segmentation algorithm to reduce TCP latency

“…Since segment loss probability has a major impact on the performance and it may depend on the segment size, we need to know the relationship between the segment loss probability and the segment size. If RED buffer management schemes are adopted and the queue length is managed in bytes, it is reasonable to assume that the segment loss probability is proportional to the segment size, as suggested in [23]. Then, for example, the probability to lose a small segment, , is about 4 times smaller than the probability of losing a full-size segment .…”

Subsidized RED: an active queue management mechanism for short-lived flows

“…Over the past a few years, much work has been done on network congestion control and different AQMs have been proposed and studied [3][4][5][6]9,[13][14][15][16][17][18]21,22,24,27,29]. Among them, RED is the earliest and best-known.…”

“…Ref. [14] studies RED behaviors with different packet sizes and points out that if the RED drop probability is independent from the packet size, RED discriminates against connections with smaller maximum transfer units (MTUs). RED can also result in a very high Packet Loss Ratio (PLR), and as a consequence low goodput, for connections with higher MTUs.…”