The Impact of Active Queue Management on Multimedia Congestion Control Wu-chi Feng Ohio State University
Multimedia Applications Recent proliferation of streaming multimedia applications VoIP, Real Audio, Real Video, etc. Increasing packet loss rates due to non-adaptive applications Potential for congestion collapse
TCP Congestion Control Instrumental in preventing congestion collapse Cooperative, window-based flow control done at every source Multiplicative decrease of window on packet loss Additive increase of window for congestion avoidance
RED Queue Management IETF/IRTF recommendation to reduce packet loss in the Internet Improves performance of TCP congestion control Mechanisms for identifying and penalizing non-adaptive applications
RED Algorithm Keep exponentially weighted moving average of queue length (EWMA) If (EWMA < min th ) { queue packet } If (min th < EWMA < max th ) –drop packet with probability p drop –p=max p *(EWMA-min th )/(max th -min th ) –p drop = p*f(count) If (EWMA > max th ) {drop packet }
RED Extensions Extensions added to penalize non- adaptive applications RED with penalty box –Keep track of last n packet losses –Penalize flows with the largest number FRED –Restrict buffer occupancy of flows to a fair share
Implications Multimedia applications will need to implement TCP-compatible rate control or else Problem: TCP + RED = Trouble for streaming multimedia applications
Why? Congestion notification (CN) random CN spread out randomly over large time scales Each TCP connection experiences bandwidth jitter over short and long time scales Detrimental to streaming multimedia
Experiment Idealized scenario simulated using ns ( 4 connections over small network –20K drop-tail queues –20K RED queues (min th =5K, max th =15K) –50K RED queues(min th =20K, max th =40K) 10Mbs 45Mbs SourceDestination Run with 50KB RED queues so that buffer size of drop- tail and RED are effectively equal. Effective size of 20KB RED queue is less than 20KB due to early detection Effective size of 50KB RED queue is at least 20KB since min th =20KB Run with 50KB RED queues so that buffer size of drop- tail and RED are effectively equal. Effective size of 20KB RED queue is less than 20KB due to early detection Effective size of 50KB RED queue is at least 20KB since min th =20KB
Drop-tail Queues (20KB) Bandwidth over 1 and 8 sec. intervals
RED queues (20KB) Bandwidth over 1 and 8 sec. intervals
RED queues (50KB) Bandwidth over 1 and 8 sec. intervals
Impact on Multimedia Apps. Increased jitter requires additional buffering at the destination to smooth out Introduces additional delay in playback of stream Detrimental to interactive applications such as video conferencing
Bandwidth Variation (4 sources)
Bandwidth Variation (10 sources) RTOs introduce significant variation
Conclusions Active queue management algorithms cause significant bandwidth variation to TCP- compatible sources Additional buffering and playback delay necessary Requires smoother rate/congestion control algorithms