Analysis of
Active Queue Management
Jae Chung and Mark Claypool
Computer Science Department
Worcester Polytechnic Institute
Worcester, Massachusetts, USA
http://perform.wpi.edu/
Active Queue Management
TCP
AQM
TCP
Advantages
• Reduce packet losses
(due to queue overflow)
• Reduce queuing delay
ACK…
Drop!!!
ACK…
Congestion
Queue
Inbound Link
Router
ACK…
Congestion
Notification…
NCA 2003
WPI
Sink
Outbound Link
Sink
Analysis of AQM #2
Related Work: Analysis of AQM

Control Theory Approach (Hollot, Infocom 01)
– Model TCP and AQM Behaviors (Laplace domain).
– Apply Classical Control Theory (through analysis).
– Params not obvious for understanding control info.

Queue Law (Firoiu, Infocom 00)
– Model Average TCP Throughput and the behavior
of Average Queue Length at congested router.
– Shows the impact of traffic parameters on AQM.
– Not suitable for system stability analysis.
– Good for analyzing and configuring AQMs using
average queue.
NCA 2003
WPI
Analysis of AQM #3
Contribution

Extend Firoiu’s to a General Queue Law
– Simplify the queue law to better illustrate the
effect of traffic parameters on AQM.
– Extend the queue law to support Explicit
Congestion Notification (ECN) and study the
impact of ECN on AQM congestion control.

Analysis of RED-Family AQM
– Evaluate RED-Family AQMs (RED, Gentle RED
and Adaptive RED)
– Demonstrate ECN gains on packet loss rates.
NCA 2003
WPI
Analysis of AQM #4
Outline
Introduction
 Queue Law
 Impact of ECN
 RED Family AQM
 Analysis of REDs
 Summary

NCA 2003
WPI
Analysis of AQM #5
Average TCP Window
(Bulk Transfers)



NCA 2003
Function of congestion notification p
robability (CNP) only.
Not affected by the number of flows
(N), RTT or Service Rate (SR).
Same for ECN and non-ECN flows.
WPI
Analysis of AQM #6
The General
Queue Law
L (1  pd ) 
L : offered traffic Load
AR : Arrival Rate
SR : Service Rate
pd : Pr( drop )
pm : Pr( mark )
N : Number of TCP flows
RTT : Round Trip Time (aggr avg)
RTLD : Round Trip Link Delay (aggr avg)
q : average queue length
N  avg_tcp_wind ( pd  pm )  (1  pd )
AR
(1  pd ) 
SR
RTT  SR
N  avg_tcp_wind ( pd  pm )  (1  pd )

1
( RTLD  q SR )  SR
q  N  avg_tcp_wind ( pd  pm )  (1  pd )  RTLD  SR
NCA 2003
WPI
Analysis of AQM #7
General Queue Law Validation



NCA 2003
Used average TCP throughput mo
del from (Padhye, Sigcomm 1998).
Works well for ECN.
Inaccurate for non-ECN: The TCP
throughput model does not effectiv
ely model Retransmission Timeout.
WPI
Analysis of AQM #8
Effect of Limiting TCP Window


NCA 2003
Window-limited (and short-lived)
flows consume less bandwidth.
Yet, they are less responsive to
congestion notification
WPI
Analysis of AQM #9
Outline
Introduction
 Queue Law
 Impact of ECN
 RED Family AQM
 Analysis of REDs
 Summary

NCA 2003
WPI
Analysis of AQM #10
Impact of ECN on AQM



NCA 2003
ECN has no signaling packet loss.
Requires higher CNP to keep the
average queue at the same level.
Helps the control system stability
as less sensitive to CNP changes.
WPI
Analysis of AQM #11
Outline
Introduction
 Queue Law
 Impact of ECN
 RED Family AQM
 Analysis of REDs
 Summary

NCA 2003
WPI
Analysis of AQM #12
RED-Family AQM
q
2 maxth
Random
Adaptive
Gentle
Adaptive
Early
RED
Detection
RED
RED
(w/ Gentle)
(RED)
Queue law: p = g(q)
RED control function: p = h(q)
maxth
Stable RED operating point
minth
max
variable
maxp
p
NCA 2003
1
WPI
p
Analysis of AQM #13
Simulation Configuration



Dumbbell topology (SR = 20Mbps, Q = 500 pkts)
Increase # of TCP flows (N = 50~300) each 50 sec.
Implicit (drop) and ECN (configurations shown)
q
N=50
N=100
N=150
N=200
N=250
N=300
s
(300)
d
Q = 500 pkts
SR = 20 Mbps
s
maxth
s
r
d
r
d
RTLD = 80 ms
s
d
s
(100)
d
minth
maxp (0.1)
NCA 2003
maxp limit (0.5)
WPI
p
Analysis of AQM #14
Congested
Router Queue
(ECN Results)
NCA 2003
WPI
Analysis of AQM #15
Packet Loss Rate
NCA 2003
WPI
Analysis of AQM #16
Contribution

The General Queue Law
– Illustrated the effect of traffic parameters on AQM.
– Discussed the impact of ECN traffic on AQM.
– Provides mean to configure RED for ECN traffic.

RED-Family AQM Evaluation
– Showed the ECN gain on packet loss rate and the
limitation of AQMs without ECN over Drop-Tail.
– Provided an analysis on RED-Family AQM
• Gentle RED: may result in an unstable queue oscillation.
• Adaptive RED: can support a wide range of traffic loads.
NCA 2003
WPI
Analysis of AQM #17
Future Work

Extend our study to a mixture of ECN and
non-ECN TCP flows.
 Building an adaptive AQM technique that
makes use of our queue law to quickly adapt
to a well-configured state in the presence of
changing network loads.
NCA 2003
WPI
Analysis of AQM #18
Analysis of
Active Queue Management
Jae Chung and Mark Claypool
Computer Science Department
Worcester Polytechnic Institute
Worcester, Massachusetts, USA
http://perform.wpi.edu/