Control Theoretic Analysis and Design of Active Queue Management

1
Control Theoretic Analysis and Design of Active
Queue Management
Presented by Honggang Zhang Computer Networks
Research Group
2
Outline

• Active Queue Management (AQM)
• Random Early Detect (RED)
• Control Theoretic Analysis of AQM
• Proportional and Integral (PI) AQM
• Self-Tuning AQM
• Differentiated Services
• Adaptive Rate Management (ARM) and AQM

3
Active Queue Management (AQM)

• Problems with DropTail Queue
• Global synchronization
• Bias against bursty traffic
• Full queue increases packet delays, jitter.
• Solution Active Queue Management (AQM)
• Inform sender about incipient congestion before
  buffer overflows so senders react accordingly.
• Maintain high throughput, low queue size

4
RED (Random Early Detect) Mechanism
RED Marking/dropping based on average queue
length x (t)
1
Marking probability p
pmax
thmin
thmax
Average queue length x
x (t) smoothed, time averaged q (t) x (ti 1)
a q (ti d) (1-a) x (ti)
5
Observations on RED

• Tuning RED is difficult
• - sensitive to packet sizes, load levels,
  round trip
• time, etc.
• discontinuity of drop function contributes to
  oscillations.
• use of variable sampling interval d causes
  oscillations.

6
Outline

• Active Queue Management (AQM)
• Random Early Detect (RED)
• Control Theoretic Analysis of AQM
• Proportional and Integral (PI) AQM
• Self-Tuning AQM
• Differentiated Services
• Adaptive Rate Management (ARM) and AQM

7
Control Theoretic Analysis of AQM

• differential equation modeling enables casting in
  classical control systems framework
• scientifically based parameter tuning

8
Control Theoretic Analysis of AQM
AQM
RED controller
9
Control Theoretic Analysis of AQM
AQM
Proportional and Integral (PI) controller
10
Control Theoretic Approach to Designing AQM
Queue length vs. Time

• Working with linearized model
• developed rules for setting RED parameters
• compared PI controller with properly adjusted RED
• ns simulation with time varying http,ftp flows
• PI controller faster response, decouples queue
  size and load level
• PI AQM has been implemented in Linux router.
  Experiments in Internet show PIs better
  performance than RED.

- PI Controller - RED Controller
queue length
time
Detailed Control Theoretic Analysis of TCP/RED
available at http//gaia.cs.umass.edu/papers/paper
s.html
11
Outline

• Active Queue Management (AQM)
• Random Early Detect (RED)
• Control Theoretic Analysis of AQM
• Proportional and Integral (PI) Controller
• Self-Tuning AQM
• Diffserv
• Adaptive Rate Management (ARM)

12
Self-Tuning AQM

• Problems with present AQM policies
• Performance sensitive to parameter settings
• Parameter settings sensitive to network
  conditions
• RTT, load (no. flows N), bandwidth (C)
• Solution Self-Tuning AQM
• Adaptive Control Approach
• Estimate C and N/R from router/AQM behavior
• Automatically tune AQM based on these estimates
• Self-Tuning mechanism applicable to all AQM
  policies (RED, PI, etc.)

13
Self-Tuning AQM
TCP/Queue
Updating Rules of Controller Parameters
Load Estimator
BW Estimator
Estimator
AQM Controller
Detailed STAQM Analysis and Design available at
http//gaia.cs.umass.edu/papers/papers.html
14
STPI vs PI
PI

• 200 second simulation experiment
• Link BW changes
• 15Mbps in 0,100sec
• 90Mbps flows in 101,200sec
• Two-way propagation delay200ms
• Buffer size 800pkts
• Load N60 flows
• ARED
• queue_ref175pkts
• STPI
• queue_ref175pkts

STPI
15
STPI vs ARED
ARED

• 500 second simulation experiment
• Load changes
• 1000 flows in 0,100sec
• 100 flows in 101,200sec
• 1000 flows in 201,500sec
• Two-way propagation delay145ms
• Buffer size 1800pkts
• Link BW300Mbps(37500pkt/sec)
• ARED
• th_min200pkts
• th_max600pkts
• STPI
• queue_ref400pkts

STPI
16
Outline

• Active Queue Management (AQM)
• Random Early Detect (RED)
• Control Theoretic Analysis of AQM
• Proportional and Integral (PI) AQM
• Self-Tuning AQM
• Differentiated Services
• Adaptive Rate Management (ARM) and AQM

17
DiffServ architecture
Edge router - aggregate traffic management -
marks packets as in-profile and out-profile
Core router - per class traffic management -
buffering and scheduling based on marking at
edge - preference given to in-profile
packets - Assured Forwarding
18
Bandwidth guaranteesExact-provisioned

• M aggregates, edge markers, target rates Ai
• single bottleneck, capacity C
• adaptive rate management (ARM) at edges
• monitor achieved thruput
• PI control to adapt ri
• multilevel PI control at routers
• target rates Ai achievable if SAi lt C

19
Future work

• Implementation of Self-Tuning PI AQM in Linux
  router
• Applications to Wireless Networks
• ......

20
The End

• Thank you!

21
Related Papers on PI AQM and STAQM

• C.V. Hollot, V. Misra, D. Towlsey, W. Gong " A
  Control Theoretic Analysis of RED" Proceedings
  of IEEE Infocom 2001
• C.V. Hollot, V. Misra, D. Towlsey, W. Gong
• "On Designing Improved Controllers for AQM
  Routers Supporting TCP Flows "
• Proceedings of IEEE Infocom 2001
• Yossi Chait, C.V. Hollot, V. Misra, D. Towsley,
  Honggang Zhang and John Lui"Providing Throughput
  Differentiation for TCP Flows Using Adaptive
  TwoColor Marking and Multi-Level AQM" Proceedings
  of INFOCOM 2002
• Honggang Zhang, C.V.Hollot, Don Towsley, Vishal
  Misra"A Self-Tuning Structure for Adaptation in
  TCP/AQM Networks" submitted to Globecom 2003