Diffusion Mechanisms for Active Queue Management

1
Diffusion Mechanisms for Active Queue Management
Rafael Nunez nunez_at_ece.udel.edu
Gonzalo Arce arce_at_ece.udel.edu

• Department of Electrical and Computer Engineering
• University of Delaware
• Aug 19th / 2004

2
Diffusion Mechanisms for Active Queue Management

• Image Processing Approaches to AQM
• There is an intimate link between printing
  technologies and Active Queue Management.

3
The Internet Today

• TCP de facto congestion control protocol.
• 90 of Internet traffic.

4
Congestion

• Desirable control distributed, simple, stable
  and fair.

5
Simplest Congestion Control Tail Dropping

• Problems with tail dropping
• Penalizes bursty traffic
• Discriminates against large propagation delay
  connections.
• Global synchronization.

6
Active Queue Management (AQM)

• Router becomes active in congestion control.
• Random Early Detection (Floyd and Jacobson,
  1993).
• RED has been deployed in some Cisco routers.

7
Random Early Detection (RED)

• Random packet drops in queue.
• Drop probability based on average queue

• Four parameters
• qmin
• qmax
• Pmax
• wq
• (overparameterized)

8
Queue Behavior in RED
9
Queue Behavior in RED (2)

• 20 new flows every 20 seconds

• Wq 0.01

• Wq 0.001

10
How to overcome these problems

• Adaptive RED, REM, GREEN, BLUE,
• Problems
• Over-parameterization
• Not easy to implement in routers
• Not much better performance than drop tail

11
REM vs. RED
12
Diffusion Mechanisms Exploiting Image Processing

• Our solution
• Based on digital halftoning
• Halftoning is a successful printing technique
  from newspapers to laser printers

13
Digital Halftoning
Error Diffusion
Original Image
Ordered Dither
14
(No Transcript)
15
Probability of Marking a Packet

• Gentle RED function closely follows

(A)
16
Evolution of the Congestion Window

• TCP in steady state

(B)
17
Traffic in the Network

• Congestion Window Packets In The Pipe Packets
  In The Queue
• Or

(C)

• From (A), (B), (C), and knowing that

where
18
Probability Function
19
AQM Dynamics with nonlinearity
20
Error Diffusion

• Packet marking is analogous to halftoning
• Convert a continuous gray-scale image into black
  or white dots
• Packet marking reduces to quantization
• Error diffusion The error between input
  (continuous) and output (discrete) is
  incorporated in subsequent outputs.

21
Diffusion Mechanism

22
Diffusion Mechanism

23
Diffusion Mechanism

24
Diffusion Mechanism

25
Diffusion Mechanism

26
Diffusion Mechanism

27
Diffusion Mechanism

28
Diffusion Mechanism

29
Diffusion Mechanism

30
Diffusion Mechanism

31
Diffusion Mechanism

32
Diffusion Mechanism

33
AQM Dynamics with nonlinearity (2)
34
Algorithm Summary

• Diffusion Early Marking decides whether to mark a
  packet or not as

Where
Remember
M2, b12/3, b21/3
35
Optimizing the Control Mechanism

• Adaptive Threshold Control
• Dynamic Detection of Active Flows

36
Adaptive Threshold Control

• Dynamic changes to the threshold improve the
  quality of the output.

37
Effects of Threshold Modulation in the Control
Mechanism
38
Dynamic Detection of Active Flows

• DEM requires the number of active flows
• Effect of not-timed out flows and flows in
  timeout during less than RTT

39
Dynamic Detection of Active Flows (2)

• The number of packets
• The number of active flows

40
Active Flows Estimate
41
Diffusion Mechanisms for Active Queue Management

• RESULTS

42
Window Size
RED
Diffusion Based
Larger congestion window ? more data!
43
Stability of the Queue
RED
Diffusion Based

• 100 long lived connections (TCP/Reno, FTP)
• Desired queue size 30 packets

44
Changing the number of flows
RED
Diffusion Based

• 20 new flows every 20 seconds

45
Long lived flows
46
Long lived flows (2)
47
Long lived flows (3)
48
Http flows - model

• PackMime traffic model
• Internet Traffic Research group at Bell Labs
• Traffic controlled by the rate parameter (the
  average number of new connections started each
  second)

49
Http flows
50
Http flows (2)
51
Http flows (3)
52
Conclusions

• Digital halftoning is a mature technique that can
  be used in AQM.
• Advantages
• Increased stability
• Simpler (only one parameter)
• Increased throughput
• Current Work
• Parameter optimization
• Complete benchmarking
• Additional traffic control applications

53
Thank you!
Rafael Nunez nunez_at_ece.udel.edu
Gonzalo Arce arce_at_ece.udel.edu

• Department of Electrical and Computer Engineering
• University of Delaware
• Aug 19th / 2004