Active Queue Management

1
Active Queue Management

• 1. RED 4. AVQ
• 2. BLUE 5. SRED
• 3. REM

2

• TCP congestion model
• Sources decide rates based on feedback
  information
• End-to-End concept
• Positive news ACKs
• Negative news Duplicate ACKs, Transmission
  Timeout, ECN
• Simple networks

3
Drawbacks of Drop Tail

• To accommodate transient congestion periods,
  queue must be large, thus causes delay
• Global synchronization
• Bias against bursty traffic

4
Global Synchronization

• When queue overflows, several connections
  decrease congestion windows simultaneously

5
Bias Against Bursty Traffic

• Bursty traffic more likely to be dropped

average queue length
V.S.
6
Congestion Avoidance

• Maintains low delay and high throughput
• Average queue size kept low
• Actual queue size grows enough to handle
• Bursty Traffic
• Transient Congestion

7
The Solutions of RED

• Global synchronization
• Transient congestion
• (short queue)
• Bursty Traffic

• Drop packets when congestion eminent
• Select packets at random
• Use average queue length as indicator of
  congestion

8
RED Calculating Average Queue Size

• Use low-pass filter (exponential weighted moving
  average)
• wq should be small enough to filter out transient
  congestion, and large enough for the average to
  be responsive

9
Drop Probability
10
Uniform Distribution of Drops

• If you just use pb
• Using pb/(1 - count x pb)

11
Drop Probability Parameters

• To accommodate bursty traffic minth must be large
• maxth depends in part on maximum average delay
  allowed
• (maxth- - minth) must be larger than typical
  increase in average queue length during an RTT

12
Simulation Results
13
Advantages of RED

• No bias against bursty traffic
• No global synchronization
• Packet marking probability proportional to
  connections share of bandwidth
• Scalable no per-connection state

14
BLUE vs RED

• RED relies on queue lengths to estimate
  congestion
• Gives little information about the number of
  competing connections sharing the link
• Requires many parameters
• BLUE relies directly on packet loss and link
  utilization
• Maintains a single probability

15
BLUE
Note d1 gtgt d2
16
REM Overview

• Goals
• Decouples congestion from Performance
• Introduction of Price
• Targets at both Bandwidth and Delay
• Based on mathematical dual model

17
Random Exponential Marking (REM)

• Link algorithm
• Update congestion measure Price
• Mark each packet with probability
• End-to-end marking probability

Clear buffer
Match rate
18
Congestion Measure in REM
19
Marking Probability in REM

• Marking probability
• End-to-end marking probability

20
AVQ

• C Actual link capacity
• C Capacity of virtual queue 0 ltC lt C
• l Actual arrival rate
• Parameters
• g desirable utilization in 0,1
• a to be determined theoretically

21
AVQ

• Each router maintains a virtual Q with 0ltCltC
  and a buffer with sizereal buffer size
• Upon arrival of a real packet, enque the packet
  in both queues if no overflow occurs in the
  v-buffer otherwise, drop the packet.
• Update dC/dt a (g C l)
• lltlt g C virtual Q is depleted quickly ? no
  packets drop
• Lgtgtg C virtual Q is depleted slowly ? drop
  packets

22
Simple Code

• Assume that previous pkt arrives at time s.
• For each arrived packet at time t, do
• VQ ? VQ C(t-s)
• If (VQ current_pkt_size gt buffer_size)
• drop packets in both VQ and real queue
• Else
• VQ ? VQ current_pkt_size
• C ? C a g C (t-s) a current_pkt_size
• s? t

23
SRED Main Idea

• Number of active flows ? number of different
  flows in the buffer
• A misbehaving flow has a lot of packets in the
  buffer
• When a packet arrives, compare it with a packet
  arrived before. If they belong to the same flow,
  a hit occurs.

24
Zombie List

• A list of M recently seen flows, zombies
• Longer memory than the buffer alone
• Information for each zombie
• Count number of packets of this zombie received
• timestamp arrival time of the most recently
  received packet

25
Zombie List Operations

• Zombie list is not full
• Insert the flow with count 0, timestamp ta
• Zombie list is full
• Randomly pick a zombie
• Hit count 1, timestamp ta
• No hit with prob. p that the zombie is replaced
• The arrived packet may be dropped no matter there
  was a hit or not

26
Hits Number of Active Flows

• Zombie list loses memory once every M/p packets
• Few active flows ? more hits
• Misbehaving flows cause more hits than
  well-behaved flows

27
Hit Frequency

• P(t) hit frequency around the time of the tth
  packet arrives at the buffer
• Hit(t) 1 when hit Hit(t) 0, otherwise
• P(t) (1 a)P(t 1) aHit(t)
• Proposition P(t)-1 is a good estimate for the
  effective number of active flows

28
Proposition Argument

• P(arrival packet belongs to flow i) pi
• P(Hit(t)1) S pi2
• 1/N ? S pi2 ? 1
• Symmetric case N flows, pi 1/N
• P(t) 1/N (exact estimate)
• Asymmetric case infinite flows, pi 2-i
• P(t) 3/16 (effective number of active flows)

29
Simple Stabilized RED

• Target buffer occupation Q0
• Set a drop probability p
• Square root law congestion window of each flow,
  cwnd ? p-1/2
• Sum of N congestion windows N p-1/2
• Q0 Np-1/2 ? p (N/Q0)2
• p is proportional to N2

30
Simple Stablized RED (cont.)

• Buffer capacity B
• Current buffer size q
• pzap psred(q)
• psred(q) pmax if 1/3B ? q lt B
• ¼ pmax if 1/6B ? q lt 1/3B
• 0 if 0 ? q lt 1/6B

31
psred(q)

• Depends on current q, not history of q
• Three levels

• Ratio 4 halving the
• congestion windows

32
pzap

• When number of flows ? 256
• pzap psred/65356 (number of flows)2
• When number of flows gt 256
• pzap psred
• Avoid pzap becomes too large
• pzap depends on q and P(t)

33
Full SRED
Increase the drop probabilities of misbehaving
flows