On Leveraging Traffic Predictability in Active Queue Management

1
On Leveraging Traffic Predictability in Active
Queue Management
Yuan Gao, Guanghui He and Jennifer Hou Computer
Science Dept. University of Illinois at
Urbana-Champaign
Presented by Xue Liu 3/5/2002
2
Outline of Presentation

• Summary of different AQM schemes
• Motivation of this paper
• Traffic prediction
• Predictive AQM (PAQM)
• Simulation and Results
• Conclusion

3
Summary of Different AQM Schemes

• AQM
• Equipping core routers with the capability to
  detect incipient congestion and to explicitly
  signal traffic sources before congestion actually
  takes place
• A taxonomy of existing AQM schemes
• Achieve Per-flow fairness
• FRED, BRED
• Decouples congestion index and performance index
• To achieve high link utilization and low queue
  length (packet delay)
• BLUE, REM, PI, AVQ
• Stabilize the instantaneous queue length
• SRED, PAQM

4
Achieve Per-flow Fairness
5
Decouples Congestion Index and Performance Index
6
Stabilizing Queue
7
Motivation (1)

• Design Goals
• Stabilize the queue
• Decouple the congestion index and performance
  index
• Use Internet Traffics LRD characteristics
• Predict traffic arrival
• Use prediction results in AQM to determine the
  packet drop rate

8
Motivation (2)- LRD characteristics

• Internet traffic is observed to exhibit long
  range dependency (LRD) characteristics.
• Burstiness at coarser time scales induces
  extended periods of either over-utilization or
  under-utilization
• Resources have to be reserved with respect to the
  peak rates over a wide range of time scales.
• The existence of nontrivial correlation structure
  at larger time scales can be judiciously
  exploited for better congestion control.
• The correlation structure present in LRD traffic
  can be detected on-line and used to predict the
  future traffic.
• The prediction results can be used to infer the
  optimal operation point. (Dropping Probability)

9
Traffic Prediction Problems

• Traffic quantity to measure and predict?
• Measure and predict the amount of traffic
  averaged over a measurement time interval of
  length t.
• t 0.02-gt0.05 seconds (LRD means t is not
  important)
• How far ahead the prediction step is?
• Two (or one) step(s) in this paper.
• How to select prediction method to use?
• Complexity
• The easier, the better
• Accuracy
• The more accurate, the better

10
Measure of Traffic Prediction and Step

• f(t) is the time series representing the amount
  of traffic between two packet arrivals
• Within the k-th time interval t, suppose there
  are m packets arrival, then
• different from original paper, since m is NOT
  constant, what need to be predicted is
• A router keeps the history of incoming traffic
  for n intervals f(k1), f(k2), ., f(kn) and
  predicts for the next two interval f(kn1),
  f(kn2)
• In fact, one step prediction is used in PAQM

11
Prediction Method-- LMMSE

• Use a Linear Minimum Mean Square Error (LMMSE)
  predictor to predict the incoming traffic for the
  next few intervals (n used as 20)

The calculation of coefficients
Ri is covariance function of time series, can be
estimated gt
12
Why Use LMMSE Predictor (1)

• 3 typical kind of Predictors
• (Fractal) Model Based Predictors
• Fractional Brownian Motion (FBM)
• Fractional AutoRegressive Integrated Moving
  Average (FARIMA)
• Linear Minimum Mean Square Error (LMMSE)
  Predictor
• How to depict the LRD characteristics of Traffic?
• Hurst Parameter H (0, 1)
• Estimation of the Hurst Parameter of Long-Range
  Dependent Time Series, O. Rose, Report No. 137,
  February 1996
• The larger H, the larger of LRD

13
Why Use LMMSE Predictor (2)

• Accuracy criteria

• The most important criterion in choosing a
  predictor is the accuracy.
• Depicts the relative mean square errors v.s. H
  under the three prediction models.
• The three curves are close to each other when H
  lt 0.85.
• In real practice, real traffics H is usually lt
  0.85

14
Why Use LMMSE Predictor (3)

• Complexity Criteria
• To implement the two fractional model-based
  predictors, we need to on-line estimate H and let
  d H-1/2
• In the FBM model, the weight coefficient is
• In the ARIMA model, the weight coefficient is
• In contrast, the LMMSE predictor calculates
  coefficient ais directly from the collected
  traffic samples.

15
Why Use LMMSE Predictor (4)

• Validation Real traffic prediction

• The actual traffic and the traffic estimated
  using LMMSE agree well under the following
  simulation scenario.
• Single bottleneck link of capacity of 20 Mbps,
  and a buffer size of 100 packets.
• Totally 60 connections are established.

16
Prediction in AQM -- Architecture

• Design objective
• Keep the queue length of a router at a stable
  level.
• Reduce the packet loss ratio while sustaining
  throughput.
• Approach used
• Predict the future traffic periodically using
  LMMSE predictor.
• Figure in the prediction result in the
  calculation of the packet dropping probability
  used for the next interval.

17
PAQM Controller Design (1)

• System Equation
• C Rt
• u(k) amount of data that should be dropped in
  the k-th interval
• Q(k) the queue length by the end of the k-th
  interval.
• and the predicted and desired
  queue length by the end of the k-th interval
• Objective Function

Or
18
PAQM Controller Design(2)

• Problem Find uopt such that J(uopt) minu J(u)
• Given the uopt, set the dropping probability in
  the router is
• Solution of uopt M 1, Qopt(k) Q, bi 0

Constraints
19
Dropping Probability Graph (1)
Dropping Prob. v.s. current queue size and
predicted incoming traffic theoretic graph
20
Dropping Probability Graph (2)
Dropping Prob. v.s. current queue size when
predicted incoming is constant a RED like graph
21
Packet Dropping Probability Graph (3)
Packet dropping probability calculated in a ns-2
simulation, as depicted by the theoretic graph
22
Simulation and Results

• Simulation Setup
• NS-2 Simulation
• 50-100 TCP connections are established over a
  bottleneck link of capacity 20 Mbps and generate
  packets using the on-off traffic model.
• Maximum buffer size of each router is 100 packets
  (each of size 1000 bytes) under PAQM, RED, and
  AVQ, and 20 packets under SRED

23
PAQM v.s. RED and SRED (1)
Queue Length Stabilization
24
PAQM v.s. RED and SRED (2)
Queue Length Deviation
25
PAQM v.s. RED and SRED (3)
Packet Loss Ratio
26
PAQM v.s. RED and SRED(4)
Attainable Throughput
27
PAQMSensitivity to Qopt
Packet loss ratio and link utilization v.s. Qopt
Under PAQM, the equilibrium value of congestion
measure is independent of equilibrium performance
measure (packet loss or attainable throughput)
28
PAQM v.s. AVQ (1)
Packet Loss Ratio
29
PAQM v.s. AVQ (2)
Attainable Throughput
30
Conclusion

• Correlation structure in LRD traffic can be
  detected on-line and used to predict the future
  traffic
• Exploiting traffic predictability to enhance the
  performance of AQM
• Stabilize queue length (v.s. SRED)
• Achieve low packet loss ratio and high link
  utilization (v.s. AVQ)
• PAQM generalized RED
• PAQM converts to RED when future traffic is
  constant
• PAQM is orthogonal to REM, PI , AVQ etc
• Future traffic prediction is a new axis (index)