26
5/4/2006 EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

  • View
    214

  • Download
    0

Embed Size (px)

Citation preview

Page 1: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

5/4/2006 EE228A – Communication Networks 1

Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications

Presented by

Phoebus Chen

Page 2: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 25/4/2006

Outline

Motivation: Sensor Network Surveillance Background: Congestion Control Difficulties with Addressing Latency Design Guidelines for Latency Congestion

Control Policies

Page 3: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 35/4/2006

Sensor Networks for Real-Time Surveillance Event Detection

bursty traffic varying importance

of data for estimation can operate with

incomplete data Low Latency

routing selective packet

delivery congestion control

Page 4: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 45/4/2006

Sample Surveillance Scenario

Multiple targets on linear trajectories

One centralized estimator per cell

Ultimate scenario:Pursuit-Evasion Games with mobile robots

Page 5: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 55/4/2006

Study focused on design of network congestion control Wireless, multi-hop channel Fixed routing Multiple sources, one sink

EstimationSensing and

Data Aggregation

Sensing and Data

Aggregation

(sink)

(source)

(source)

(network)

Page 6: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 65/4/2006

Performance Metric: Estimator Linear System Dynamics

driven by a white noise process Simple linear measurement model

Estimation via Kalman Filter

Check performance under different traffic patterns

Page 7: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 75/4/2006

Background on Congestion Control [1] [2] Flow model Network Optimization Problem

[1] R. Srikant, The Mathematics of Internet Congestion Control, ser. Systems & Control: Foundations & Applications. Birkhauser Boston, 2004.

[2] F. P. Kelly, A. K. Maulloo, and D. K. H. Tan, “Rate control for communication networks: shadow prices, proportional fairness and stability,” Journal of the Operational Research Society, vol. 49, no. 3, pp. 237–252, March 1998.

Page 8: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 85/4/2006

Various User Utility Functions Weighted Proportional Fairness

Minimum Potential Delay Max-Min Fair General Utility Function [3]

For max-min fairness

[3] J. Mo and J. Walrand, “Fair end-to-end window-based congestion control,” IEEE/ACM Transactions on Networking, vol. 8, no. 5, pp. 556–567, Oct 2000.

Page 9: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 95/4/2006

Primal Algorithm and Controller Primal Algorithm (Lyapunov Function)

Flow Controller

kr(xr) > 0 is a non-decreasing, continuous function

Assume prices react instantaneously

Page 10: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 105/4/2006

Dual Algorithm and Controller Dual Algorithm

Price Controller

hl(pl) > 0 is a non-decreasing continuous function

Assume flows react instantaneously

Page 11: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 115/4/2006

Primal-Dual Algorithms and other variants Can combine primal and dual controllers, and

prove via a Lyapunov function that the algorithm is globally, asymptotically stable

Other variants exist Calculate prices using a weighted average of the flow

at a link over time Setting prices based on fullness of a virtual queue

(Adaptive Virtual Queue, or AVQ) Prices are marking probabilities of packets

Page 12: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 125/4/2006

Examples of Congestion Control Analysis Convergence Rate

Linearize about equilibrium Look at smallest eigenvalue of dynamics matrix

Time-delay Stability Analysis Linearize about equilibrium Look at transfer function in the frequency domain and apply

Nyquist stability criterion Stochastic Stability

Linearize about equilibrium Look at Brownian motion perturbations, check induced

covariance of fluctuations

Page 13: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 135/4/2006

Applying TCP/IP congestion control to wireless sensor networks Does not account for

wireless networks with: interference from

neighboring paths physical channel errors Hard to address both, first

pass is to treat as constant error disturbance like [4] [5]

[4] M. Chen, A. Abate, and S. Sastry, “New congestion control schemes overwireless networks: stability analysis,” in Proceedings of the 16th IFACWorld Congress, 2005.[5] A. Abate, M. Chen, and S. Sastry, “New congestion control schemesover wireless networks: delay sensitivity analysis and simulations,” inProceedings of the 16th IFAC World Congress, 2005.

Page 14: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 145/4/2006

Properties of Utility and Pricing Functions

Assumptions on Ur(xr), r is a non-decreasing, continuously

differentiable, strictly concave functionUr(xr) - as xr 0

Assumptions on prices pl() l is a non-decreasing, continuous function such

that

Page 15: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 155/4/2006

Incorporating Latency into Utility Assign a utility to each packet

Sigmoidal function for differentiability

Page 16: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 165/4/2006

Incorporating Latency into Utility (2) Integrate delay utility of each packet with

flow

non-decreasing, continuously differentiable, strictly concave (assuming additional flow only come with greater delay)

May not be able to meet constraint

Ur(xr) - as xr 0

Page 17: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 175/4/2006

Flow Rate vs. Delay and Packet Drop Rate Delay is a function of

queuing delay Congestion Errors from wireless channel CSMA contention

transmission delay (number of hops)

Do not have a good/simple model of CSMA contention at the MAC layer Without knowing we have a hard time knowing

for our optimization problem

Congestion at merge points In routing tree

Page 18: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 185/4/2006

Hope? Congestion control policies as an optimization solver with a black box Some optimization solvers only needs a black box Make delay part of objective function Know general trend D = g(x), delay increases with more flow Treat channel contention, lossy wireless link, inteference, as noise

Lossy Communication

Channel

Lossy Communication

Channel

SourceNodes

RelayNodes

D = g({xr})

Congestion Black Box

{xr} Delay Noise

DD

pl

pl

Page 19: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 195/4/2006

Design Guidelines for Packet Drop Policy May want to use a LIFO queue on a node, to get

latest packets delivered (least delay) Fairness for packets from different merging

routes suggests round robin service over many queues May want to prioritize based on time to last delivered

packet Need to design policy on when to purge LIFO

queues, and how many LIFO queues Parameters of policy set by messages from sink Given vehicle dynamics, sink can determine how

many targets it can track well

Page 20: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 205/4/2006

Design Guidelines for Congestion Feedback Policy Since low network bandwidth, may not want

end-to-end acknowledgement Sparse end-to-end acknowledgement means cannot

adapt to network changes as quickly Types of Information

Queue lengths Number of hops to congestion point Delay on packets delivered

Interfering nodes may want to share information about their respective flow rates and packet delays

Page 21: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 215/4/2006

Design Guidelines for Rate Adaptation Policy Slow start phase? May want evenly spaced samples for

Kalman Filter If within delay constraints, may want to queue

packets to accommodate channel fluctuations How to decode multiple congestion

indicators from relay nodes (queue length, delay, number of hops)?

Page 22: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 225/4/2006

Future Work

Fix a model for simulating the network Design a congestion control scheme via

heuristics, and simulate If I can get a mathematical model, analyze

its stability and convergence

Page 23: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 235/4/2006

Extra Slides

Page 24: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 245/4/2006

Definition of Max-Min Fair

Page 25: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 255/4/2006

What pursuers really see

Page 26: 5/4/2006EE228A – Communication Networks 1 Congestion Control to Reduce Latency in Sensor Networks for Real-Time Applications Presented by Phoebus Chen

EE228A – Communication Networks 265/4/2006

Sensor net increases visibility