R. Srikant University of Illinois at Urbana-Champaign

Slides:



Advertisements
Similar presentations
The Capacity of Wireless Networks Danss Course, Sunday, 23/11/03.
Advertisements

The Capacity of Wireless Networks
Mobility Increase the Capacity of Ad-hoc Wireless Network Matthias Gossglauser / David Tse Infocom 2001.
Delay Analysis and Optimality of Scheduling Policies for Multihop Wireless Networks Gagan Raj Gupta Post-Doctoral Research Associate with the Parallel.
Gibbs sampler - simple properties It’s not hard to show that this MC chain is aperiodic. Often is reversible distribution. If in addition the chain is.
Resource Allocation in Wireless Networks: Dynamics and Complexity R. Srikant Department of ECE and CSL University of Illinois at Urbana-Champaign.
EE 685 presentation Optimal Control of Wireless Networks with Finite Buffers By Long Bao Le, Eytan Modiano and Ness B. Shroff.
DYNAMIC POWER ALLOCATION AND ROUTING FOR TIME-VARYING WIRELESS NETWORKS Michael J. Neely, Eytan Modiano and Charles E.Rohrs Presented by Ruogu Li Department.
Entropy Rates of a Stochastic Process
CS774. Markov Random Field : Theory and Application Lecture 04 Kyomin Jung KAIST Sep
1 Cooperative Communications in Networks: Random coding for wireless multicast Brooke Shrader and Anthony Ephremides University of Maryland October, 2008.
1 Mazes In The Theory of Computer Science Dana Moshkovitz.
*Sponsored in part by the DARPA IT-MANET Program, NSF OCE Opportunistic Scheduling with Reliability Guarantees in Cognitive Radio Networks Rahul.
ADCN MURI Tools for the Analysis and Design of Complex Multi-Scale Networks Review September 9, 2009 Protocols for Wireless Networks Libin Jiang, Jiwoong.
CSE 421 Algorithms Richard Anderson Lecture 4. What does it mean for an algorithm to be efficient?
Dept. of Computer Science Distributed Computing Group Asymptotically Optimal Mobile Ad-Hoc Routing Fabian Kuhn Roger Wattenhofer Aaron Zollinger.
CISS Princeton, March Optimization via Communication Networks Matthew Andrews Alcatel-Lucent Bell Labs.
Lecture 11. Matching A set of edges which do not share a vertex is a matching. Application: Wireless Networks may consist of nodes with single radios,
Complexity 1 Mazes And Random Walks. Complexity 2 Can You Solve This Maze?
Sampling and Approximate Counting for Weighted Matchings Roy Cagan.
Stability and Fairness of Service Networks Jean Walrand – U.C. Berkeley Joint work with A. Dimakis, R. Gupta, and J. Musacchio.
Distributed Scheduling Algorithms for Switching Systems Shunyuan Ye, Yanming Shen, Shivendra Panwar
Mixing Times of Markov Chains for Self-Organizing Lists and Biased Permutations Prateek Bhakta, Sarah Miracle, Dana Randall and Amanda Streib.
Mixing Times of Self-Organizing Lists and Biased Permutations Sarah Miracle Georgia Institute of Technology.
Decentralised load balancing in closed and open systems A. J. Ganesh University of Bristol Joint work with S. Lilienthal, D. Manjunath, A. Proutiere and.
EE 685 presentation Distributed Cross-layer Algorithms for the Optimal Control of Multi-hop Wireless Networks By Atilla Eryılmaz, Asuman Özdağlar, Devavrat.
Computing and Communicating Functions over Sensor Networks A.Giridhar and P. R. Kumar Presented by Srikanth Hariharan.
Delay Analysis for Maximal Scheduling in Wireless Networks with Bursty Traffic Michael J. Neely University of Southern California INFOCOM 2008, Phoenix,
Adaptive CSMA under the SINR Model: Fast convergence using the Bethe Approximation Krishna Jagannathan IIT Madras (Joint work with) Peruru Subrahmanya.
1 Introduction to Approximation Algorithms. 2 NP-completeness Do your best then.
CS774. Markov Random Field : Theory and Application Lecture 21 Kyomin Jung KAIST Nov
1 Oblivious Routing in Wireless networks Costas Busch Rensselaer Polytechnic Institute Joint work with: Malik Magdon-Ismail and Jing Xi.
1 Oblivious Routing in Wireless networks Costas Busch Rensselaer Polytechnic Institute Joint work with: Malik Magdon-Ismail and Jing Xi.
Solution to HW1. Problem 1 Need to find shortest path from a single source s to a single destination d. Have a condition in the Dijkstra algo loop which.
Node-based Scheduling with Provable Evacuation Time Bo Ji Dept. of Computer & Information Sciences Temple University Joint work.
Delay-Based Back-Pressure Scheduling in Multi-Hop Wireless Networks 1 Bo Ji, 2 Changhee Joo and 1 Ness B. Shroff 1 Department of ECE, The Ohio State University.
Markov Chains and Random Walks. Def: A stochastic process X={X(t),t ∈ T} is a collection of random variables. If T is a countable set, say T={0,1,2, …
Graph Colouring L09: Oct 10. This Lecture Graph coloring is another important problem in graph theory. It also has many applications, including the famous.
R. Srikant Coordinated Science Laboratory and Department of Electrical and Computer Engineering University of Illinois at Urbana-Champaign Joint work with.
Approximate Inference: Decomposition Methods with Applications to Computer Vision Kyomin Jung ( KAIST ) Joint work with Pushmeet Kohli (Microsoft Research)
Seminar on random walks on graphs Lecture No. 2 Mille Gandelsman,
Stochastic Optimization for Markov Modulated Networks with Application to Delay Constrained Wireless Scheduling Michael J. Neely University of Southern.
Complexity and Efficient Algorithms Group / Department of Computer Science Testing the Cluster Structure of Graphs Christian Sohler joint work with Artur.
Date: 2005/4/25 Advisor: Sy-Yen Kuo Speaker: Szu-Chi Wang.
2/14/2016  A. Orda, A. Segall, 1 Queueing Networks M nodes external arrival rate (Poisson) service rate in each node (exponential) upon service completion.
NOTE: To change the image on this slide, select the picture and delete it. Then click the Pictures icon in the placeholder to insert your own image. Fast.
TU/e Algorithms (2IL15) – Lecture 12 1 Linear Programming.
Theory of Computational Complexity Probability and Computing Lee Minseon Iwama and Ito lab M1 1.
Theory of Computational Complexity Probability and Computing Chapter Hikaru Inada Iwama and Ito lab M1.
Markov Chains and Random Walks
Markov Chains and Mixing Times
Balaji Prabhakar Departments of EE and CS Stanford University
Approximating the MST Weight in Sublinear Time
Dynamic Graph Partitioning Algorithm
DTMC Applications Ranking Web Pages & Slotted ALOHA
Path Coupling And Approximate Counting
R. Srikant University of Illinois at Urbana-Champaign
IEEE Student Paper Contest
Hidden Markov Models Part 2: Algorithms
Depth Estimation via Sampling
Instructor: Shengyu Zhang
Topology Control and Its Effects in Wireless Networks
Throughput-Optimal Broadcast in Dynamic Wireless Networks
Introduction Wireless Ad-Hoc Network
Capacity of Ad Hoc Networks
Balaji Prabhakar Departments of EE and CS Stanford University
Javad Ghaderi, Tianxiong Ji and R. Srikant
Communication Networks
Optimal Control for Generalized Network-Flow Problems
Presentation transcript:

Mixing TIMES and QuEUE LENGTH BOUNDS FOR GLAUBER DYNAMICS-BASED CSMA Scheduling R. Srikant University of Illinois at Urbana-Champaign Joint work with Libin Jiang, Mathieu Leconte, Jian Ni and Jean Walrand (Berkeley)

Introduction Glauber Dynamics Inspired CSMA Scheduling Algorithms Low complexity, fully distributed Can achieve maximum throughput However, queueing performance is not well understood The mean queue length can be related to the amount of time (mixing time) that the Glauber dynamics takes to converge to equilibrium. We study the mixing time as a function of the nodes in the network.

Conflict Graph of a Wireless Network Each vertex in the conflict graph represents a wireless link. An edge connects two vertices if the corresponding wireless links interfere with each other. Feasible schedule: a set of vertices (links) which are not neighbors in the conflict graph (an independent set). 5 2 7 1 4 6 3 Example of feasible schedule: {1, 4, 7} Represented by a binary vector x = (1, 0, 0, 1, 0, 0, 1) xi=1 if link i is included in the schedule and 0 otherwise

Throughput Optimality Associate each link i with a weight wi . The weight of a schedule x is w(x) = i2x wi. Want the following probability of picking schedule x: In wireless networks, if weights are chosen as appropriate functions of queue lengths, an algorithm which chooses schedules from this distribution is throughput-optimal.

Parallel Glauber Dynamics (PGD) Randomly select an update set m with probability qm: a set of vertices that are allowed to change their states; other vertices do not change their states. For each vertex v 2 m do If no vertices in its neighborhood N(v) were active in the previous slot, v will decide to become active with probability pv=v/(1+v) xv=1 inactive with probability1-pv: xv=0 Else, v will be inactive: xv=0

Illustration of PGD-CSMA Current schedule: x(t)={1, 5} Select an update set: m={3, 5, 6} Allowed decisions for links in m: link 3: x3=0 (no choice) Link 5: x5=0 (w.p. 1-p5) link 6: x6=1 (w.p. p6) Other links’ states unchanged New schedule: x(t+1)={1, 6} 5 2 7 1 4 6 3

Dynamics of Schedules x(t) evolves as a Discrete-Time Markov Chain (DTMC) Proposition (Ni & S.) If the probability of updating every link is positive, the steady-state probability of using schedule x has the following product-form: By letting i =exp(wi), we have Discrete-time version of the algorithm studied by Jiang and Walrand

Main Result Theorem 1: Consider a network with n links and maximum interference degree . If the arrival rate vector  2  for some  < 1/ which is also independent of n, then there exist fugacities  such that the queue lengths are stable, and the expected queue length per link is O(log n) under PGD-CSMA. For comment 2, the trick is to select rho’such that rho < rho’<1/Delta, (rho’/rho)v still strictly lies in (1/Delta) Lambda.

Queue Length Evolution ai(t) xi(t) ai(t+1) xi(t+1) Qi(t) Qi(t+1) slot t slot t+1 slot t+2 Time slotted system; fugacities chosen to ensure stability ai(t): # of packets arriving at link i in slot t xi(t): scheduling variable 2 {0,1} (determined by PGD-CSMA) Qi(t): queue length of link i at the end of slot t Queue dynamics: Exact computation of mean queue lengths appears to be hard because the schedules are correlated both spatially and temporally. So we need to analyze how fast the PGD Markov chain converges to steady-state (mixing time).

Drift Analysis T slots Sample the system once every T=Tmix/ slots (Tmix is the mixing time ; T is sufficiently large for the Markov chain to reach steady-state; thus, the empirical average service given to a link over T time slots is “close” to the steady-state mean) Test function Compute drift: Setting E(drift)=0 yields a bound on E(Qi).

Queue Length Bound Expected queue length is upper bounded by a linear function of the mixing time: Often the conductance method is used to estimate Tmix. Leads to bounds which grow exponentially in n, the number of links in the network. We prove a logarithmic bound on Tmix for graphs with bounded degree using the coupling method

Mixing Time of a Markov Chain Roughly speaking, the time required to reach steady-state The variation distance between two distributions ,  is defined as: The mixing time Tmix of the MC is the time required for the MC to get close to the stationary distribution:

Coupling (X(t), Y(t)) is a coupling if both {X(t)} and {Y(t)} are two “copies” of a Markov chain, and once X(t)=Y(t), then X(t+1)=Y(t+1) henceforth. X(t) 1 1 1 1 1 1 2 2 2 2 2 2 … 3 3 3 3 3 3 Y(t) 4 4 4 4 4 4 t=0 t=1 t=2 t=3 t=4 t=5

Coupling Theorem Let d(x,y) be some distance metric between two states of the Markov chain (NOT the variation distance). Theorem: Suppose there exist a constant  < 1 and a coupling (X(t),Y(t)) of the MC such that Then the mixing time is bounded by Tmix · log(De)/(1-) where D is the ratio between max and min distances between two distinct states.

Path Coupling Theorem Under the coupling theorem, we have to check the condition for all pairs of schedules to estimate . Bubley&Dyer ’97 introduced the path coupling theorem, under which, in our context, we only need to check those x and y which are different at only one link, for example x = (1, 0, 0, 0, 1, 1, 0) y = (1, 0, 0, 0, 1, 0, 0)

Coupling on Conflict Graph Distance metric: weighted Hamming distance with weights f(v) for all vertex v. Coupling: both chains select the same update set and use the same coin toss when a vertex in the update can be added to both schedules X(t) = (1, 0, 0, 0, 1, 1, 0) X(t+1)= (1, 0, 0, 0, 0, 1, 0) Y(t)= (1, 0, 0, 0, 1, 0, 0) Y(t+1)=(1, 0, 0, 0, 0, 1, 0) 5 5 2 2 7 7 1 1 4 4 6 6 3 3 d(X(t+1), Y(t+1)) = 0 d(X(t), Y(t)) = f(6)

Useful Lemma Lemma: Consider a pair of adjacent schedules x and y that differ only at v, we have If v is selected to update (with prob. qv), distance will decrease by f(v) If a neighbor w 2 N(v) is selected to update (with prob. qw) and w decides to become active (with prob. w/(1+w), distance will increase by f(w)

Fast Mixing Theorem: For any weight function f(v)>0 of v2V, let m = min f(v), M = max f(v), D=M/m, if then the mixing time of PGD is bounded by

Condition for Fast Mixing Choose f(v)=dv/qv where dv is the degree of vertex v in the conflict graph, then if v < 1/(dv-1) for all v where M = max f(v), m = min f(v), D = M/m.

Proof of Main Result For bounded-degree conflict graphs (dv · ), using a simple distributed randomized scheme, qv can be lower bounded by some constant (i.e., independent of n), so both M and D can be upper bounded by some constants. When arrival rate vector  2  for some  < 1/ where  is independent of n, then v · 1/( -1)- for some constant , so  can also be lowered bounded by some constant. Therefore, Tmix=O(log n), and by our previous queue length analysis, E[Qi] = O(log n) for every link i.

Summary The average queue length (delay) grows logarithmically with the size of the network when the arrival rate lies in a fraction of the capacity region Fast mixing of Parallel Glauber Dynamics The fraction is lower bounded by 1/, where  is the maximum vertex degree in the conflict graph Polynomial growth in delay can also be shown for the case where the weights are chosen adaptively as a function of queue lengths CSMA can stabilize the network queues for all arrival rates in the capacity region. capacity region Low-delay region When the arrival rate lies in this region, the delay grows logarithmically with the size of the network under CSMA.