Distributed CSMA/CA Algorithms for Achieving Maximum Throughput in Wireless Networks

Jian Ni and R. Srikant Coordinated Science Laboratory and Department of Electrical and Computer Engineering University of Illinois at Urbana-Champaign jianni@illinois.edu; rsrikant@illinois.edu August 1, 2009

arXiv:0901.2333v1 [cs.NI] 15 Jan 2009

Abstract Recently, it has been shown that CSMA-type random access algorithms can achieve the maximum throughput in wireless ad hoc networks. Central to these results is a distributed randomized algorithm which selects schedules according a product-form distribution. The product-form distribution is achieved by considering a continuous-time Markov model of an idealized CSMA protocol under which collisions cannot occur. In this paper, we present an algorithm which achieves the same product-form distribu- tion in a discrete-time setting where collision of data packets is avoided through the exchange of control messages (however, the control messages are allowed to collide as in the 802.11 suite of protocols). In our discrete-time model, each time slot consists of a few control mini-slots followed by a data slot. We show that two control mini-slots are sufficient for our distributed scheduling algorithm to realize the same steady-state distribution as in the continuous-time case. Thus, the overhead can be as low as twice the ratio of a control mini-slot to a data slot.

Introduction

In wireless communication networks with limited resources, efficient resource allocation (e.g., power control, link scheduling, routing, rate control) plays an important role in achiev- ing good performance and providing satisfactory quality of service. In this paper, we study link scheduling for wireless networks, where all links (node pairs) may not be able to simul- taneously transmit due to transceiver constraints and/or radio interference. A scheduling algorithm determines which links can transmit data at each time instant so that no two active links interfere with each other. The performance metric of interest here is throughput and we restrict our attention to MAC layer (or link-level) throughput as opposed to end-to-end throughput. It is well known that the queue-length based Maximum Weighted Scheduling (MWS) algorithm is throughput optimal [21], in the sense that it can stabilize the queues in the network for all
Research supported by NSF Grant CNS 07-21286, AFOSR Grant FA9550-08-1-0432 MURI BAA 07-036.18.

and ARO

traffic rates in the capacity region of the network. However, MWS has a high computational complexity (for general interference graphs finding a maximum weighted schedule is an NP- Hard problem). In addition, MWS is not amenable to distributed implementation. These drawbacks greatly limit the deployment of MWS in real networks. Maximal scheduling is a low-complexity alternative to MWS but it may achieve only a small fraction of the capacity region [4, 23]. The Greedy Maximal Scheduling (GMS) algorithm, also known as the Longest-Queue-First (LQF) algorithm, is another natural low-complexity alternative to MWS which performs better than simple maximal scheduling. GMS proceeds in a greedy manner by sequentially adding a link with the longest queue to the schedule and disabling all its interfering links until the schedule is maximal [5, 10, 24]. It was proved in [5] that if the network satisfies the so-called localpooling condition, then GMS is throughput optimal. However, for networks with general topology, GMS may only achieve a fraction of the capacity region [10]. Moreover, while the computational complexity of GMS is low, the signaling and time overhead of decentralization of GMS can scale with the size of the network in the worst-case (depending upon the topology). Another class of scheduling algorithms are CSMA (carrier sense multiple access) type random access algorithms. Under CSMA, a node (sender of a link) will sense whether the channel is busy before it transmits a packet. When a node detects that the channel is busy, it will wait for a random backoff time. Since CSMA type algorithms can be easily implemented in a distributed manner, they are widely used in practice (e.g., IEEE 802.11 MAC protocol). An early paper which derived an analytical model of a CSMA-type algorithm to design link service rates to meet the traffic demand was reported in [2]. The authors of this paper showed that the Markov chain describing the evolution of schedules has a product-form stationary distribution under an idealized CSMA model (zero propagation delay between neighboring nodes which implies zero sensing time, no hidden terminals) where collisions cannot occur. Then they proposed a heuristic algorithm to select the CSMA parameters so that the link service rates are equal to the link arrival rates which were assumed to be known. No proof was given for the convergence of this algorithm. The insensitivity properties of such a CSMA algorithm have been recently studied in [12]. Based on the product-form distribution of the idealized CSMA Markov chain, a distributed algorithm has been developed in [9] to adaptively choose the CSMA parameters to meet the traffic demand without explicitly knowing the arrival rates. The result in [9] makes a time-scale separation assumption, whereby the CSMA Markov chain converges to its steady-state distribution instantaneously compared to the time-scale of adaptation of the CSMA parameters. Then the authors suggest that this time-scale separation assump- tion can be justified using a stochastic-approximation type argument which is verified in [14]. Preliminary ideas for a related result was reported in [17] where the authors study distributed algorithms for optical networks. But it is clear that their model also applies to wireless networks with CSMA. In [18], a slightly modified version of the algorithm proposed in [17] has been shown to be throughput optimal. The key idea in [18] is to choose the CSMA parameters to be a specific function of the queue lengths to essentially separate the time scales of the queue length and the CSMA dynamics. Further, the result in [18] assumes that the maximum queue length in the network is known. In practice, however, the (random) backoff times are not continuous but are measured in multiples of mini-slots. The length of a mini-slot is determined by the typical propagation delay (sensing time) between neighboring nodes, which cannot be chosen arbitrarily 2

small. In this case, collisions are possible which will degrade the throughput performance of the idealized CSMA algorithms in [9] and [18]. The authors of [9] proposed a heuristic to limit

the impact of collisions by placing upper bounds on the CSMA parameters, but the loss in throughput due to this design choice is hard to quantify. In [14], the authors note that if the collision parameters (the inverse of the transmission probability and the packet transmission duration) in [9] are made large, then the impact of collisions on the resulting system can be made arbitrarily small. Our approach is different in that we are able to directly analyze our discrete-time system and quantify the loss in throughput as the ratio of the duration of a control slot to the duration of a data slot (these quantities will be defined later in the paper). While the results in our paper are most closely related to the works in [9] and [18], we also note an important contribution in [15] who showed that CSMA type algorithms are asymptotically throughput optimal for large networks under the primary interference constraint, with many small flows and a small sensing time. This result and the results in [6, 16, 3] also make the important connection between random access algorithms and stochastic loss networks. In this paper, we propose a class of throughput-optimal scheduling algorithms with constant overhead, independent of the size of the network. As in earlier works, our goal is to design a collision-free schedule. However, we allow for collisions during the process of finding a collision-free schedule, thus relaxing the perfect CSMA assumption. We assume a time-slotted system, where each time slot consists of a control slot and a data slot. In the control slot, a random set of links will be selected and only those links have the opportunity to change their states (from inactive to active, or from active to inactive), while other links will keep their state unchanged. The decision of a selected link only depends on its own queue length and the transmission status of its interfering links (which can be obtained by carrier sensing). The scheduling algorithm has the following nice properties: The transmission schedule (the set of links that are active for data transmission) in each time slot evolves as a discrete-time Markov chain (DTMC) when the link activation probabilities are fixed. The DTMC is reversible and has a productform stationary distribution (see Section 3). The scheduling algorithm can be easily implemented in a distributed manner. We propose one distributed implementation in Section 4. This implementation requires constant signalling and time overhead for networks with arbitrary topology. In par- ticular, two control mini-slots are sufficient to realize the product-form stationary distribution. (We note that a constant overhead algorithm that achieves an arbi- trary fraction of the capacity region has been proposed in [19] but that algorithm is applicable only to wireless networks under the primary interference constraint.) Under some assumptions, the scheduling algorithm can be made throughput optimal by appropriately choosing the link activation probabilities as functions of the queue lengths (see Section 5).

2 Model

Network

We model a wireless network by a graph G = (V, E), where V is the set of nodes and E is the set of links. Nodes are wireless transmitters/receivers and there exists a link between 4

two nodes if they can directly communicate with each other. For any link l E, we can

define the set of its interfering links as N (l) = {l E |l interfers with l}. (1) We assume a time-slotted system. A feasible schedule of G = (V, E) is a subset of links that can be activated at the same time according to the interference constraint, i.e., no two links in a feasible schedule interfere with each other. We assume that all links have unit capacity, i.e. a scheduled link can transmit one packet in one time slot. A schedule is represented by a rate vector x {0, 1}|E | . The l th element of x is equal to 1 (i.e., xl = 1) if link is set in the schedule; x 0 otherwise. With that a little bit abuse of notation, we also treat l = x as l if a and write lk x if x(l) 1. Note a feasible schedule has of the property l = that l, k x, then / N and l / N (k). Let M be the set all feasible schedules of the network. A scheduling algorithm is a procedure to decide which schedule should be used (i.e., which subset of links should be activated) in every time slot for data transmission. In this paper we focus on the MAC layer so we only consider single-hop traffic. The capacity region of the network is the set of all arrival rates for which there exists a scheduling algorithm that can stabilize the queues, i.e., the queues are bounded in some appropriate stochastic sense depending on the arrival model used. For the purposes of this paper, we will assume that if the arrival process is stochastic, then the resulting queue length process admits a Markovian description, in which case, stability refers to the positive recurrence of this Markov chain. It is proved in [21] that the capacity region is given by = { | C o(M), < }, (2) where C o(M) is the convex hull of the set of feasible schedules in M. When dealing with vectors, inequalities are interpreted component-wise. We say that a scheduling algorithm is throughput optimal, or achieves the maximum throughput, if it can keep the network stable for all arrival rates in .

The Basic Algorithm

We divide each time slot t into a control slot and a data slot. (Later, we will further subdivide the control slot into control mini-slots.) The purpose of the control slot is to determine the transmission schedule x(t) M used for data transmission in the data slot. To determine the transmission schedule x(t), in the control slot the network first selects a set of links that do not interfere with each other, denoted by m(t). Note that these links also form a feasible schedule, but it is not the schedule used for data transmission. We call m(t) the decision schedule in time slot t. Let M0 M be the set of possible decision schedules. The network selects a decision schedule according to a randomized algorithm, i.e., it selects m M0 with some positive P probability (m), mM0 (m) = 1. Then, the transmission schedule is chosen as follows. For any link l in m, if no links in N (l) are active1 in the previous data slot, then link 6

l is chosen to be active with a certain probability pl and inactive with probability 1 pl in the current data slot. If some link in N (l) is active in the previous data slot, then l will be inactive in the current data slot. Any link not selected by m will maintain its state
We say that a link is active in a data slot if it is included in the transmission schedule for that data slot and inactive otherwise.
1

(active or inactive) from the previous data slot. activation probabilities below.

Conditions on M0 and the link

pl s will be specified later. The algorithm is summarized

Basic Scheduling Algorithm (in Time Slot t) In the control slot, randomly select a decision schedule m(t) M0 with probability (m(t)). l m(t): If no links in N (l) are active in the previous data slot, i.e., j N (l), xj (t 1) = 0 (a) x (t) = 1 with probability pl ,p 0< pl < 1; l = 1 pl .(b)l xl (t) = 0 with probability Else (c) xl (t) = 0. l / m(t) : (d) xl (t) = xl (t 1). In the data slot, use x(t) as the transmission schedule.

First we will show that if the transmission schedule used in the previous data slot and the decision schedule selected in the current control slot both are feasible, then the transmission schedule selected in the current data slot is also feasible. Lemma 1 If x(t 1) M and m(t) M, then x(t) M. Proof Note that x M if and only if l such that xl = 1: xj = 0 for all j N (l). For any l / m(t) such that xl (t) = 1, xl (t 1) = 1 and hence xj (t 1) = 0, j N (l). Now for any j N (l): (1) if j / m(t), then xj (t) = xj (t 1) = 0 based on Step (d) of the scheduling algorithm above; (2) if j m(t), then since l N (j) and xl (t 1) = 1, xj (t) = 0 based on Step (c). Also, note from the scheduling algorithm that, for any l m(t), xl (t) = 1 only if xj (t 1) = 0, j N (l). Since l m(t) and m(t) is feasible, we know N (l) m(t) = . Therefore, for any j N (l), xj (t) = xj (t 1) = 0 based on Step (d). QED Because x(t) only depends on the previous state x(t 1) and some randomly selected decision schedule m(t), x(t) evolves as a discrete-time Markov chain (DTMC). Next we will derive the transition probabilities between the states (i.e., transmission schedules). Lemma 2 A state x M can make a transition to a state x M if and only if x x M and there exists a decision schedule m M0 such that x x = (x \ x ) (x \ x) m, and in this case the transition probability from x to x is given by: X Y m l x\ x p(x, x ) = (m) 8

pl

Y
k x \x

pk Y
j m\(xx )\N (xx )

Y
im(x

x )

pi (3)

pj .

Proof (Necessity ) Suppose x is the current state and x is the next state. x \ x = {l : xl = 1, xl = 0} is the set of links that change their state from 1 (active) to 0 (inactive). x \ x = {k : xk = 0, xk = 1} is the set of links that change their state from 0 to 1. From the scheduling algorithm, a link can change its state only if the link belongs to the decision schedule. Therefore, x can make a transition to x only if there exists an m M0 such that the symmetric difference x x = (x \ x ) (x \ x) m. In addition, since x, x , and x x all belong to M, we have x x M. (Sufficiency ) Now suppose x x M and there exists an m M0 such that x x m. Given m is the selected decision schedule, we can calculate the (conditional) probability five cases: that x makes a transition to x , by dividing the links in m into the following (1) l x \ x : l decides to change its state from 1 to 0, this occurs with probability pl based on Step (b) in the scheduling algorithm; (2) k x \ x: k decides to change its state from 0 to 1, this occurs with probability pk based on Step (a); (3) i m (x x ): i decides to keep its state 1, this occurs with probability pi based on Step (a); (4) e m N (x) where N (x) = lx N (l): e has to keep its state 0, this occurs with probability 1 based on Step (c); (5) j m \ (x x ) \ N (x): j decides to keep its state 0, this occurs with probability pj based on Step (b). Note that m N (x \ x) = because x \ x m, we have m \ (x x ) \ N (x) = m \ (x x ) \ N (x x ). Since each link in m makes its decision independently of each other, we can multiply these probabilities together. Summing up all possible decision schedules, we get the total transition probability from x to x given in (3). QED Proposition A necessary and sufficient the DTMC of the transmission schedule to be1irreducible and aperiodic is condition mM0 m for = E , and in this case the DTMC is reversible and has the following product-form stationary distribution: (x) Z = 1 Y pi , Z ix p
i

(4) (5)

X Y pi = . p
xM ix i

Proof If mM0 m = E, suppose l / mM0 m, then from state 0 the DTMC will never reach a feasible schedule including l (there exists at least one such schedule, e.g., the schedule with only l being active). On the other hand if mM0 m = E, then using Lemma 2 it is easy to verify that state 0 can reach any other state x M with positive probability in a finite number of steps and vice versa. To prove this, suppose x = {l1 , l2 , ..., lm }. Define xi = {l1 , ..., li } for i = 0, ..., m. Note that x0 = 0 and xm = x. Now for 0 i m 1, xi xi+1 = xi+1 M and xi xi+1 = {li+1 }. Since mM0 m = E, there exists an m M0 such that {li+1 } m. Then by Lemma 2, xi can make a transition to xi+1 with positive probability as given

in (3) (where x = xi and x = xi+1 ), hence 0 can reach x with positive probability in a finite

number of steps. The reverse argument is similar. Therefore, the DTMC is irreducible and aperiodic. In addition, if state x can make a transition to state x , then we can check that the distribution in (4) satisfies the detailed balance equation: (x)p(x, x ) = (x )p(x , x), (6)

hence the DTMC is reversible and (4) is indeed its stationary distribution (see, for example, [11]). QED

A Distributed Implementation

In this section we describe a distributed implementation of the basic algorithm. Note that in the control slot of a time slot, once a link knows whether it is included in the decision schedule, it can determine its state for the data slot based on local neighborhood information (i.e., whether its interfering links are active in the previous data slot) which can be acquired by carrier sensing. The key idea is to construct a distributed randomized mechanism to select a (feasible) decision schedule in the control slot. In achieving this we further subdivide the control slot into control mini-slots. Distributed Scheduling Algorithm (at Link l in Time Slot t) 1. l selects a random (integer) backoff time Tl uniformly in [0, Wl 1] and waits for Tl control mini-slots. (The idea is that l will send a message announcing its INTENT to make a decision at the expiry of this backoff time subject to the constraints described in the next steps.) 2. IF l hears l an INTENT message2 a llink Nx(l) before the (Tl + 1)-th control mini-slot, will not be included in from m(t). willin set l (t) = xl (t 1). 3 IF l does not hear an INTENT message from any link in N (l) before the (Tl + 1)-th control mini-slot, l will broadcast an INTENT message to all links in N (l) at the beginning of the (Tl + 1)-th control mini-slot. 3.1 If there is a collision (i.e., if there is another link in N (l) broadcasting an INTENT message in the same control mini-slot), l will not be included in m(t). l will set xl (t) = xl (t 1). 3.2 If there is no collision, l will be included in m(t) and decide its state as follows: if no links in N (l) are active in the previous data slot xl (t) = 1 with probability pl , 0 < pl < 1; xl (t) = 0 with probability pl = 1 pl . else xl (t) = 0. 4 If xl (t) = 1, l will transmit a packet in the data slot.

For example, the INTENT message can be an RTS/CTS pair exchanged by the sender and receiver of a link, and we select the duration of a mini-slot such that every link can hear the INTENT message from its interfering links within one mini-slot.

Lemma 3 Let m(t) produced by of the above distributed scheduling by algorithm is a feasible schedule. M0 be the window set all decision schedules above scheduling algorithm. If the size Wl 2 for everyproduced l, then M0 the =M and distributed mM0 m = E. Proof Under the distributed scheduling algorithm, l will be included in the decision schedule m(t) in the control slot only if it has the (unique) minimum backoff time among its interfering links and successfully broadcasts an INTENT message to all links in N (l) without a collision. This will shut up all links in N (l) so those links will not be included in m(t). Hence m(t) is a feasible schedule. Now for any m M, note that m will be selected in the control slot if T = 0, l m, and Tj = 1, j / m. This occurs with positive probability if the window lsize Wl 2 for every l, i.e., Y 1 (m) Pr Tl = 0, l m; Tj = 1, j / m = > 0. lE Wl Therefore, if the window size Wl 2 for every l, M0 = M and hence mM0 m = E. QED Combining Proposition 1 and Lemma 3 we have the following main result of the paper.

Proposition 2 The 1 distributed scheduling algorithm achieves the product-form distribution given in Proposition if the window size W l 2, l.

Comments on Throughput Optimality

Using the product-form distribution in Proposition 1 or Proposition 2, one can then proceed as in [9] (under a time-scale separation assumption) or as in [18] (without such an assumption) to establish throughput optimality of the scheduling algorithm. We do not pursue such a proof here, but instead we point out an alternative simple proof of throughput optimality under the time-scale separation assumption in [9]. We associate each link l E with a nonnegative weight wl (t) in time slot t. The Maximum Weighted Scheduling (MWS) algorithm selects a maximum weighted schedule in every time slot, i.e., it selects a schedule x (t) in time slot t such that X X wl (t) = max wl (t). (7)
lx (t) xM l x

Let ql (t) be the queue length of link l in the beginning of time slot t. It was proved in [21] that MWS is throughput optimal if we let wl (t) = ql (t). The result is generalized in [7] as follows. Suppose fl : [0, ] [0, ], l E are functions that satisfy the following conditions: (1) fl (ql ) is a nondecreasing, continuous function with limql fl (ql ) = . (2) Given any M1 > 0, M2 > 0 and 0 < < 1, there exists a Q < , such that for all ql > Q and l, we have (1 )fl (ql ) fl (ql M1 ) fl (ql + M2 ) (1 + )fl (ql ). (8)

Now for all l E, let wl (t) = fl (ql (t)). The following result was established in [7]. Theorem 1 ([7]) For a scheduling algorithm, if given any and such that 0 , < 1, there exists a B > 0 such that the scheduling algorithm satisfies the following condition: in any time slot , with a schedule x(t) tM thatprobability satisfies: greater than 1 , the scheduling algorithm chooses X X wl (t) (1 ) max wl (t). lx(t) (9)
xM l x

whenever ||q(t)|| > B , where q(t) = (ql (t) : l E). throughput optimal.

Then the scheduling algorithm is

We assume that, by appropriately choosing the link weight functions fl s, the DTMC converges much faster than the dynamics of the link weights. For example, fl (ql ) = ql with a small is suggested as a heuristic to satisfy the time-scale separation assumption in [9] and fl (ql ) = log log ql is used in the proof of throughput-optimality in [18] to essentially separate the time scales. Here, as in [9], we simply assume the DTMC is in steady-state in every time slot. Proposition 3 Suppose the basic scheduling algorithm presented in Section 3 (resp., the pl distributed scheduling algorithm presented in Section 4) satisfies mM0 m = E . pl = Let ewl (t) , i.e., pl =
wl (t) e wl (t) e +1

. Then the scheduling algorithm is throughput optimal.

Proof We prove the proposition using P Theorem 1. Now given any and such that 0 , < 1. Let W (t) = maxxM lx wl (t). Define n X X = xM : wl (t) < (1 )W o (t) .
l x

Since mM0 m = E, by Proposition 1 the DTMC has the product-form stationary distribution in (4), we have X (X ) = (x)
xX

= =

X 1 Y p l Z Xe
lx P
l x

xX

pl
wl (t)

xX

Z
(t)

|X |e(1)W Z |E | 2 < (t) W e

(10)
P
l x

where (10) is true because |X | |M| 2|E | , and Z > emaxxM

wl (t)

eW

Therefore, if W (t) > 1 1 |E | log 2 + log ,

(t)

. (11)

then (X ) < . Since W (t) is a continuous, nondecreasing function of ql (t), with lim||q(t)|| W (t) = , there exits a B > 0 such that whenever ||q(t)|| > B, (11) holds and then (X ) < . Hence the scheduling algorithm satisfies the condition of Theorem 1 and is throughput optimal. QED Note that pl in the above proposition is identical to the link activation probability in the Glauber dynamics mentioned in [18]. Thus, our algorithm is a generalization of Glauber dynamics where multiple links are allowed to make decisions in a single time slot.

Discussion

The basic idea behind our algorithm can be described in words as follows. Each time slot is divided into a number of control mini-slots followed by a data slot. Each link chooses a random backoff time (measured in control mini-slots). When its backoff time expires, the link decides to grab the channel with a certain probability related to its queue length if no other link in its neighborhood was transmitting in the previous data slot and no other link in its neighborhood had an earlier backoff time. In addition, links which were transmitting in the previous data slot continue to transmit unless they have the smallest backoff time in their neighborhood (in which case, they make a decision to either transmit or not as above). The idea is thus closely related to the ideas in [9, 18] but we explicitly allow for collisions in the control slots. Our distributed scheduling algorithm can be approximated in practice by modifying the IEEE 802.11 CSMA/CA MAC protocol as follows: each sender exchanges an RTS/CTS pair (an approximation of the INTENT message) with its receiver after that channel has been sensed to be idle for a sufficient amount of time (an approximation of giving priority to transmissions in the previous time slot as discussed in the previous paragraph); then the sender will decide to transmit a packet to the receiver with a queue-length based probability as opposed to grabbing the channel definitely in normal 802.11. Similar modifications to 802.11 have been studied in [1, 22] where the back-pressure on a link is used as the weight instead of the queue length. Note that a backoff window size larger than 1 is sufficient for the throughput optimality of the scheduling algorithm. So for a small fixed backoff window size (e.g., 2), the scheduling overhead will be quite small. Our model also allows for different window sizes in different time slots and different links. Thus, the model is also applicable even when an adaptive mechanism is used to change the backoff window size like the binary exponential backoff scheme in 802.11. We believe that it should be straightforward to extend our algorithms to be applicable to networks with multi-hop traffic and congestion-controlled sources (see [13, 8, 20] for related surveys).

References
[1] U. Akyol, M. Andrews, P. Gupta, J. Hobby, I. Saniee, and A. Stolyar. Joint scheduling and congestion control in mobile ad-hoc networks. In Proceedings of IEEE INFOCOM, April 2008. 10 10

[2] R. R. Boorstyn, A. Kershenbaum, B. Maglaris, and analysis in multihop CSMA packet radio networks. Communications, 35(3):267274, March 1987.

V. Sahin. Throughput IEEE Transactions on

[3] C. Bordenave, D. McDonald, and A. Proutiere. Performance of random multi-access algorithms, an asymptotic approach. In Proceedings of ACM Sigmetrics, June 2008. [4] P. Chaporkar, K. Kar, and S. Sarkar. Throughput guarantees through maximal schedul- ing in wireless networks. In Proceedings of 43rd Annual Allerton Conference on Com- munication, Control and Computing, 2005.

[5] A. Dimakis and J. Walrand. Sufficient conditions for stability of longest-queue-first scheduling: Second-order properties using fluid limits. Advances in Applied Probabilities, 38(2):505521, 2006. [6] M. Durvy and P. Thiran. Packing approach to compare slotted and non-slotted medium access control. In Proceedings of IEEE INFOCOM, April 2006.

[7] A. Eryilmaz, R. Srikant, and J. R. Perkins. Stable scheduling policies for fading wireless channels. IEEE/ACM Transactions on Networking, 13(2):411424, April 2005. [8] L. Georgiadis, M. Neely, and L. Tassiulas. Resource allocation and cross-layer control in wireless networks. Foundations and Trends in Networking, 2006.

[9] L. Jiang and J. Walrand. A distributed CSMA algorithm for throughput and utility maximization in wireless networks. In Proceedings 46th Annual Allerton Conference on Communication, Control and Computing, September 2008. [10] C. Joo, X. Lin, and N. B. Shroff. Understanding the capacity region of the greedy maximal scheduling algorithm in multi-hop wireless networks. In Proceedings of IEEE INFOCOM, April 2008. [11] F. Kelly. Reversibility and Stochastic Networks. Wiley, Chichester, 1979. [12] S. C. Liew, C. Kai, J. Leung, and B. Wong. Back-of-the-envelope computation of throughput distributions in CSMA wireless networks. Submitted for publication, http://arxiv.org//pdf/0712.1854. [13] X. Lin, N. B. Shroff, and R. Srikant. A tutorial on cross-layer optimization in wireless networks. IEEE Journal on Selected Areas in Communications, 2006. [14] J. Liu, Y. Yi, A. Proutiere, M. Chiang, and H. V. Poor. Maximizing utility via random access without message passing. Microsoft Research Technical Report, September 2008. [15] P. Marbach, A. Eryilmaz, and A. Ozdaglar. Achievable rate region of CSMA schedulers in wireless networks with primary interference constraints. In Proceedings of IEEE CDC, December 2007.

[16] A. Proutiere, Y. Yi, and M. Chiang. Throughput of random access without message passing. In Proceedings of CISS, March 2008. [17] S. Rajagopalan and D. Shah. Distributed algorithm and reversible network. In Pro11 11

ceedings of CISS, March 2008.

12 12

[18] S. Rajagopalan, D. Shah, and J. Shin. Aloha that works, November 2008. Submitted. [19] S. Sanghavi, L. Bui, and R. Srikant. Distributed link scheduling with constant overhead. In Proceedings of ACM SIGMETRICS, 2007. [20] S. Shakkottai and R. Srikant. Network optimization and control. Foundations and Trends in Networking, pages 271379, 2007. [21] L. Tassiulas and A. Ephremides. Stability properties of constrained queueing systems and scheduling policies for maximal throughput in multihop radio networks. IEEE Transactions on Automatic Control, 37(12):19361948, December 1992. [22] A. Warrier, S. Janakiraman, and I. Rhee. DiffQ: Practical differential backlog congestion control for wireless networks. In Proceedings of IEEE INFOCOM, 2009. [23] X. Wu, R. Srikant, and J. R. Perkins. Queue-length stability of maximal greedy sched- ules in wireless networks. IEEE Transactions on Mobile Computing, pages 595605, June 2007. [24] G. Zussman, A. Brzezinski, and E. Modiano. Multihop local pooling for distributed throughput maximization in wireless networks. In Proceedings of IEEE INFOCOM, April 2008.

13 13