Sie sind auf Seite 1von 16

1 Article

2 Distributed Channel-Assignment and Scheduling


3 Algorithms with Low Complexity and Guaranteed
4 Throughput for Multi-Channel Wireless Networks
5 Fan Zhang 1,3,*, Xinyan Xu 2, Liguo Xie 1 and Qing Liu 1
6 1 Postdoctoral Innovation Practice Base, Shandong Cable Interactive Service Ltd, Jinan 250014, China;
7 ccfov@126.com; jngdxlg@163.com; liuqing0385@163.com
8 2 Department of Computer Science and Technology, Shandong College of Electronic Technology, Jinan

9 250200, China; xuxinyan@sdcet.cn


10 3 School of Information Science and Engineering, Shandong University, Qingdao 226237, China;

11 kangli@sdu.edu.cn
12 * Correspondence: ccfov@126.com; Tel.: +86-137-9318-3075
13

14 Abstract: Equipping multi-channel in wireless networks brings not only an improved capacity
15 region but also high challenge in designing joint channel-assignment and scheduling algorithms.
16 With regard to the distributed and online strategies in the context of multi-channel scenarios, the
17 implementation complexity is non-negligible and much higher than that of single-channel cases. In
18 this paper, a distributed resource allocation policy for single-hop traffic flows and its multi-hop
19 multi-path extension are put forward in multi-channel wireless networks, whose complexity is low
20 and independent of both the size of network and the number of available channels. Through
21 theoretical analysis and simulation experiments, it is proved that the proposed algorithms are
22 throughput guaranteed and in some network scenarios, the achieved capacity region can be larger
23 than that of other comparable distributed channel-assignment and scheduling algorithms in the
24 literature.

25 Keywords: multi-channel; channel-assignment; scheduling algorithm; complexity; throughput


26 guaranteed; capacity region.
27

28 1. Introduction
29 In wireless networks, datagrams are transmitted from source to destination by routing selection
30 in network layer associated with scheduling in MAC layer. When the paths for all users are fixed,
31 scheduling policies play a key role in utilizing limited bandwidth effectively to achieve better
32 throughput performance. There have been a number of findings and research results on the topic of
33 improving throughput capacity region via designing efficient scheduling algorithms. It has been
34 demonstrated that the maximum capacity region can be guaranteed by the throughput-optimal
35 scheduling algorithms such as max-weight scheduling [1] for single-path traffic and back-pressure
36 scheduling [2] for multi-path cases. However, such policies require centralized management along
37 with high complexity and as a consequence, it is difficult to implement in real large-scale systems. In
38 order to reduce the computational complexity, a series of distributed maximal scheduling (MS)
39 algorithms that are more applicable in practice have been proposed under various interference
40 models [3, 4]. It has been proved that MS is throughput guaranteed. That is, it can achieve at least a
41 certain fraction of the maximum capacity region. Even so, such maximal-matching-type algorithms
42 need to execute a maximal matching finding process at each time-slot, which requires a remarkable
43 number of iterations increasing logarithmically with the size of the network (i.e., the number of links
44 in the system) [5] and consequently causes non-trivial implementation complexity. Considering such
45 drawback, designing distributed scheduling algorithms with low complexity worth further research.
46 The authors in [6] presented a distributed Q-SCHED algorithm based on random access and backoff
47 time techniques [7], whose throughput performance could be arbitrarily close to that of MS. The
48 complexity of Q-SCHED is low and independent of the network size, which argues its applicability
49 for large-scale network systems.
50 The achievements mentioned above are acquired on single-channel conditions. Recently, the
51 multi-channel technology has attracted a lot of attention since it can significantly increase the
52 capacity region. At present, IEEE 802.11 series has provided resource base for the application of
53 multi-channel, such as 802.11a with 12 non-overlapping channels operating on 5GHz band and
54 802.11b/g with 3 channels on 2.4GHz. In multi-channel environments, channel-assignment and
55 scheduling coordinate and interact with each other to jointly affect the system performance [8-10].
56 Such feature challenges the designing of fully online and distributed algorithms that solve the
57 channel-assignment and scheduling issues. Based on this motivation, in [9], a provably efficient
58 algorithm (referred therein as SP) and its multi-path extension MP have been developed for
59 multi-channel multi-interface (MC-MI) networks. Aiming at achieving reliable throughput capacity
60 region, each link collects queue length and channel rate information from neighborhood in the
61 data-distribution process of SP, which leads to extra overhead inevitably. For the purpose of solving
62 the coupled resource allocation problems systematically in the MC-MI networks, Yu Cheng et al.
63 raised another method, named tuple-based MS, by transforming an MC-MI network node into
64 multiple node-radio-channel tuples [11]. Such framework enables the single-channel MS to be
65 extended straightforward to MC-MI wireless networks with guaranteed throughput performance.
66 Compared with the link-based SP, the tuple-based MS has been shown to ensure a larger capacity
67 region with lower average backlog. However, it could be observed that, with regard to SP, MP and
68 tuple-based MS, they all necessarily attempt to find a maximal schedule at each time-slot, which
69 incurs significant complexity. In fact, in order to compute the maximal matching for all queues, one
70 would have to conduct at least O(log|L||C|) number of iterations [12], where |L| and |C| denote
71 respectively the numbers of links and available non-overlapping channels in the system. Obviously,
72 the complexity of such mechanism also increases with the size of the network and the number of
73 available channels.
74 In view of the above-discussed facts, the goal of this paper is to devise distributed algorithms
75 for wireless multi-channel networks with low complexity. We present a joint channel-assignment
76 and scheduling algorithm, named LDCS, for cases with single-interface nodes. In order to keep the
77 complexity low and independent of both the network size and the number of orthogonal channels,
78 LDCS employs a rate-proportional approach for data flow allocation and extends the idea of the
79 above-mentioned Q-SCHED to fit into multi-channel circumstances. It is theoretically proved that
80 LDCS is throughput guaranteed and the achieved capacity region can be larger than that of other
81 comparable distributed algorithms mentioned above in a variety of scenarios. After that we extend
82 the policy to the cases of multi-path routing with multi-hop flows. The simulation experiments
83 verify our theoretical analysis.
84 So far, there has been a series of developments emerged to design channel-assignment or
85 scheduling algorithms for multi-channel networks from various perspectives. In [13], a DES-Chan
86 framework for distributed channel-assignment based on 2-hop interference model was presented.
87 Such framework did not focus on throughput or scheduling, which differentiates DES-Chan from
88 our work. Some other efforts have been made massively on designing algorithms for OFDM-based
89 multi-channel downlink relay networks [14]-[16]. However in OFDM-based systems, each channel
90 can be occupied by only one transmission link at each time-slot because of opportunism. In contrast,
91 our results are obtained under more generalized models where different links can be operated on
92 the same channel unless they interfere with each other. In addition, for centralized schemes, some
93 appealing results have been achieved with small delay or low complexity [17-18]. Differently, we
94 only consider the distributed and online fashion that is more suitable for large-scale wireless
95 networks in this paper.
96 The rest of the paper is organized as follows. System model and notations are introduced in
97 Section 2. In Section 3, we propose the LDCS algorithm and prove that it can guarantee a certain
98 fraction of the maximum throughput capacity region with low complexity. We next extend our
99 results to multi-hop multi-path cases in Section 4. Simulation results are given in Section 5. Finally,
100 we conclude in Section 6.

101 2. System Model and Notations


102 We consider a single-interface multi-channel wireless network where each node is equipped
103 with one interface which is able to switch among channels dynamically if necessary. Let L and C
104 denote the set of all links and available non-overlapping channels respectively. For any set Γ, |Γ|
105 refers to the cardinality of Γ. With single-hop data flows, each link represents a transmission path
106 for a pair of source and receiver. For each link l, let b(l) and e(l) denote the sending node and the
107 receiving node respectively. Let E(g) be the set of links incident on node g. Radio packets leave the
108 system once they reach the destination. The time is slotted and synchronized for all links in the
109 network. During each time-slot, channel-assignment and scheduling policy jointly decide the sets of
110 scheduled links and occupied channels. Relevant definitions and terminologies are introduced now.
111 A general collision model that has been widely used in practice is employed. That is, each link l
112 corresponds to an interference set Il that includes links interfering with l over the same channel. For
113 convenience, we adopt the convention that l∈Il. If l and another link in the set Il operate on the
114 same channel simultaneously, neither of them can transmit successfully. Assume the interference
115 relationship is symmetric, i.e., l∈Ik if and only if k∈Il. Define the system interference degree K as
116 the maximum number of links that can be activated on a particular channel at the same time in the
117 interference set of any link in the network. Furthermore, due to the single-interface configuration,
118 adjacent links cannot transmit simultaneously even they are scheduled on different channels. Hence,
119 Il can be divided into two incompatible subsets named as interface interference set IInt(l) and channel
120 interference set ICha(l), where IInt(l) consists of links adjacent to l, i.e., IInt(l) = {E(b(l)) ∪ E(e(l))} ∖ {l} and
121 ICha(l) = Il ∖ IInt(l). Here define K1 = maxl∈L K1(l) which can be understood as the interface interference
122 degree of the system. Similarly, in the set ICha(l), use K2(l) to denote the maximum number of links
123 that can operate at the same time using any of the available channels. Let K2 = maxl∈L K2(l) that can
124 be interpreted as channel interference degree.
125 The transmission rate at which link l can operate when scheduled on channel c is denoted by
126 rl . It is assumed that rcl > 0 for any l∈L and c∈C. Because of the presence of channel diversity, each
c

127 link has different rates on different channels and different links have different rates on the same
128 channel. The stochastic arrival process is defined as {Al(n)}, where Al(n) denotes the number of
129 packets arriving at link l at time-slot n. We make some simple assumptions that the arrival process
130 {Al(n)} is i.i.d. across time with average λl, the second moments of the arrival process and the
131 arrivals for any link at each time-slot are both upper bounded. In fact, our method can be extended
132 to suit more general arrival processes. For simplicity, this paper focus only on situations with the
133 assumptions mentioned above. Each link maintains one channel queue for each channel, denoted
134 by (l,c), c∈C. Define ε(l,c) as the set of all channel queues that interfere with (l,c). That is

ε( l ,c)  (k , c) : k  ICha (l)  ( j , c) : j  I Int (l), c  C.

135 The queue length of (l,c) at time-slot n is denoted by qcl (n). Arrivals are distributed to channel
136 queues according to data-flow-allocation algorithm. Use Acl (n) to represent the number of packets
137 assigned by link l to channel c at time-slot n and the expectation is then denoted by λcl = E[Acl (n)].
138 Hence, the evolution of the channel queue (l,c) is given as

qlc (n  1)  [qlc (n)  Alc (n)  Dlc (n)]

139 where [⋅]+ denotes the projection to [0, +∞) and Dcl (n) is the number of packets served by (l,c) in
140 time-slot n via scheduling policy. It is easy to find that Dcl (n) = rcl if (l,c) is scheduled at n, otherwise,
141 Dcl (n) = 0. For convenience, in what follows, we will use the term queue to represent the channel
142 queue. Stability of the system refers to that for any υ > 0, there exists a constant Ψ > 0 such that [19]
 c
2 
lim Pr 
n 
 q (n)
l
   υ
 l L cC 

143 where Pr{S} represents the probability of the event S. As we know, system stability implies 100%
144 throughput with all queue lengths finite. And it can be immediately obtained that if the Markov
145 chain {qcl (n): l∈L, c∈C} is positive recurrent, then the system is stable.
146 Rate vector λ = [λ1 , λ2 ,…, λL ] is exactly the average load of the network. The capacity region is
147 defined as the set of λ that stabilizes the network system under a particular scheduling algorithm.
148 The maximum capacity region achieved by the centralized throughput optimal scheduling [17] is
149 denoted by Ω, which is the union of capacity regions of all policies. For some constant γ∈(0, 1], if
150 the network system remains stable for any λ∈γΩ under a certain algorithm, then the efficiency
151 ratio of the algorithm is said to be γ.

152 3. LDCS and Performance Analysis


153 At each time-slot in LDCS, instead of waiting in a common queue for assignment, packets are
154 distributed to channel queues immediately when arriving, which will lead to a reduced queuing
155 delay [17]. On the other hand, the scheduling of LDCS applies random access and backoff time
156 techniques without finding maximal schedules. The LDCS algorithm consists of two parts given as
157 follows.

158 3.1. Data Flow Allocation


159 For data flow allocation, a local rate-proportional-based mechanism is employed as

rlc
Alc (n)  Al (n) (1)
 dC l
rd

160 which can be realized by using a probability approach. It is easy to see that the process {Acl (n)} is
161 also i.i.d. across time. The allocation process is shown in Figure 1.

162
163 Figure 1. Data flow allocation based on channel queues.

164 Note that each link neither needs to maintain an extra queue to store and relay radios nor
165 requires information from other links, which can decrease effectively the network overhead.

166 3.2. Channel-assignment and Scheduling


167 After allocating the traffic load to different channel queues, the next step is to decide if a queue
168 should be scheduled for transmission in each time-slot. To make this decision, each time-slot is
169 divided into two sub-slots: scheduling-slot and transmission-slot, either of which has a fixed length.
170 In scheduling-slot, each queue decides whether or not to work. In transmission-slot, the queues that
171 are scheduled send their messages. The scheduling-slot is further divided into M mini-slots. Each
172 queue (l,c) selects a backoff time from the set {1, 2,…, M+1} with probability as follows

Pr I c  M  1  e  plc
 
 l
  pc
m 1
 pc
m (2)
 
Pr I lc  m  e l M  e l M

m  1, 2,..., M

173 where Icl represents the backoff time picked by (l,c) and the parameter pcl is computed as

qlc (n)
rlc
plc  α (3)
 qkc (n) 
max   ( k ,c)ε 
( j , d )ε( l ,c )

( j ,d ) rkc 

174 where α = log M. Choosing backoff time M+1 implies that the queue will not be scheduled in this
175 time-slot. When the backoff time expires, the queue begins to enter the transmission-slot to serve
176 unless at least one of its interfering queues has already transmitted. If two or more queues pick the
177 same backoff time, none of them will accomplish transmission successfully. Note that in order to
178 figure out pcl , each queue (l,c) requires the queue length and transmission rate information of all
179 queues in the interference sets of (j,d)∈ε(l,c), which needs information exchange before scheduling
180 implemented by transmitting additional control packets.

181 3.3. Stability Region


182 To establish stability, we introduce two lemmas and construct the Lyapunov function as

 q dj (n) 
V (n)  max  ( j ,d)ε .
lL , cC

( l ,c )
rjd 

183 Lemma 1: At each time-slot n, for any μ > 0, ϑ∈[0,1] and constants C1, C2 > 0, there exists a
184 constant G > 0 such that if V(n) ≥ G, then for any queue (j,d) satisfying

qlc (n)
    V (n)  C1  C2μ  , (4)
( l , c )ε( j ,d )
rlc

185 the LDCS policy guarantees


 log M  1 
 Pr {Slc }    1   μ (5)
( l ,c )ε( j ,d )
 M 
c
186 where Sl denotes the event that queue (l,c) is scheduled. The proof of Lemma 1 can be found in
187 Appendix A.
188 Lemma 2: For any ς > 0, the LDCS ensures that if the offered load satisfies

λcl log M  1
 ( l , c )ε( j ,d )
rlc
 1
M
 4μ

189 for every queue (j,d), then there exists some positive integer H > 0 and a constant Φ > 0 such that if
190 V(n) ≥ Φ, the following inequality

 qlc ( n  H ) 
Pr  (l ,c )ε c
 V (n)  Hμ  1  ς

( j ,d ) rl 
191 holds for all queue (j,d). We present the proof in Appendix B. Now based on Lemma 1 and Lemma 2,
192 we state the following theorem which shows that the system remains stable at average load λ in a
193 certain value range.
194 Theorem 1: Using LDCS mechanism, the irreducible aperiodic discrete state Markov chain
195 {qcl (n): l∈L, c∈C} is positive recurrent if for some μ > 0 and for all queue (j,d),

λcl log M  1
 ( l , c )ε( j ,d ) c
rl
 1
M
 4μ.

196 The proof is presented in Appendix C. Hence, the throughput region guaranteed by LDCS is then
197 obtained through the following proposition.
198 Proposition 1: Appling single-interface wireless nodes, the efficiency ratio γ of LDCS is

τ  log M  1 
γ 1   4μ  (6)
K 2  K1 C  M 

199 where τ  min lL  cC rlc max cC rlc  and the parameter μ can be arbitrarily positive.
200 Proof: According to the definitions of capacity region and efficiency ratio in Section 2, we say
201 that LDCS can achieve γΩ stability region if for any offered load λ such that there exists some other
202 algorithm which can stabilize the system at λ∕γ, the LDCS can ensure the network stable for λ.
203 Under our scenarios, the presence of some algorithm that can stabilize the network at λ∕γ implies
204 the existence of xcl ∈[0, rcl ] for each (l,c) such that for some   0 and for all links l [10],

2 λl
1    x c (7)
γ  cC l

x kc
 r
kIChannel ( l ) cC
c
 K2 (8)
k

x ck
 r
kI Interface ( l ) cC
c
 K1
(9)
k

205 where xcl can be interpreted as an average service rate of the queue (l,c) during a long time. Note
206 that the inequality (7) is due to the rate constraint, while (8) and (9) are due to the interference and
207 the interface constraint respectively. Set β = ( 1   )2∕γ and replace γ with the right side of (6), we
208 have according to the rate-proportional data flow allocation and the inequalities (7)-(9): for all (l,c)
209 and some μ > 0,

λck λck λck


 c kI  (l) rc kI  (l) c
( k , c)ε( l , c ) rk
 
C rk
c
Channel k Interface

λk C λk
  c
  c
kIChannel ( l)  r
cC k kI Interface ( l )  r
cC k

1  λk λk 
 
τ  kIChannel (l ) max rk c
C 
kI Interface ( l ) max rk 
c

 cC cC 
  k 
x c

 k 
x c

11
  (l) max r c kI  (l) max r c 
τ β  kIChannel
cC
 C cC

k k
 
Interface
cC cC

11 x c x ck 
 

    ck  C  r c

τ β  kIChannel ( l ) cC rk kI Interface ( l) cC k 
 
11
  K  C K1 
τ β 2
1  log M  1 
 1 2
M
 4μ 
1     
log M  1
 1  4μ.
M

210 Thus, from Theorem 1, we have proved Proposition 1.


211 3.4. Performance Analysis
212 Firstly, we can infer from (6) that the efficiency ratio of LDCS policy can arbitrarily approach
213 τ∕(K2 + K1|C|) with M→∞ and μ→0. As discussed above, μ can be arbitrarily selected close to 0 and
214 the length of each mini-slot can be controlled to be as small as possible by improving hardware
215 performance so that M→∞ since the length of scheduling-slot is fixed. Therefore, it can be argued
216 that the proposed LDCS can achieve a throughput region of τΩ∕(K2 + K1|C|) approximately.
217 Then we compare the throughput performance of LDCS with other existing distributed and
218 online algorithms. According to [11], tupled-based MS attains a efficiency ratio of 1∕ > 1∕(K + 2),
219 where denotes the interference degree over the tuple-based equivalent model and 1∕(K+2) is the
220 efficiency ratio of SP algorithm [9]. Hence, one can conclude that as M→∞, the capacity region of
221 LDCS is larger than that of tupled-based MS or SP algorithm on condition that the network scenario
222 satisfies τ∕(K2 + K1|C|) > 1∕ > 1∕(K + 2). A simple topology in Figure 2 with single-hop flows and
223 2-hop interference model is illustrated. Suppose there are 14 nodes and 4 available channels with
224 rates equal to 1, 1, 2 and 2packets∕time-slot respectively for all links. Thus, it is not hard to figure
225 out that the efficiency ratio of LDCS equals to 1∕4, whereas that of tupled-based MS and SP are
226 respectively 1∕5 and 1∕6, we have 1∕4 > 1∕5 > 1∕6.

227
228 Figure 2. Network topology with single-hop flows and 2-hop interference model

229 Secondly, the complexity of LDCS is low and independent of the size of the network system as
230 well as the number of available channels. As described before, at each time-slot, to compute a
231 schedule, LDCS merely requires a scheduling-slot of a fixed length independent of both the size of
232 network and the number of available channels. In contrast, the implementation complexity of the
233 maximal-matching-based algorithms [9-10] increases logarithmically with the numbers of links and
234 channels.
235 It is worth mentioning that LDCS only ensures the weakly stability [19] of the system with a
236 possibly unbounded delay experienced by customers and it is designed for multi-channel networks
237 with single-interface wireless nodes.

238 4. Multi-path Extension


239 The discussion above focused on the single-hop flows. For multi-hop flows with multiple paths,
240 a routing selection problem needs to be resolved to ensure effectiveness and fairness. For multi-hop
241 multi-path cases, the arrival process and the maximum capacity region are redefined as follows.
242 Suppose a system composed of S users with I(s) alternate paths for each user s. Let As(n) denote the
243 number of packets offered by user s at time-slot n. It is further assumed that As(n) is i.i.d. across
244 time with average λs. Define Hlsi as the routing indicator variable. If path i of user s passes through
245 link l, Hlsi = 1; otherwise, Hlsi = 0. Thus, the arrivals at link l in time-slot n can be given by
S I ( s)
Al (n)   s  1  i  1 H sil As (n)Psi (n) (10)

246 where Psi(n) denotes the fraction of data packets assigned to path i by user s during time-slot n. The
247 extended maximum throughput region Ω for multi-hop with multiple paths is redefined as the set
248 of [λ1 , λ2 ,…, λS ] such that there exists for all user s and path i satisfying
 S I ( s)
Hl λ P   
  s 1  i 1 si s si  (11)
I ( s)
s.t.  P 1
i  1 si

249 where Psi can be interpreted as the average fraction of traffic assigned to path i on a long-term basis
250 and Ω has been defined by link-based rate vectors λ in Section 2. We now combine the MAC-layer
251 LDCS policy with a routing selection mechanism so that the joint routing, channel-assignment and
252 scheduling algorithm, which is called M-LDCS, can guarantee a certain fraction of the extended
253 maximum throughput region Ω. The same data-flow-allocation strategy as LDCS is utilized and
254 the joint algorithm is performed in two steps as
255 Step 1: Each user computes the fraction vector Ps(n) = [Ps1(n), Ps2(n),…, PsI(s)(n)] at each time-slot
256 n as the solution to the following optimization problem

φ s I ( s) 2
max 
Ps ( n)

2 i 1
 Psi (n) 
θ
I ( s)
H sil  qkc (n) 
  Psi (n) b    c
 (12)
i 1 l L cC  rl ( j , d )ε( l ,c ) 
( k ,c)ε( j ,d ) rk 
bC
I ( s)
s.t.  P (n)  1, P (n)  0
i 1
si si

257 for some θ > 0, where φs is a positive constant chosen for each user s. In fact, the transmitting path is
258 selected by (12). Specifically, when a data packet is generated by user s, it will be assigned to path i
259 with probability Psi(n).
260 Step 2: The same channel-assignment and scheduling algorithm as in Section 3 is employed to
261 determine the set of operating queues and the queue is updated by

S I ( s)
 
  H sil As (n)Psi (n) 
qlc (n  1)  qlc (n)  s 1 i 1  Dl
c
( n)  . (13)
  rlb 
 bC 
 

262 Note that the first term of (12) is to avoid an oscillation problem in routing selection process
263 [20]. In the following proposition, we demonstrate the provably efficiency of M-LDCS.
264 Proposition 2: For any μ > 0, there exists some θ0 > 0 such that for any θ > θ0, the joint routing
265 and LDCS algorithm can ensure γΩ capacity region where

γ 
1  μ  τ  log M  1 
1 . (14)
1  μ   K  K 2 1 
C  M 

266 The proof is presented in Appendix D.

267 5. Simulation Results


268 In this section, we evaluate the performances of the proposed algorithms through simulation
269 experiment. The topology we adopted is shown in Figure 3 which contains 36 nodes (denoted by
270 circles), 60 links (denoted by dashed lines) and 4 available channels with rates equal to 1, 1, 2 and
271 2packets/time-slot respectively. Each node is configured with a single interface that can switch
272 among channels without hindrance. 2-hop interference model is used for simulations. In order to
273 attain the throughput capacity region, we gradually increase the input rates (or offered loads) and
274 observe the experimental data of average backlog. The unit of input rate is packets/time-slot. We
275 first consider the single-hop scenario and compare the developed LDCS (M = 103) with existing SP
276 and tuple-based MS algorithms. As illustrated in Figure 3, there exist 16 traffic flows (represented
277 by arrows) in our environment. Assume that the input rates of all flows are the same, denoted by λ.
278 The average backlog is defined as the mean total backlog in the system divided by the number of
279 traffic flows. Figure 4 shows the performance comparison of the three protocols. As we can see from
280 Figure 4, along with the increasing of offered load, queues and the network become accordingly
281 congested. At the same time, the average backlog grows sharply to infinity when λ approaches a
282 certain value. Such inflection point can be regarded as the throughput capacity region achieved by
283 the corresponding algorithms. In addition, Figure 4 shows the superiority of the proposed LDCS in
284 terms of throughput performance under our scenario, as expected. In fact, it can be said that LDCS
285 guarantees larger capacity region provided that τ∕(K2 + K1|C|) > 1∕ and M→∞. The model we use
286 for experiment is one of the scenarios who satisfy such restriction.

287
288 Figure 3. Network topology for experiments

289
290 Figure 4. Average backlog versus input rates in single-hop scenario

291 We next investigate the performance of the multi-hop multi-path extension (ME) of LDCS
292 using the same topology in Figure 5. Differently, let four source-destination pairs randomly picked
293 across multiple hops replace the 16 single-hop flows. Assume that there exist three alternative paths
294 for each user. Source node selects forwarding path for every data packet according to (12). We let θ
295 be 103 to make sure μ is small enough based on Proposition 2 and set φs = 105 so that the route fraction
296 vectors calculated by (12) are not too sensitive to the queue length updates. In this simulation, the
297 comparison objects are MP and the ME of Tuple-based MS. In practice, multi-path extension
298 implies a cross-layer control method. That is, for the purpose of ensuring throughput, link layer and
299 network layer exchange information to solve a joint routing, channel-assignment and scheduling
300 problem. Figure 4 displays the comparison of three solutions. As expected, ME of LDCS performs
301 much better than MP since it achieves larger capacity region, which confirms theoretical analysis.
302 As we can observe, the capacities attained by ME of LDCS and Tuple-based MS are fairly close.
303 Furthermore, for the same input rates, ME of Tuple-based MS achieves lowest average backlog. The
304 reason for this phenomenon is that ME of Tuple-based MS additionally considers transmission
305 delay and solve the cross-layer problem with path selection convex optimization technique. For MP
306 algorithm, it utilizes a two-stage queuing structure that stores datagrams in a common link-queue
307 firstly and relays them to channel-queues, which aggravates congestion and leads to imperfect
308 performance.

309
310 Figure 5. Average backlog versus input rates in multi-hop multi-path scenario

311 It is worth noting that the number of interfaces and channels would impact the system capacity
312 region inevitably. This paper purely focuses on single-interface scenarios. In reality, most existing
313 communication or sensor systems employed single-interface nodes at present.

314 6. Conclusions
315 This paper presents distributed algorithms for multi-channel single-interface networks with
316 both single-hop and multi-hop multi-path flows, named LDCS and ME of LDCS. It is theoretically
317 demonstrated that the proposed algorithms can achieve guaranteed throughput capacity regions
318 that are comparable with other distributed maximal-matching-based algorithms (such as SP, MP
319 and Tuple-based MS) with low implementation complexity.
320 LDCS utilize a rate-proportional-based mechanism to allocate data locally, and applies random
321 access technique based on probability to complete the channel-assignment and scheduling. Such
322 designing avoids executing maximal-matching-finding process and attains lower complexity that is
323 independent of the numbers of links and channels. It is further proved that LDCS can achieve better
324 throughput performance in some specific situations. LDCS is then extended to be appropriate for
325 multi-hop multi-path scenarios and a cross-layer solution is given. Simulation experiments confirm
326 our analysis. In the future research, we will be interested in improving LDCS to be able to maintain
327 the network under a strongly stable status with bounded average delay for both single-interface and
328 multi-interface cases.
329 Acknowledgments: This work was jointly supported by the National Natural Science Foundation of China
330 under Grant 61501283 and the Shandong Provincial Natural Science Foundation of China under Grant
331 ZR2017BF013.

332 Appendix A
333 For any queue (l,c), according to the backoff time mechanism of LDCS, it will be operated to
334 transmit at each time-slot n if and only if it attempts to participate in the schedule while other queues
c
335 in its interference set picks larger backoff times. Let Sl be the event that (l,c) is scheduled. Hence,
336 from (2), for the event Scl , a lower bound can be confirmed by
M
Pr{Slc }   Pr I lc  m
m 1
  ( j , d )ε( l ,c )
( j ,d) (l ,c )

Pr I jd  m 
Plc [
m
 Pjd ]
(15)
M M
( j , d )ε( l ,c )
 (e M  1) e ,
m 1

337 where the term  ( j ,d)ε Pjd can be upper bounded using the definition of Pjd by
( l ,c )

 ( j , d )ε( l ,c )
Pjd  α. (16)

338 Plug (16) into (15), we have


Plc  m   m 
M  M α Plc M  M α 
Pr {Slc }  (e M  1) m 1 e 
M  m 1

 e .

339 The sum of all Pr{Scl } over (l,c)∈ε(j,d) is then given by


 m 
M  M α   Plc 
 ( l ,c )ε( j ,d )
Pr {Slc }   m 1 e  
  (l,c)ε( j ,d) M  . (17)
 
340 From the assumption (4) made in Lemma 1, we have

qlc (n)
 ( l ,c )ε( j ,d )
rlc V (n)  C1  C2μ (18)
 Plc  α  α .
( l , c )ε( j ,d ) V (n) V (n)

341 Combining (17) and (18) with α = logM, one can have a lower bound of  (l ,c)ε Pr {Slc } by
( j ,d )

 log M  1   C1  C 2μ 
 Pr {Slc }    1    1  V (n)  .
( l , c )ε( j ,d )
 M  

342 For any positive constant G ≥ (C1+C2μ)∕μ, if V(n) ≥ G, then it is easy to verify that

 log M  1   log M  1 
 ( l ,c )ε( j ,d )
Pr {Slc }    1 
M   1  μ    1 
M
 μ.
   
343 This completes the proof of Lemma 1.

344 Appendix B
345 Let B be the upper bound of the difference between the summation of queue lengths in the
346 interference set of any queue at adjacent time-slot. That is,

qlc ( n  1) qlc ( n)
 ( l , c )ε( j ,d )
rlc
  ( l , c )ε( j ,d )
rlc
 B,

347 for all (j,d) and n. Then for any queue (j,d) and any positive integer H, we discuss in two cases. If (j,d)
348 satisfies at time-slot n,

qlc (n)
  V (n)  H  B  μ  ,
( l , c )ε( j ,d )
rlc

349 it is then easy to get that Lemma 2 holds. For the case when

qlc (n)
  V (n)  H  B  μ ,
( l , c )ε( j ,d )
rlc

350 for any time-slot t∈{n+1,…,n+H}, we have


qlc (t)
  V (n)  H  B  μ   B  t  n 
( l , c )ε( j ,d )
rlc
 V (t)  H  3B  μ  .

351 Employ Lemma 1 and let ϑ = 1, one can draw a conclusion that there exists a constant Φ > 0 such
352 that if V(n) ≥ Φ, then
log M  1
 Pr {Slc }  1  μ (19)
( l , c )ε( j ,d )
M
353 holds for all t∈{n+1,…,n+H}. Clearly, Φ is chosen as a function of H. Define an indicator variable
354 M(l,c)(t) such that M(l,c)(t) = 1 in case of queue (l,c) is scheduled at time-slot t, and otherwise M(l,c)(t) = 0.
355 Then from (19), a probability model can be described by

 
E   (l ,c)ε M(l ,c) (t) M( j,d)()
 ( j ,d )
 ( j , d)ε( l ,c ) ,   n  1,..., t  1 

  (l ,c)ε Pr {Slc } (20)
( j ,d )

log M  1
 1  μ.
M
356 Suppose a random variable defined as
M(l ,c ) (t)  1 2 
ρt   (l ,c )ε  0, , ,...,1
( j ,d ) κ  κ κ 

357 where κ is the maximum number of queues that can be scheduled simultaneously in the interference
358 set of any queue in the network. For any φ ≥ 0, based on (20) and the convexity of the exponential
359 function f(x) = e  φx [21], it is not hard to obtain that

1 log M  1 
E  e
 φρt
ρn  1 ,..., ρt 1   1   1 
κ M
 μ  1  e φ .

  (21)

360 Define    tnnH1 ρt . Now we aim to obtain an upper bound of E[ e φ ]. It is well-known that the
361 Chernoff bound for E[ e φ ] can be obtained if the variables ρt’s were independent across time.
362 Obviously, based on the above-mentioned definition of ρt and the Markov characteristics of the
363 random process {M(l,c)(t)}, ρt’s are not independent across time. In order to upper bound E[ e φ ], we
364 introduce a new random variable    tnnH1 χt , where χt ’s are independent and take values of 0 or 1
365 and with probability

1 log M  1 
Pr χt  1   1  μ .
κ M 

Obviously, E[ e φ ] can be upper bounded through Chernoff bound [22] as follows

366
 H  log M  1  

E[ e  φ ]  exp    μ  1 1  eφ  .   (22)
κ  M  

367 Furthermore, based on (21), we can obtain E[ e  φ ]  E[ e  φ ] by induction [23]. So far, we have upper
368 bounded E[ e φ ] by

 H  log M  1  
E[ e  φ ]  exp    μ  1 1  eφ  .   (23)
κ M  
369 Hence, from Markov inequality and (23), we get
  log M  1  
Pr κ  H  1   2μ  
  M  
 H   log M  1   log M  1   
 exp   φ    μ  1  1  e φ φ   
 2μ    .
 κ   M   M   

370 It can be seen that for some φ ≥ 0, there exists a constant ω1 > 0 such that

  log M  1  
Pr κ   H  1   2μ    e  Hω1 . (24)
  M  

371 Similarly, for the arrivals, let

n H Alc (t )
Z   ( l ,c )ε  .
( j ,d ) tn1
rlc

372 Considering the limitation of the offered load formulated in Lemma 2, we can infer for some ω2 > 0,

  log M  1  
Pr Z  H  1   3μ    e  Hω2 . (25)
  M  

373 Hence, based on (24) and (25), we have

 qlc (n  H ) 
Pr  ( l ,c)ε c
 V (n)  Hμ 
 ( j ,d )
rl 
c
 ql (n) 
 Pr Z  V (n)   (l ,c)ε c
 κ  Hμ 
 ( j ,d ) rl 
 Pr Z  κ  Hμ
 1  e  Hω1  e  Hω2 .

374 Thus, by appropriately choosing H such that e  Hω1  e  Hω2  ς , then we can obtain

 qlc (n  H ) 
Pr   (l ,c )ε c
 V ( n)  Hμ   1  ς.

( j ,d ) rl 
375 This ends the proof of Lemma 2.

376 Appendix C
377 Lemma 2 implies that for any ς > 0, there exists a positive constant Φ and a positive integer H
378 such that if V(n) ≥ Φ, we have Pr{V(n + H) − V(n) ≤ −Hμ|q(n)} ≥ 1 − ς. Since the upper bounded
379 difference between the sum of queue lengths in the interference set at adjacent time-slot,

qlc ( n  1) qlc (n)


 (l ,c)ε( j ,d) r c   (l ,c)ε( j ,d) r c  B
l l

380 holds for all (j,d) and n. Hence, we have V(n + H) – V(n) ≤ HB and it can be inferred that


Pr Hμ  V (n  H )  V (n)  HB q(n)  ς. 
381 It is then deduced that

E V (n  H )  V (n) q(n)   Hμ(1  ς)  HBς

382 where q(n) denotes the vector of all the queues. If we choose ς < μ∕(B +μ), we will obtain

E V (n  H)  V (n) q(n)  0


383 for V(n) ≥ Φ. Now the positive recurrence of the Markov chain {q(n)} follows [24] and the system is
384 then stable.

385 Appendix D
386 Referring to the proof process of Proposition 1 and the corresponding definition of the extended
387 maximum capacity region for multi-hop flows with multi-path in (11), the efficiency ratio in (14)
I ( s)
388 implies there is Psi  0 for all user s and path i satisfying  i 1 Psi  1 such that

S I ( s)
  H sil λs Psi 1μ log M  1 
 ( l , c )ε( j ,d )
s 1 i 1
b

1  μ 
1
M . (26)
 bC l
r 

389 For multi-hop multi-path case, the following Lyapunov function is used,
1 θ
1  q dj ( n) 
θ  1  lL  cC   ( j ,d)ε( l ,c) rjd 
U (n)    .
 
390 From (1), (10), (13) and the fact that E[As(n)Psi(n)] = λsPsi(n), we thus have

E U(n  1)  U(n) q(n)


θ
 q dj (n)   Ajd (n) 
  lL  cC   ( j ,d)ε
 d

rj  
  ( j , d )ε
rj d
  ( j ,d )ε Pr Sdj   

( l ,c ) ( l ,c ) ( l ,c )

θ
 q dj (n) 
  lL  cC o   ( j ,d)ε  ,


( l ,c )
rjd 

391 where o(⋅) means higher order. Further on the basis of Lemma 1, it can be easily to obtain that

q dj (n)
 rjd 
( j , d )ε( l ,c )
log M  1 
 ( j , d )ε( l ,c )  
Pr Sdj 
V (n) 

1
M 

392 for all queues (l,c), where V(n) has already been defined before in Section 3. In addition, by the
393 binomial expansion [6], it is not difficult to obtain that for any μ > 0, there exists some θ0 > 0 such that
(1  θ)
 q d (n) 
 lL  cC  ( j ,d)ε(l ,c ) jr d   q dj (n) 
θ

 j   (1  μ) lL  cC   ( j ,d)ε 


V (n) 

( l ,c )
rjd 

394 holds for all θ > θ0. Hence, one can obtain
θ
 q dj (n)  
 lL  cC  ( j ,d)ε(l ,c ) r d   ( j ,d)ε( l ,c) Pr Slc 
  
 j 
 log M  1 
1 M   q dj (n) 
θ 1

 

V (n)  lL  cC  ( j ,d)ε(l ,c ) r d 
 j 
θ
 log M  1   qdj (n) 
 1  μ   1    lL  cC   ( j ,d)ε( l,c) d  .

 M  rj 

395 Then utilizing (26), we have
E U(n  1)  U(n) q(n)
 
 S I ( s) S I ( s) 

θ
qdj (n)    H sij λsPsi(n)  H sij λsPsi 

    
d    s 1 i 1  1 μ    s 1 i 1

lL cC ( j ,d )ε( l,c ) rj
   
( j , d)ε(l ,c ) b C
rlb ( j ,d)ε
(l ,c )  bC l  r b

 
θ
 qdj(n) 
   ο ( j,d)ε 
lL cC  (l ,c ) rjd 
 
 θ

S  I( s) H sil  qkc(n) 
 
 λs  Psi (n)  rb  
 rkc 
s 1  i 1

lL cC  bC l ( j ,d )ε(l ,c )  ( k ,c)ε( j ,d ) 
θ θ
I ( s)
H sil 
 qkc(n)    qdj (n) 
 1  μ   Psi      ο   .
i 1 lL cC  r b 
bC l ( j ,d )ε(l ,c )  ( k ,c)ε( j ,d )
rkc  d
 lL cC ( j ,d)ε(l ,c) rj 
 

396 Based on the routing selection mechanism in (12), it is not hard to found that
θ
I(s)
Hsil  q ck (n) 


i 1
Psi (n) b  
 
lL cC  bC rl ( j , d)ε( l , c ) ( k ,c)ε( j ,d ) rk

c 
 
θ
I ( s) 
H sil qkc (n) 
 1  μ   Psi  b    c 

i 1

lL cC  bC rl ( j , d)ε( l ,c ) ( k , c)ε( j ,d) rk
 
φ s I ( s) 2 φ s I ( s) 2
    Psi (n)     Psi   φ s I( s).
2 i 1 2 i 1

397 Therefore, the Lyapunov drift is negative and is given by

E U(n  1)  U(n) q(n)


 
S I ( s)
 qdj (n) 
θ
 Hsij λsPsi
  s 1 i 1
 μ  
lL cC ( j ,d)ε( l ,c) rj
d  
( j ,d)ε bC rlb
  ( l ,c )
θ
S  qdj (n) 
 λsφs J( s) 
 ο  d
 .

s 1 lL cC  ( j ,d)ε( l ,c) rj 
 

398 The stability of the system then follows [25].

399 References
400 1. Tassiulas, L.; Ephremides, A. Stability properties of constrained queueing systems and scheduling policies
401 for maximum throughput in multihop radio networks. IEEE Trans. Autom. Control 1992, 37, 1936-1948.
402 10.1109/9.182479.
403 2. Tassiulas, L. Scheduling and performance limits of networks with constantly changing topology. IEEE
404 Trans. Inf. Theory 1997, 43, 1067-1073. 10.1109/18.568722.
405 3. Lin, X.; Shroff, N.B. The impact of imperfect scheduling on cross-layer congestion control in wireless
406 networks. IEEE/ACM Trans. Netw. 2006, 14, 302-315. 10.1109/TNET.2006.872546.
407 4. Joo, C.; Lin, X.; Shroff, N.B. Understanding the Capacity Region of the Greedy Maximal Scheduling
408 Algorithm. IEEE/ACM Trans. Netw. 2009, 17, 1132-1145. 10.1109/TNET.2009.2026276.
409 5. Israeli, A.; Itai, A. A fast and simple randomized parallel algorithm or maximal matching. Inf. Process. Lett.
410 1986, 22, 77-80. 10.1016/0020-0190(86)90144-4.
411 6. Gupta, A.; Lin, X.; Srikant, R. Low-Complexity Distributed Scheduling Algorithms for Wireless Networks.
412 IEEE/ACM Trans. Netw. 2009, 17, 1846-1859. 10.1109/TNET.2009.2021609.
413 7. Lin, X.; Rasool, S.B. Constant-time distributed scheduling policies for ad hoc wireless networks. IEEE
414 Trans. Autom. Control 2009, 54, 231-242. 10.1109/TAC.2008.2010888.
415 8. Kyasanur, P.; So, J.; Chereddi, C.; Vaidya, N.H. Multichannel mesh networks: challenges and protocols.
416 IEEE Commun. Mag. 2006, 13, 30-36. 10.1109/MWC.2006.1632478.
417 9. Lin, X.; Rasool, S.B. Distributed and Provably-Efficient Algorithms for Joint Channel-Assignment,
418 Scheduling and Routing in Multi-Channel Ad Hoc Wireless Networks. IEEE/ACM Trans. Netw. 2009, 17,
419 1874-1887. 10.1109/TNET.2009.2021841.
420 10. Bhandari, V.; Vaidya, N.H. Scheduling in multichannel wireless networks. Distrib. Comput. Netw. 2010,
421 5935, 6-17. 10.1007/978-3-642-11322-2_6.
422 11. Cheng, Y.; Li, H.; Shila, D.M.; Cao, X. A Systematic Study of Maximal Scheduling Algorithms in
423 Multiradio Multichannel Wireless Networks. IEEE/ACM Trans. Netw. 2015, 23, 1342-1355.
424 10.1109/TNET.2014.2324976.
425 12. Jung, K.; Shah, D. A distributed matching algorithm; Technical Report; MIT Press, Cambridge, MA, USA,
426 2006.
427 13. Juraschek, F.; Güneş, M.; Philipp, M.; Blywis, B.; Hahm, O. DES-Chan: A Framework for Distributed
428 Channel Assignment in Wireless Mesh Networks. Proceedings of the IEEE Telecommunication Networks
429 and Applications Conference (ATNAC), Melbourne, VIC, Australia, 9-11 Nov. 2011.
430 14. Bodas, S.; Shakkottai, S.; Ying, L.; Srikant, R. Low-complexity scheduling algorithms for multi-channel
431 downlink wireless networks. IEEE/ACM Trans. Netw. 2012, 20, 1608-1621. 10.1109/TNET.2012.2185709.
432 15. Moharir, S.; Shakkottai, S. MaxWeight Versus BackPressure: Routing and Scheduling in Multichannel
433 Relay Networks. IEEE/ACM Trans. Netw. 2015, 23, 1584-1598. 10.1109/TNET.2014.2343992.
434 16. Ji, B.; Gupta, G.R.; Lin, X.; Shroff, N.B. Achieving Optimal Throughput and Near-Optimal Asymptotic
435 Delay Performance in Multichannel Wireless Networks with Low Complexity: A Practical Greedy
436 Scheduling Policy. IEEE/ACM Trans. Netw. 2015, 23, 880-893. 10.1109/TNET.2014.2313120.
437 17. Choi, J.G.; Choi, G.S. Optimal Scheduler with Simplified Queue Structure for Multi-Channel Wireless
438 Networks. IEEE Commun. Lett. 2011, 15, 962-964. 10.1109/LCOMM.2011.072011.110852.
439 18. Li, M.; Salinas, S.; Li, P.; Huang, X.; Fang, Y.; Glisic, S. Optimal Scheduling for Multi-Radio Multi-Channel
440 Multi-Hop Cognitive Cellular Networks. IEEE/ACM Trans. Mobi. Comput. 2015, 14, 139-154.
441 10.1109/TMC.2014.2314107.
442 19. Leonardi, E.; Mellia, M.; Neri, F.; Marsan, M.A. On the stability of input-queued switches with speed-up.
443 IEEE/ACM Trans. Netw. 2001, 9, 104-118. 10.1109/90.909028.
444 20. Lin, X.; Shroff, N.B. Utility maximization for communication networks with multipath routing. IEEE
445 Trans. Autom. Control 2006, 51, 766-781. 10.1109/TAC.2006.875032.
446 21. Boyd, S.; Vandenberghe, L. Convex Optimization; Cambridge University Press: Cambridge, U.K., 2004;
447 0-521-83378-7.
448 22. Motwani, R.; Raghavan, P. Randomized Algorithms; Cambridge University Press: Cambridge, U.K., 1995;
449 978-052-14746-58
450 23. Zhang, F.; Cao, Y.; Wang, D. A Note on “Low-Complexity Distributed Scheduling Algorithms for
451 Wireless Networks”. IEEE/ACM Trans. Netw. 2015, 23, 1367-1369. 10.1109/TNET.2014.2323998.
452 24. Asmussen, S. Applied Probability and Queues; Springer-Verlag Press: New York, USA, 2003; 0-387-00211-1.
453 25. Georgiadis, L.; Neely, M.J., Tassiulas, L. Resource allocation and cross-layer control in wireless networks.
454 Foundations and Trends in Networking 2006, 1, 1-149. 10.1561/1300000001.

Das könnte Ihnen auch gefallen