This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication.

ROP: Near-Optimal Rendezvous for Dynamic

Spectrum Access Networks
ChunSheng Xin, Member, IEEE,, Min Song, Senior Member, IEEE, Liangping Ma, Member, IEEE,
Chien-Chung Shen, Member, IEEE

Abstract—This paper presents a rendezvous scheme, termed and a receiver. That is, when a transmitter node wants to send
Rendezvous with near-Optimal Performance (ROP), for dynamic traffic to a receiver node, how does the former find the channel
spectrum access (DSA) networks. ROP distributes nodes in pairs of the latter. To address this issue, most MAC protocols pro-
onto different rendezvous channels without using a control
channel, and achieves near-optimal throughput when the network posed for DSA networks, e.g., [5], assume a common control
is fully loaded, as well as the minimum expected time to rendezvous channel for nodes to negotiate data communication channels.
(ETTR) and channel load. For large networks, we further While the common control channel streamlines rendezvous
propose an efficient variant of ROP, termed fast ROP (FROP), for DSA networks, it is vulnerable to both traffic congestion
to significantly reduce the convergence time while still maintains and security attacks such as jamming. Hence there have been
near-optimal performance. Moreover, the simulation results have
illustrated that the convergence time of FROP is not only short, quite a few solutions proposed for rendezvous without using
but also independent of the network size, and hence FROP a control channel [2], [6]–[10]. The interested readers is
has good scalability. We have developed mathematical models referred to [11] for a detailed introduction on rendezvous
to analyze the throughput and the convergence time of ROP and and performance comparisons. Typically, these schemes utilize
FROP. Both the analytical and the simulation results validate the channel hopping (CH), i.e., every node independently hops
near-optimal performance of ROP and FROP.
over channels in a channel hopping sequence (CHS). A CHS
is a sequence of <channel, time slot number> tuples that
I. I NTRODUCTION indicate the channel a node needs to hop to in the correspond-
Dynamic spectrum access (DSA) has been widely studied ing time slot in a frame. Two nodes rendezvous whenever
due to its prospect to significantly enhance spectrum access they hop to the same channel in a time slot. Nevertheless,
efficiency [2]–[4]. With DSA, unlicensed or secondary users most of such schemes target rendezvous for two nodes, and
(SUs) dynamically detect and access the idle licensed spectrum may result in sub-optimal performance when considering ren-
bands (called as channels in this paper) that are not being used dezvous for multiple nodes altogether. For instance, many CH
by the licensed or primary users (PUs). The dynamic avail- schemes have optimal channel load for rendezvous between
ability of channels forces the SU nodes in DSA networks to two nodes, i.e., each channel has the same probability to be
dynamically change their operational channels. Furthermore, the rendezvous channel for the two nodes. However, when
these SU nodes typically operate on different channels due considering rendezvous for multiple nodes with these schemes,
to the spatial heterogeneity of channel availability and the the rendezvous node pairs are typically not equally distributed
requirement to minimize co-channel interference. Thus DSA on each channel, and hence the channel load can be poor.
poses a great challenge for rendezvous between a transmitter In this paper, we take a approach different from channel
hopping to achieve rendezvous without using a common
Copyright (c) 2013 IEEE. Personal use of this material is permitted. control channel, and target rendezvous for multiple nodes with
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending a request to [email protected]. near-optimal performance. We propose a rendezvous scheme
Manuscript received Oct. 11, 2012; revised Feb. 12, 2013; accepted Mar. 10, termed Rendezvous with near-Optimal Throughput (ROP).
2013. The associate editor coordinating the review of this paper and approving With ROP, SU nodes are distributed into available channels in
it for publication was Y. Cheng. A preliminary version of this work was
presented in IEEE Globecom 2011 [1]. pairs such that each rendezvous channel contains two nodes.
C. Xin is with the Department of Computer Science, Norfolk State ROP can achieve near-optimal performance, i.e., near-optimal
University, Norfolk, VA 23504, USA (email: [email protected]). The research throughput, minimum expected time to rendezvous (ETTR),
of C. Xin is supported in part by NSF under grants CNS-1017172, CNS-
1217668, and ECCS-1247853. and optimal channel load. Here, the channel load is the average
M. Song is with the Department of Electrical Engineering and Com- fraction of nodes falling onto a rendezvous channel, and ETTR
puter Science, University of Toledo, Toledo, OH 43606, USA (e-mail: is the average number of time slots between two rendezvous
[email protected]). The research of M. Song is supported in part by
NSF CAREER Award CNS-0644247 and NSF IPA Independent Research of two given nodes. For large networks, the convergence time
and Development (IR/D) Program. Any opinion, finding, and conclusions or of ROP (to be discussed in Section III-C) may be long. Hence,
recommendations expressed in this material; are those of the author and do we further propose a variant of ROP, termed fast ROP (FROP),
not necessarily reflect the views of the National Science Foundation.
L. Ma is with InterDigital Communications, San Diego, CA 92121, USA to reduce convergence time while still maintain near-optimal
(email: [email protected]). performance. In fact, the convergence time of FROP is not
C. Shen is with the Department of Computer and Information Sciences, Uni- only short, but also independent of the network size, and hence
versity of Delaware, Newark, DE 19716, USA (email: [email protected]).
The research of C. Shen is supported in part by NSF under grant CNS- FROP has good scalability.
1016841. The remainder of the paper is organized as follows. Sec-

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

tion II describes ROP. Section III proves ROP’s throughput Algorithm 1 Operations of node i in time slot t with ROP
optimality and develops an analytical model to compute the 1. Set the time slot t as seed, repeatedly generate node
convergence time of ROP. Section IV presents the design states [g1 , g2 , . .P
. , gN ] through
§ ¨a pseudo-random number
N
of FROP and analyzes its throughput. Section V presents generator until i=1 gi = N2 .
performance evaluation, and Section VI concludes the paper. 2. if gi = 1 then
3. Select the home channel as L(i, t, Ci ) and return.
4. end if
II. R ENDEZVOUS WITH NEAR -O PTIMAL T HROUGHPUT
5. Let P denote the set of passive nodes that node i has
We assume that each SU node is equipped with a cognitive packets for.
radio for communication, and a fast wideband spectrum sensor 6. repeat
for spectrum analysis. The nodes are assumed to operate in 7. success = true.
a time slotted mode. The cognitive radio of each node is 8. Randomly pick a node j in P, and estimate its home
assumed to be equipped with a GPS waveform, so that the channel as L(j, t, Ci ).
nodes are synchronized in time through receiving GPS signals. 9. Switch the radio to channel L(j, t, Ci ), and begin
A node uses its spectrum sensor to detect accessible channels neighbor discovery.
in each time slot. From the perspective of SUs, a licensed 10. if no other node on the channel then
channel alternates between the accessible (without PU activity) 11. success = false
and the inaccessible (with PU activity) states. The inaccessible 12. else if found other active node on the channel then
duration is the period of transmitting an entire PU data flow, 13. The active nodes coordinate to randomly pick
including the short periods between PU packet transmissions. one node to stay with the passive node.
Hence we assume that the accessible and the inaccessible 14. if Node i is not selected to stay then
durations of a channel are much longer than a time slot. 15. success = false
In this paper, we consider multi-user rendezvous in a single- 16. end if
hop network where all nodes can directly communicate with 17. end if
each other. Rendezvous in large-scale multi-hop networks will 18. until success = true or has selected all passive nodes
be studied in our future work. Let N denote the number of 19. Start packet transmission to peer nodes on the channel if
(SU) nodes in the network. Let M denote the total number of any.
both accessible and inaccessible channels, which is assumed
known to every node in advance. In this paper, we assume
that the accessible channels are already detected by SU nodes Algorithm 2 L(i, t, C), to compute the home channel for node
through spectrum sensing, and study the rendezvous between i in time slot t
nodes using these accessible channels. 1. Set i + t as the seed for the pseudo-random number
generator Z(•).
Due to the spatial heterogeneity of channel availability as
2. Let H = {1, . . . , M}
well as spectrum sensing imperfection, the sets of channels
3. repeat
accessible to different nodes are usually different. Let Ci
4. k = Z(|H|) // Generate k with 1 ≤ k ≤ |H|
denote the set of accessible channels detected by node i
5. h = H(k) // H(k) is kth channel in H
(1 ≤ i ≤ N ) in the current time slot. Algorithm 1 describes
6. H = H\{h} // Remove h from H
the operations of node i under ROP. As each node is assumed
7. until h ∈ C
to have only one radio, to ensure successful rendezvous, when
8. Return h // Selected home channel
a transmitter node tries to rendezvous with a receiver node, it
is desirable that the receiver radio is fixed on a certain channel.
Hence, in each slot, we classify nodes into passive or active
states, which is done in a decentralized manner. Specifically, in L(i, t, C) which selects node i’s home channel in time slot
time slot t, each node uses a pseudo-random number generator t from a channel set C.
to generate N states [g1 , g2 , . . . , gN ], where gi denotes the Next we discuss how a node j rendezvous with another node
state of node i. Note that each node obtains the same state i without exchanging control messages or any information,
vector [g1 , g2 , . . . , gN ]. For node i, if gi = 1, then node i is a before nodes j and i actually rendezvous on some channel.
passive node in time slot t and stays on a channel referred to as When node j has packets for node i, node j estimates the
its home channel to listen to incoming packets. If gi = 0, then home channel of node i by calling function L(i, t, Cj ), i.e.,
node i is an active node; it selects a passive node that node i node j uses its own accessible channels set Cj as a substitute
has packets destined to, and switches to the home channel of for Ci . It has been shown in [12] that the success probability
the selected passive node for packet transmission. To maximize of channel estimation is high.
throughput, we balance the number of passive nodes and We do not require coordination among active nodes to
the number of active nodes, by letting each node repeatedly choose passive nodes, since such coordination usually de-
generate [g1 , g2 , . . . , §gN ]¨ until the number of passive nodes, mands a common control channel and incurs coordination
P N N
i=1 gi , is equal to 2 . We use an algorithm adapted from overhead. Hence it is possible that some passive nodes are
[12] for a node to select its home channel, as described in selected by several active nodes while other passive nodes are
Algorithm 2. The algorithm can be viewed as a function not selected by any active node. However, this would result

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission
2 must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

in co-channel interference among these nodes. To reduce co- A. Throughput

channel interference and improve throughput, when multiple We assume that CAH p is reasonably large, and the number
active nodes select the same passive node, after they switch to of accessible channels M ≫ N . These two assumptions are
the home channel of the selected passive node and complete reasonable in most scenarios. First, the nodes that we consider
neighbor discovery, the active nodes coordinate to randomly for rendezvous are the nodes in a single-hop network that
pick one active node to stay with this passive node, e.g., let the can directly communicate with each other. The heterogeneity
active node with the largest ID pick one active node and notify of channel availability should not be such dramatic that a
all nodes on the channel. (Note that this requires exchange of node has only a small percentage of channels common to
control messages between these nodes on the same channel. a neighbor node. Therefore, p should be large. Second, the
Nevertheless, since they are already on the same channel, number of nodes N in a single-hop network is limited by the
they can easily exchange control messages.) All active nodes transmission range and hence is not too large. On the other
that are not picked then re-select passive nodes. Moreover, an hand, many studies have shown that there is a large number of
active node may incorrectly estimate the home channel of the spectrum bands that are idle at any time and any location [13].
selected passive node and thus switches to a channel that does Therefore, the number of accessible channels M is usually
not have the passive node. In this case, the active node also re- much more than the number of nodes N , i.e., M ≫ N .
selects a passive node. An active node continues to re-select a Next, we introduce a lemma from [12] regarding the number
passive node whenever it encounters either scenario. Note that of distinct home channels§ selected by passive nodes.
when an active node selects a passive node, it always selects Lemma 1: Let Ñ = N2 denote the number of passive
¨
one of the passive nodes that this active node has traffic for. nodes. The mean number of (distinct) home channels among
If all such passive nodes have been previously selected, the the Ñ passive nodes, denoted as Υ, is given as
active node stops passive node re-selection. µ ¶
At last, we briefly discuss the time complexity of Al- 1 M Ñ
£ ¡ ¢¤
Υ = M 1 − 1 − M 1 − (1 − p) . (1)
gorithm 1. On average, Step 1 takes 1/B(N/2 : N, 0.5)
iterations, where B(•) denotes the probability mass function
From Eq. (1), for large M , and Ñ ≪ M , we have
of a Binomial distribution with parameters N and 0.5. Since
generating random numbers is quick, the actual time spent on 1 Ñ Ñ
£ ¤
Υ ≈ M (1 − 1 − M ) ≈ M (1 − e− M ) ≈ Ñ , (2)
Step 1 is very small. The number of iterations for the loop
in step 6–18 is the convergence time of ROP, which is to be i.e., the number of distinct home channels is approximately
analyzed in Section III-C. equal to the number of passive nodes.
We assume that all nodes operating on the same channel use
an existing multiple-access MAC protocol, e.g., IEEE 802.11,
III. A NALYSIS OF ROP to access the channel. As a general practice in throughput
analysis, we are interested in the saturation throughput, which
As discussed earlier, in DSA networks, the set of chan- is defined as the limit reached by the system throughput as the
nels accessible to different nodes is usually different. We offered load increases, and represents the maximum load that
use the following model to characterize the heterogeneity of the system can carry in stable conditions. The reader is referred
the available channels at different nodes. Let M denote the to [14] for an insightful discussion of saturation throughput.
total number of accessible channels that are accessible by at An important implication of saturation throughput is that each
least one node in the current time slot, and these channels node is assumed to always have traffic to every other node, and
are denoted as {1, . . . , M }. We use a channel availability hence in ROP, an active node may select any passive node for
heterogeneity (CAH) parameter p (0 ≤ p ≤ 1) to model that packet transmission. Let T (n) denote the saturation throughput
an accessible channel k (1 ≤ k ≤ M ) is probabilistically on one channel when there are n nodes on the channel, which
available to each node with probability p. (Spectrum sensing can be computed from an analytical model for the MAC on a
imperfectness can be also accommodated by the CAH model. single channel, e.g., the model from [14].
For instance, if due to spectrum sensing imperfectness, an We model the passive node selection process in ROP
available channel can be detected only with probability r, as a discrete time
¥ ¦Markov chain (DTMC) with state space
then we can use a new CAH p′ = pr, to accommodate {n | 0 ≤ n ≤ N2 }, where state n denotes the number of
the spectrum sensing imperfectness.) We assume that p is a active nodes that need to re-select passive nodes. The initial
state of the DTMC is the total number of active nodes, N2 ,
¥ ¦
constant to make the analysis tractable. As to be shown in
Section V, our model also performs well when p varies for since all active nodes need to select passive nodes in the
different nodes and channels. With this CAH model, given beginning. A transition at state n is modeled as the concurrent
that a channel is available to either node i or node j, which passive node re-selection by all n active nodes. We call this
has a probability of 1 − (1 − p)2 , the conditional probability concurrent passive node re-selection at a state as a round of
that £ this channel¤ is commonly available to both nodes is ROP. Therefore, in a round with n active nodes, it is possible
p2 / 1 − (1 − p)2 = 2/ (2 − p) − 1. The smaller the p, the that k (0 ≤ k ≤ n) active nodes each selects a distinct passive
more heterogeneous the channel availability is. Therefore, node that has not been selected by any active node before, and
the CAH model effectively characterizes the heterogeneity of hence the DTMC may transit from state n to any lower state
channel availability between different nodes. n − k for 0 ≤ k ≤ n. On the other hand, the DTMC does

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission
3 must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

not transit from state n to state n + k for k > 0, since the N/2 n n-1 n-k 0
re-selection of passive nodes by the n active nodes at state n
will not result in more than n active nodes to re-select passive
Fig. 1. Transition from state n to other states in the DTMC of ROP
nodes in the next round. Fig. 1 illustrates the DTMC transitions
from state n to other states. N/2 - n occupied boxes n empty boxes
Theorem 1: ROP converges in a finite number ¥ ¦ of rounds,
and after it converges, its throughput S ≈ N2 T (2), which
¥ ¦ throughput with N nodes, i.e., the N nodes
is the optimal
k n-k
occupies N2 channels, with each channel having one active Among n balls reselecting boxes, j balls fall into k empty boxes
node and one passive node. In the case that N is an odd
number, one passive node would have no pairing active node Fig. 2. Boxes status at state n with a possible transition to state n − k
and is alone on its home channel.
Proof: From Fig. 1, clearly the DTMC has one absorbing
state, state 0. All the other states are transient states. This each pair occupy a separate channel in a time slot. Therefore
is a finite-state DTMC and hence it will settle down to the there are N2 number of rendezvous in a time slot. Assuming
absorbing state in a finite time. After the DTMC enters the that each node has traffic to all the other nodes, for N nodes
absorbing state, no active node needs¥ ¦ to re-select a passive to rendezvous with each other once, we need at least N (N2−1)
node. In other words, each of the N2 active nodes is paired number of rendezvous. Therefore, the minimum ETTR for a
pair of nodes is N (N2−1) N2 = N − 1.
±
with one (distinct) passive node and successfully switches to
the home channel of the selected passive ¥ node. For large p, From Theorem 1 and Eq. (2), after ROP terminates, the
N , M , and Ñ ≪ M , from Eq. (2), the N2 selected passive nodes are distributed into approximately Ñ channels, each
¦

nodes are all on different channels with a high ¥probability. In containing two nodes (assuming an even N ). Hence the
other words, with a high probability, there are N2 channels channel load is approximately 1 = N2 . As discussed above,
¦
Ñ
each having one passive¥ N ¦node and one active node. Thus the any rendezvous scheme with the optimal throughput has to
N
total throughput S ≈ 2 T (2). If N is an odd number, then
distribute N nodes into 2 channels, each with a pair of nodes,
and hence the channel load is N2 .
§N ¨
Ñ = 2 . Thus one passive node is not paired with any active
node, and is alone on its home channel.
In the ensuing discussions, without loss of generality, we C. Convergence Time
assume¥ N¦to be an even number so that the number of active We now analyze the convergence time of ROP, or the time to
nodes N2 = N2 = Ñ . The analysis for the case that N is an absorption of the DTMC, which will be used interchangeably
odd number is similar and omitted due to the space limitation. in the ensuing discussion. The convergence time is the number
of rounds for the nodes to re-select passive nodes within one
B. Expected Time to Rendezvous and Channel Load time slot. The passive node re-selection by active nodes at
Next we examine ETTR between two given nodes, which a given round of ROP is similar to placing balls into boxes
is the average number of time slots for the two nodes to meet as discussed below, with passive nodes as boxes and active
on some channel. From Theorem 1, after ROP terminates, nodes as balls. A box is occupied if a ball is successfully
every active node is paired with only one passive node. Since placed into it, i.e., an active node selects the passive node and
we consider the saturation throughput, each node always has correctly estimates the home channel of this passive node.
traffic to any other node, and hence an active node has the A ball may fail to be placed into a box (due to incorrect
same chance to be paired with any passive node. Thus an channel estimation) and is then placed out of all the boxes.
active node has approximately 1 chance to rendezvous with Fig. 2 illustrates the occupation status of boxes at state n and a
Ñ
a given passive node. In a given time slot, two given nodes can transition to state n−k, which happens when among the n balls
be in four joint states: PP, PA, AP, and AA, where ‘P’ denotes needing to re-select boxes, j balls fall into k (k ≤ j) boxes
passive and ‘A’ denotes active. When the two nodes are in that are previously empty. Let H(j, k) denote the number of
states PA and AP, they have probability 1 to rendezvous with ways to place j balls into k boxes such that all the k boxes
Ñ
each other. When they are in states PP or AA, the probability are occupied. H(j, k) is given as
of rendezvous is very small, and hence omitted in our analysis. Xk
(−1)v kv (k − v)j .
¡ ¢
Therefore, the probability of rendezvous between two nodes H(j, k) =
v=0
in a given time slot is lower-bounded by
¡ N −2 ¢ Let A(j, k) denote the event that among the n balls re-
1 2 N/2−1 1 1 selecting boxes, j balls fall into k (k ≤ j) boxes that are
Pr(PA or AP) = ¡ N ¢ × N = . previously empty. The probability of A(j, k) is
Ñ N/2 2
N −1
¡n¢¡n¢ n−j
The ETTR is then upper-bounded by N − 1. j k H(j, k)(Ñ − n)
Pr(A(j, k)) = ,
The ETTR upper-bound of ROP is the same as the minimum Ñ n
ETTR for any rendezvous scheme with the optimal throughput where in the numerator, the first term is the number of ways
N
2 T (2). To achieve the optimal throughput, a rendezvous to select j balls out of the n balls, the second term is the
scheme has to schedule the N nodes into N2 pairs and let number of ways to select k boxes from the n empty boxes,

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission
4 must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

N/2 – (n-k)-i occupied boxes n-k empty boxes

After obtaining P ′ , we remove the first row and the first
column (corresponding to the absorbing state 0), and denote
′
i boxes become empty due
the remaining matrix as P̄ . Let mn denote the mean time
to channel estimation error to absorption when the DTMC is at state n and m =
{m1 , . . . , mN/2 }. Let I denote the identity matrix and 1
Fig. 3. State transition from state n to state n − k
denote a column vector with N2 number of 1. Then m is
obtained as [15, Ch. 3]
the third term is the number of ways to place j balls into these ′
m = (I − P̄ )−1 1.
k boxes such that all the k boxes are occupied, and the last
term (Ñ − n)n−j is the number of ways for the remaining mN/2 is the time to absorption of the DTMC, i.e., the
n − j balls to select the Ñ − n occupied boxes. The transition convergence time of ROP.
probability from state n to state n − k is Next we discuss how to compute F (j, k, i). Let Q(j, k)
Xn denote the number of ways to place j balls into k boxes
P (n, n − k) = Pr(A(j, k)). (3) such that each box contains two or more balls. F (j, k, i) is
j=k obtained as follows. First, select i boxes from the k boxes and
However, due to the channel estimation error, the transition select i balls from the j balls; then place the i balls into the
from state n to state n−k may turn out to become a transition i boxes such that each box contains one ball; at last, place
from state n to state n − k + i, if among the k boxes, there the remaining (j − i) balls into the remaining (k − i) boxes
are i boxes where all the balls (active nodes) fail to correctly such that each box contains two or more balls, which has
estimate the channels and thus are placed out of any box, Q(j − i, k − i) number of ways. Therefore we have
 ¡k¢¡j ¢
as depicted in Fig. 3. In other words, some amount of the  i i i!Q(j − i, k − i), if
transition probability from state n to state n − k is ‘leaked’ to F (j,k,i) = max(0, 2k−j) ≤ i ≤ max(k−1, 2k−j) (5)
the transition probability from state n to state n − k + i. Let 
0, otherwise
α denote the success probability of channel estimation, which
is given as follows [12], By the definition of F (j, k, i) and H(j, k), we have
p ¡ Xk
1 − (1 − p)2M , F (j, k, i) = H(j, k).
¢
α= (4)
2−p i=0

where p is the CAH. Hence Q(j, k) can be computed as follows.

To make the analysis tractable, we make two assumptions. Xk
First, when a box has been selected by two or more balls, Q(j, k) = F (j, k, 0) = H(j, k) − F (j, k, i)
i=1
we neglect the probability that this box will become empty Xmax(k−1,2k−j)
= H(j, k) − F (j, k, i) (6)
due to channel estimation error, as this probability is less than i=max(1,2k−j)
(1−α)2 , which is small for reasonably large p. Hence, we can F (j, k, i) and Q(j, k) can be efficiently computed using
focus our analysis on the boxes selected by only one ball. The dynamic programming on Eqs. (5) and (6).
lower bound of the number of such boxes is max(0, k − (n −
k)) and the upper bound is max(k − 1, k − (n − k)). Second,
IV. FAST ROP (FROP)
when a ball fails to be placed into one of these k boxes due
to channel estimation error, we also neglect the probability Although ROP is guaranteed to converge in a finite number
that the ball will be placed by accident into any of the n − k of rounds by Theorem 1, the convergence time can be long
empty boxes (see Fig. 3), since the incorrectly estimated home for large networks, to be shown in Section V. This is because
channel can be assumed to be randomly any one of the M − k when the DTMC enters lower states, it becomes more and
channels, and hence this probability is M n−k more difficult for the few balls to find empty boxes, because
−k , which is also
small for Ñ ≪ M . there exist only a few empty boxes.
Let F (j, k, i) denote the number of ways to place j balls Can we design a rendezvous scheme that converges quickly
into k boxes such that all the k boxes are occupied and there while the throughput is still close to the one of ROP, i.e.,
are i boxes each containing one ball (the remaining k − i
boxes each contains two or more balls). Then, given that all Algorithm 3 Adjust transition probabilities
the k boxes are occupied when placing j balls into these k
boxes, the probability that there are i boxes each containing P′ = P
for n = 0 to N2 and k = 0 to n do
¥ ¦
one ball is F (j, k, i)/H(j, k). Given that there are i boxes
each containing one ball, the number of these boxes which for i = max(0, 2k − n) to max(k − 1, 2k − n) do
fail to receive the balls due to channel estimation error follows for h = 0 to i do
a Binomial distribution B(h : i, 1 − α) for 0 ≤ h ≤ i. P ′ (n, n−k+h) = P ′ (n, n−k+h)+P (n, n−
Based on the above discussion, after computing the transition k)B(h : i, 1 − s)F (j, k, i)/H(j, k)
probabilities matrix P = {P (n, n − k)} (0 ≤ n ≤ N2 , end for
0 ≤ k ≤ n) as in Eq. (3), we can obtain an adjusted transition end for
probabilities matrix P ′ by Algorithm 3. end for

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission
5 must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

absorbing states Algorithm 4 Operation of an active node in FROP

N/2 U +1 U 0 1. Compute the mean convergence time τ based on
N, M, p, γ.
2. while t < τ σ and the pairing passive node has more than
Fig. 4. State transition diagram of the FROP DTMC
one active node do
3. Re-select a passive node and switch to the estimated
with a near-optimal throughput? Fortunately, this is possible. home channel of the passive node.
Next we describe this scheme, termed fast ROP (FROP). Let γ 4. end while
denote the performance threshold that FROP has to maintain, 5. Start data communication with peer nodes in the same
which means the throughput of FROP, denoted as SFROP , has channel.
to be within the γ fraction of the throughput of ROP, denoted
as SROP , i.e., SFROP ≥ γSROP . Let U denote ⌊Ñ (1 − γ)⌋,
for the ease of description. We let states 0, 1, . . . , and U of the and passive nodes. We propose the following approach for an
underlying DTMC of ROP be absorbing states. In fact, states active node to stop re-selecting passive nodes in FROP. Each
0, 1, . . . , and U can be combined into a single absorbing state. node computes the mean convergence time τ of FROP, based
Fig. 4 illustrates the state transition diagram of the underlying on N, M, p and γ, which are known in advance by every node,
DTMC of FROP. After FROP converges, i.e., its underlying using a similar model in Section III-C with states 0, 1, . . . , and
DTMC enters the (new) absorbing state, at most U passive U as absorbing states. Let σ denote the duration needed for
nodes do not have pairing active nodes. Hence, at least Ñ − U one round of passive node re-selection and channel switching.
passive nodes have pairing active nodes. Since each passive An active node knows that on average, after duration τ σ from
node is very likely on a different channel, then there are the beginning of a time slot, the DTMC enters an absorbing
approximately Ñ − U utility channels. (We call a channel that state. Hence the node can stop re-selecting passive nodes at
includes at least one passive node and one active node a utility this time. Algorithm 4 illustrates how an active node operates
channel.) As a comparison, after ROP converges, the number in FROP.
of utility channels is approximately Ñ . On the other hand, the Next we derive an analytical model to compute the through-
throughput is proportional to the number of utility channels, put of FROP. Again, we assume that the number of nodes N
as the throughput on each utility channel is approximately is an even number, and hence the number of passive nodes
equal, independent of the number of nodes on on each channel Ñ = N2 . (If N is an odd number, the throughput is the same
(e.g., see Table I in [12]). Therefore, the throughput of a DSA as the one with N − 1 nodes, since there will be one passive
network is approximately proportional to the number of utility node without a pairing active node by Theorem 1.) We first
channels. Therefore we have introduce two lemmas from [12].
Lemma 2: Let β denote the probability that passive node
SFROP Ñ −U Ñ −⌊Ñ (1−γ)⌋ i selects channel k as its home channel using Algorithm 2.
≥ = ≈ γ.
SROP Ñ Ñ Then β does not depend on i or k, and is given as
Hence FROP does achieve the goal of SFROP ≥ γSROP . On
1
1 − (1 − p)M .
¡ ¢
the other hand, FROP also minimizes the convergence time. β= M (7)
This is because, to further reduce convergence time, we would
Lemma 3: Let φ(ñ) denote the probability that an active
have to let more states, e.g., state U +1, be an absorbing state.
node estimates the home channel of its selected passive node
However, it is then possible to have only Ñ − U − 1 passive
to be a given channel, given that there are ñ passive nodes on
nodes with pairing active nodes. Thus we would have
this channel. Then φ(ñ) is given as
Ñ −U −1 Ñ −⌊Ñ (1−γ)⌋−1
Ñ
= Ñ
< γ, h i
ñ ñ 1 M −1
φ(ñ) = Ñ
α + (1 − Ñ
)(1 − α) M −1 (1−(1−p) ) (8)
which does not meet the goal of SFROP ≥ γSROP . Hence,
FROP minimizes the convergence time subject to the perfor- The number of passive nodes selecting the same channel
mance threshold γ. as their home channels follows a Binomial distribution B(
After showing that FROP can significantly reduce conver- ñ : Ñ , β). For a given utility channel, the number of passive
gence time while achieving near-optimal throughput, there still nodes selecting it as their home channels follows the condi-
remains a challenge: how an active node knows the DTMC tional probability distribution B( ñ : Ñ , β)/(1 − (1 − β)Ñ )
enters the absorbing state and hence stops re-selecting passive with ñ ≥ 1.
node1 . From the perspective of an individual node, it does To obtain the exact throughput of FROP, we would need
not know the state of the DTMC, since it does not have the the state distribution of the DTMC after τ rounds of passive
global knowledge of the pairing relationship between active nodes re-selections, where τ is the mean convergence time of
FROP. To obtain τ , we will need to compute P τ , where P is
1 Note that ROP does not have this problem as an active node knows to the transition probabilities matrix of the DTMC. While it is
stop re-selecting passive node as long as there is no other active node pairing technically feasible, we would like to reduce the complexity,
with the passive node selected by this active node. In contrast, with FROP,
an active node may need to stop re-selecting passive node even if there are and aim to obtain an approximation throughput. We notice
other active nodes pairing with the passive node selected by this active node. that after τ rounds of passive node re-selections, the number of

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission
6 must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

active nodes that need to re-select passive nodes is expected to If M = M , ETCH achieves the same ETTR, channel load,
be U . Suppose we let FROP continue one more round. Given and throughput as ‘Optimal’ by letting M = N2 (assuming an
that there are ñ passive nodes on a utility channel k, among even N ). Comparing our schemes with the CH schemes, the
the U active nodes, the number of nodes switching to channel latter are suitable for scenarios where M is small, while our
k follows a Binomial distribution B(n : U, φ(ñ)). Therefore, schemes are suitable for large M , i.e., when there are plenty
the throughput for utility channel k is of PU channels available for DSA. The latter is usually true
as many studies have shown that there are a large number of
Ñ U
−U X spectrum bands that are unused in both temporal and spatial
B(ñ:Ñ ,β) domains [13].
X
Sk = B(n : U, φ(ñ))T (2ñ + n),
(1−(1−β)Ñ ) Next, we evaluate the performance of ROP and FROP
ñ=1 n=0

since it contains ñ passive nodes and ñ + n active nodes. experimentally. The nodes operating on the same channel are
Among the ñ + n active nodes, ñ active nodes are paired with assumed to use the IEEE 802.11b MAC protocol to access
the ñ passive nodes, and the other n active nodes are from the channel. The analytical model of [14] is used to calculate
the U active nodes that need to re-select passive nodes and the saturation throughput T (n) on a single channel when
have switched to channel k either intentionally or accidentally. there are n nodes operating on the channel. We assume a
Since each passive node is very likely on a different channel, channel data rate of 5.5 Mbps and the base access scheme. The
and the number of passive nodes without pairing active nodes packet payload is assumed to be 2304 bytes. The CWmin and
is approximately U , then there are about Ñ − U utility CWmax are assumed to be 32 and 1024, respectively. Other
channels. As such the total throughput is S ≈ (Ñ − U )Sk . IEEE 802.11b parameters use standardized values. Based on
these parameters, T (n) is computed as T (2) = 4.45, T (3) =
4.39, T (4) = 4.32, T (5) = 4.25, etc. The total number of
V. P ERFORMANCE E VALUATION channels M is assumed to be 400 and the total number of
In this section, we first give a theoretical comparison accessible channels M is assumed to be 200. The CAH p is
between ROP/FROP and other schemes, and then present assumed 0.9. The results with p set to other value between 0.8
an experimental evaluation of ROP/FROP. Table I compares and 1 have similar trends and observations, and are omitted
the performance of ROP and FROP with the optimal ren- due to space limitation. In practice, p can be different for
dezvous scheme and the state-of-the-art channel hopping (CH) different nodes, and hence we have also studied ROP and
schemes, where M denotes the total number of candidate FROP through simulations with p being a uniform random
PU channels, M ≤ M denotes the number of available variable in the range [0.85, 0.95], for each channel and each
channels for secondary users, P denotes the smallest prime node. The results are plotted on the figures as ‘sim-vp’.
number ≥ M , and N denotes the number of SU nodes. For FROP, the performance threshold γ is assumed 0.9. For
The performance for the CH schemes are from [2], [6]. The simulation, the simulation time is 5000 seconds with time slot
‘Optimal’ scheme is the rendezvous scheme that has optimal of 20 milliseconds. The simulation results are averaged over
throughput for N nodes, and the minimum ETTR and channel all time slots.
load given that the throughput is optimal. The performance Fig. 5 plots the throughput of ROP and FROP in simulation,
of ROP/FROP has an assumption that M is large compared where ‘optimal’ indicates the optimal throughput ⌊ N2 ⌋T (2) =
with N . There are a few observations. First, ETTR of the 4.45⌊ N2 ⌋, and ‘ana’ or ‘sim’ denotes the analytical or sim-
CH schemes is a function of the number of channels, M, M , ulation results, respectively. The optimal throughput is the
or P . Second, among the CH schemes, ETCH has the best maximum throughput for any rendezvous scheme with the
performance with regard to both ETTR and channel load. We control overhead excluded. We observe that the throughput of
note that in [2], the ETTR of ETCH (and most other CH ROP is very close to the optimal throughput. As expected,
schemes) is computed from the beginning of a frame to the the throughput of FROP is approximately within the γ =
rendezvous time slot in the frame, which is different from
our definition above. With our definition, the ETTR of ETCH
computed in [2] should be multiplied by 2 to become 2M − 1. 180
Optimal
160 ROP-sim
Throughput (Mbps)

140 ROP-sim-vp
TABLE I FROP-ana
P ERFORMANCE COMPARISON 120 FROP-sim
100 FROP-sim-vp
Scheme ETTR channel load optimal throughput 90%ROP-thp
80
JS [6] ∼ 5P
3
+ 11
3
N/A no
GOS [8] Θ(M 2 ) 1
no 60
M
2P 2 40
MC [8] P −1
N/A no
2M−1 √ 1 20
L-QCH [7] 2 2M−1
no
2M−1 1
ETCH [2] 2 M
conditional 10 20 30 40 50 60
2
ROP ≤N −1 ∼ N approximate Number of Nodes (N)
2
FROP ≤N −1 ∼ γN γ-approximate
2
Optimal N −1 N
yes Fig. 5. Throughput

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission
7 must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

80 60

Convergence Time (slots)

ROP-ana
70 ROP-sim 50
60 ROP-sim-vp

ETTR (slots)
FROP-ana 40
50 FROP-sim
40 FROP-sim-vp 30
30 Bound
20 FROP-sim
20 FROP-sim-vp
10 10 ROP-sim
ROP-sim-vp
0 0
10 20 30 40 50 60 10 20 30 40 50 60
Number of Nodes (N) Number of Nodes (N)

Fig. 6. Convergence time Fig. 7. Expected time to rendezvous (ETTR).

0.9 threshold of the ROP throughput, which is indicated as throughput, and minimum ETTR and channel load given that
‘90%ROP-thp’ in the figure. Although there is some overes- the throughput is optimal. FROP has a short convergence time,
timation, the analytical results match the simulation results while there is a slight performance depreciation compared with
well. The overestimation is primarily due to the fact that in ROP. Both the analysis and the simulation results validate
the analysis, we assume that the DTMC of FROP enters an the near-optimal performance of ROP/FROP. For our future
absorbing state, while in Algorithm 4, FROP stops after the directions, we will extend the proposed rendezvous schemes
mean convergence time, which does not guarantee that the for large-scale multi-hop networks.
DTMC enters an absorbing state. Nevertheless, as indicated
by the simulation results, the FROP throughput by simulation R EFERENCES
is close to the one by analysis that assumes convergence of
[1] C. Xin, M. Song, L. Ma, and C.-C. Shen, “An approximately optimal
the DTMC. Hence when it is in operation, FROP should either rendezvous scheme for dynamic spectrum access networks,” in Proc.
have entered or be close to an absorbing state after the mean IEEE GlobeCom, 2011.
convergence time. [2] Y. Zhang, Q. Li, G. Yu, and B. Wang, “ETCH: Efficient channel hopping
for communication rendezvous in dynamic spectrum access networks,”
The convergence time is plotted in Fig. 6. In Section III, in Proc. IEEE Infocom, 2011.
we have made an assumption that M ≫ N for the analysis [3] S. Huang, X. Liu, and Z. Ding, “Distributed power control for cognitive
model. We see that even when N = 60, the analysis result user access based on primary link control feedback,” in Proc. IEEE
INFOCOM, 2010.
matches the simulation result well. Hence the analysis model [4] Y. Shi and T. Hou, “A distributed optimization algorithm for multi-hop
is valid even when M/N = 3.3. The convergence time of cognitive radio networks,” in Proc. IEEE Infocom, 2008, pp. 1292–1300.
ROP increases linearly as a function of the number of nodes [5] J. Jia, Q. Zhang, and X. Shen, “HC-MAC: A hardware-constrained
cognitive MAC for efficient spectrum management,” IEEE J. Sel. Areas
N , and becomes large for large networks. This is the cost to Commun., vol. 26, no. 1, pp. 106–117, Jan. 2008.
achieve the near-optimal throughput. On the other hand, the [6] Z. Lin, H. Liu, X. Chu, and Y.-W. Leung, “Jump-stay based channel-
convergence time of FROP is short and approximately flat, hopping algorithm with guaranteed rendezvous for cognitive radio
networks,” in Proc. IEEE Infocom, 2011, pp. 2444–2452.
i.e., independent of the number of nodes. This is particularly [7] K. Bian, J.-M. Park, and R. Chen, “A quorum-based framework for
interesting because it means that no matter what the network establishing control channels in dynamic spectrum access networks,” in
size is, each node can simply run a certain rounds of passive Proc. ACM MobiCom, 2009, pp. 25–36.
[8] N. Theis, R. Thomas, and L. DaSilva, “Rendezvous for cognitive radios,”
node re-selection, and the DTMC would then likely enter an IEEE Trans. Mobile Comput., vol. 10, no. 2, pp. 216–227, Feb. 2011.
absorbing state, to achieve a near-optimal throughput, e.g., [9] G. Chang, W. Teng, H. Chen, and J. Sheu, “Novel channel-hopping
90% of the optimal throughput. In fact, in our simulations, schemes for cognitive radio networks,” IEEE Trans. Mobile Comput.,
2012, to appear.
after 10 rounds, additional passive node re-selections offer [10] K. Bian and J. Park, “Maximizing rendezvous diversity in rendezvous
little help to further increase throughput. We can see that the protocols for decentralized cognitive radio networks,” IEEE Trans.
throughput quickly increases in the first few rounds, and after Mobile Comput., 2012, to appear.
[11] H. Liu, Z. Lin, X. Chu, and Y.-W. Leung, “Taxonomy and challenges
about 10 rounds, the throughput increases only marginally. of rendezvous algorithms in cognitive radio networks,” in Proc. Inter-
At last, we examine the time to rendezvous. Fig. 7 illustrates national Conference on Computing, Networking and Communications
the ETTR between a pair of nodes with ROP and FROP, (ICNC), 2012.
[12] C. Xin, M. Song, L. Ma, and C.-C. Shen, “Performance analysis of a
together with the upper-bound N − 1. The ETTR of ROP control-free dynamic spectrum access scheme,” IEEE Trans. Wireless
is slightly shorter than the upper-bound, when N is large. The Commun., vol. 10, no. 12, pp. 4316–4323, Dec. 2011.
ETTR of FROP is slightly longer than the one of ROP because [13] M. A. McHenry, P. A. Tenhula, D. McCloskey, D. A. Roberson, and
C. S. Hood, “Chicago spectrum occupancy measurements & analysis
some passive nodes may not have pairing active nodes, and and a long-term studies proposal,” in Proc. 1st International Workshop
hence they cannot rendezvous with other nodes in a time slot. on Technology and Policy for Accessing Spectrum (TAPAS), 2006.
[14] G. Bianchi, “Performance analysis of the IEEE 802.11 distributed
coordination function,” IEEE J. Sel. Areas Commun., vol. 18, no. 3,
VI. C ONCLUSION AND F UTURE D IRECTIONS pp. 535–547, Mar. 2000.
This paper presented two rendezvous schemes, ROP and [15] O. C. Ibe, Markov Processes for Stochastic Modeling. Academic Press,
2008.
FROP, for DSA networks. ROP/FROP achieves near-optimal

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission
8 must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

Chunsheng Xin is currently an Associate Professor

in Norfolk State University. He received his Ph.D. in
Computer Science and Engineering from the State
University of New York at Buffalo in 2002. From
PLACE 2000 to 2002, he was a Research Co-Op in Nokia
PHOTO Research Center, Boston. In Fall 2002, he joined the
HERE Computer Science Department at Norfolk State Uni-
versity, and became early tenured in Fall 2005. His
research interests include cognitive radio wireless
networks, dynamic spectrum access, performance
evaluation and modeling. His research is supported
by several NSF grants and published in leading technical journals and
conferences. He has contributed several book chapters on professional books.
He is a member of IEEE.

Min Song is a Professor in the Electrical Engi-

neering and Computer Science Department at the
University of Toledo. Professor Song is currently
serving the National Science Foundation as a Pro-
PLACE gram Director of CNS/CISE. He received his Ph.D.
PHOTO in Computer Science from the University of Toledo
HERE in 2001. Over the years, he has secured more than
$2 million research funding from NSF, DOE, and
NASA, and published more than 120 technical pa-
pers. Dr. Song is the recipient of NSF CAREER
Award. Dr. Songs professional career is comprised
of a total 23 years of work experience in academia, government, and industry.
He was the Founding Director of a Networking System Division in an IT
company, and launched an international journal and served as the Editor-in-
Chief. Dr. Song has acted as an Editor or Guest Editor of 13 international
journals, and served as a General chair, Technical Program Committee chair,
and Session chair for numerous international conferences. Dr. Song is an IEEE
Senior member.

Liangping Ma received his B.S. degree in physics

from Wuhan University, Hubei, China in 1998, and
his Ph.D. degree in Electrical Engineering from
the University of Delaware, DE, in 2004. He is
PLACE with InterDigital Communications, San Diego, CA.
PHOTO Prior to 2009, he was with San Diego Research
HERE Center, Inc. and Argon ST, Inc. His research interests
include algorithm and protocol design for cognitive
radios at both the physical layer and above, and radio
resource management and QoS for cellular networks
and in particular for LTE/LTE-A.

Chien-Chung Shen received his B.S. and M.S. de-

grees from National Chiao Tung University, Taiwan,
and his Ph.D. degree from UCLA, all in computer
science. He was a senior research scientist at Bell-
PLACE core (now Telcordia) Applied Research working on
PHOTO control and management of broadband networks. He
HERE is now an associate professor in the Department
of Computer and Information Sciences of the Uni-
versity of Delaware. His research interests include
ad hoc and sensor networks, dynamic spectrum
management, control and management of broadband
networks, distributed object and peer-to-peer computing, and simulation. He
is a recipient of NSF CAREER Award, and his research has been sponsored
by NSF, NASA, Army Research Lab, RAND, and industrial companies. He
is a member of both ACM and IEEE.

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission
9 must be obtained from the IEEE by emailing [email protected].