Dynamic Open Spectrum Sharing MAC Protocol for ...

1

Dynamic Open Spectrum Sharing MAC Protocol for Wireless Ad Hoc Networks Liangping Ma, Xiaofeng Han, Chien-Chung Shen Abstract— The static spectrum allocation scheme used by the legacy wireless communication systems results in spectrum under-utilization. To make the best of the precious spectrum resource, any chunk of idle spectrum should be allowed to be used as a communication channel, subject to certain physical constraints, and this is the essence of the open spectrum paradigm. Allowing an arbitrary communication channel requires coordination between the sender and the receiver(s) such that the receiver(s) can readily receive the transmitted signal. This paper proposes the Dynamic Open Spectrum Sharing (DOSS) MAC protocol, a distributed protocol that allows for dynamic control channels and arbitrary data channels. The control channels are robust to jamming and are adaptive to traffic load, while the arbitrary data channels maximize the benefit of the available spectrum. In addition, this protocol supports efficient multicast, needs no synchronization, and provides an option that eliminates the hidden and exposed terminal problems. We conduct theoretical analysis of the protocol, study its performance via simulations, and discuss related implementation issues. Index Terms— Medium Access Control (MAC), ad hoc networks, open spectrum, dynamic spectrum allocation, multicast, jamming, load adaptive, radio receiver.

I. I NTRODUCTION The static spectrum allocation scheme used in legacy wireless communication systems could be quite wasteful, as shown in a recent study made by the Federal Communications Commission (FCC) that even in major urban areas, the allocated spectrum may be under utilized [1]. On the other hand, there is little spectrum left for allocation to meet the increasing spectrum demand of the emerging wireless applications. One approach to solving this paradox is opening up most of the spectrum for unlicensed spectrum users in ways that co-exist with legacy users. The potential of this approach has been demonstrated by the success of the IEEE 802.11 standard, which opens up from about 75 MHz to about 300 MHz (depending on specific sub-standards and countries) of the spectrum for unlicensed users to share with legacy users. However, what Liangping Ma is with San Diego Research Center, Inc., San Diego, CA 92121 USA (e-mail: [email protected]), Xiaofeng Han and Chien-Chung Shen are with the Department of Computer and Information Sciences, University of Delaware, Newark, DE 19716, (email:{han,cshen}@eecis.udel.edu).

the 802.11 standard does is not enough. First, the spectrum allocation is still based on the traditional notion of fixed channels, which is not flexible enough to achieve maximum spectrum utilization. For example, in 802.11, even if interference is only partially present over a channel, the entire channel is considered unusable, and is thus wasted. Second, the spectrum used for opportunistic sharing is insignificant compared to the entire spectrum that is suitable for wireless communications. With more spectrum opened up for opportunistic sharing, we can take full advantage of the more technically attractive wideband spread spectrum technologies, such as Ultra Wide Band (UWB) and CDMA [2]. In addition, the legacy spectrum applicationand-approval process is time consuming, and thus an automated and dynamic approach that offers real-time spectrum allocation and access is much desired. In this paper we consider MAC protocols for wireless ad hoc networks in the context of opportunistic open spectrum sharing. Such protocols are important since the network performance of wireless ad hoc networks is severely constrained by limited by the amount of the used spectrum. Gupta and Kumar [3] have shown that the capacity of each node in a wireless network is on the order of √ W/ n bits/sec under optimal circumstances, where W is the capacity of the wireless medium in bits/sec and n is the number of nodes. By Shannon’s channel capacity formula [4], W is proportional to the bandwidth of the channel B, i.e., W = B log2 (1 + SNR), where SNR is the signal to noise ratio. Thus, with more open spectrum available to opportunistic sharing, wireless ad hoc networks will perform better and become a more viable technology. Multi-channel Medium Access Control (MAC) protocols are proposed in the literature to improve the performance of wireless ad hoc networks through utilizing more spectrum. The most straightforward approach might be using a dedicated control channel together with multiple data channels. Example protocols of this approach include the Dynamic Channel Assignment (DCA) algorithm [5] where the access to the data channels is coordinated through control message (RTS/CTS) exchanging over the dedicated control channel. Another example protocol is the Common Spectrum Coordination Channel (CSCC) protocol [6], an extension of the DCA protocol that allows different types

2

of wireless devices to access the shared radio spectrum without conflict. Another approach is to divide the time into control periods during which all the nodes tune to the control channel for control message exchanging, and data periods during which data transfer takes place. The Multichannel Access Protocol (MAP) [7] is an example of this approach. A third approach is channel hopping, in which time is slotted and nodes hop through all the channels following a common schedule or their respective pseudo random schedules. If a node wants to communicate with another node, it follows the other one’s schedule. If they successfully exchange control information, they will stay on that channel to complete the data transfer. Example protocols in this category include the Channel Hopping Multiple Access (CHMA) [8] and the Slotted Seeded Channel Hopping (SSCH) algorithm [9]. These algorithms, however, do not take full advantage of the open spectrum paradigm. They break up a certain spectrum band into a number of fixed channels, which may result in low spectrum utilization because of the notion of unbreakable channel quantum. To see this, first note that only one of the data channels can be used by a node at any time (assuming that there is only one interface available) even though there are several contiguous data channels available. Second, a channel is considered busy even if a small fraction of it is being occupied (by legacy spectrum users or hostile interferences). Therefore, for efficient spectrum utilization, nodes shall be flexible in selecting the spectrum so as to take full advantage of all spectrum opportunities. To be specific, the nodes shall be able to (1) use multiple channels simultaneously, (2) use part of a channel, and (3) combine the available parts of adjacent channels into a single wide channel. With these functions, the center frequency and the bandwidth of the RF signal may be different from packet to packet. In fact, they may take arbitrary values within a certain range. Another problem with these multi-channel MAC protocols is that they assume that at some point of time there exists a common channel over which each of the involved communication nodes can both transmit and receive. This assumption, however, is not necessarily valid due to the the heterogeneity of the spectrum availability within a wireless network. To see this, consider the two nodes A and B in Fig.1. The shade indicates that a piece of spectrum is jammed. In this case, none of these multi-channel MAC protocols can enable communications between these two nodes. At first glance, it seems there is nothing we can do, but a careful thought reveals that such communications is possible: node A transmits over channel 3, over which node B is listening, and node B transmits over channel 4, over which node A is listening. We call this communica-

tion mechanism ”cross channel communication”. Node A

3

1

0

2

4

3

1

0

2

4

Node B

Fig. 1. A common channel is not available for nodes A and B.

In addition, these multi-channel MAC protocols also have individual protocol specific shortcomings. For dedicated control channel protocols such as DCA and CSCC, they cannot solve the potential control channel saturation problem. For time divisioned protocols such as MAP, CHMA and SSCH, network wide time synchronization is critically important, which, unfortunately, is not an easy task in a distributed wireless network [10]. To overcome the shortcomings of the above-mentioned MAC protocols, we propose the Dynamic Open Spectrum Sharing (DOSS) MAC protocol for wireless ad hoc networks. As a distributed protocol, DOSS allows for dynamic control channels and arbitrary data channels. The control channels are robust to jamming and are adaptive to traffic load, while the arbitrary data channels maximize the benefit of the available spectrum. In addition, this protocol supports efficient multicast, requires no synchronization, and provides an busy tone [11] extension based option which eliminates the hidden and exposed terminal problems. The remainder of the paper is organized as follows. Section II presents the DOSS protocol, Section III conducts theoretical analysis, Section IV provides simulation results and discuss some implementation issues. Section V concludes the paper, and the Appendix gives a more detailed discussion on the radio receiver design. II. T HE DOSS P ROTOCOL In this section we first give the details of the Dynamic Open Spectrum Sharing (DOSS) protocol, and then discuss its pros and cons. In the presentation of the data channel algorithm, we use an extensition of the single channel busy tone [11] approach as an example implementation, but note that it can be implemented using other medium access control mechanisms as well, such as CSMA/CA, for a tradeoff between cost and performance. Likewise, the implementation of the control channel algorithm is not unique either. For the DOSS protocol to work properly, the DOSS nodes should be able to detect the presence of the primary

3

spectrum users. Subject to spectrum availability and physical constraints such as the bandwidth of the transceiver antenna, DOSS nodes should confine their activities within certain reasonable operational frequency bands. We now elaborate on these matters in the below. A. Set-up of Operational Frequency Bands DOSS nodes are only secondary spectrum users, and they can use the spectrum only when the primary licensees are not using it. A critical enabling technology of this is a radio that is capable of processing potentially multigigahertz-wide bandwidth over the open spectrum and reliably detecting the presence of primary users in a very short period of time, which is challenging to the state of the art wireless technologies. It is shown, however, that wideband spectrum sensing like this is feasible when using a digital signal processing technique called cyclostationary feature detection [12]. With the wideband spectrum sensing capability, a wireless ad hoc network running the DOSS protocol continuously monitors the open spectrum over the geographic area of deployment for some time, mark the spectrum being used by the primary spectrum licensees, and decide on the spectrum that is can be used for opportunistic sharing. For reasons to be clear shortly, we divide the entire operational spectrum into two bands: a control band1 , which hosts dynamic control channels that are adaptive to exterior interference and interior traffic load, and a data band, which accommodates virtually any arbitrary dynamic data channel. Note that a wireless ad hoc network is distributed in nature. Therefore, it is undesirable to have a centralized controller for dynamic spectrum access in a wireless ad hoc network. The DOSS protocol is distributed, and it has two key algorithms: a control channel algorithm and a data channel algorithm. B. The Control Channel Algorithm The control channel algorithm distributes the nodes of a wireless network into multiple control channels and tries to keep the number of control channels as low as possible. Each control channel is dynamic in both center frequency and bandwidth, not only adapting to the interference (e.g., jamming, primary users) from the outside of the network but also the traffic load from the inside. There are two reasons for having a control channel algorithm like this. First, to make the best of the precious spectrum resource, any chunk of spectrum should be allowed to be 1 We distinguish frequency band and channel here. A band is a range of spectrum where a channel can sit, and frequencies within a band share similar RF properties.

used as a communication channel. As such, a node can select the chunk of spectrum from a broader range and thus maximizes its communication effectiveness, which is evidenced by the attained throughput, the delay, or the bit error rate. However, noticing that a radio receiver is in general much more complicated than a radio transmitter, an arbitrary communication channel puts a lot of signal processing burden on a radio receiver, which is discussed in detail in the Appendix. To support arbitrary communication channels while avoiding such signal processing burden, care must be taken to coordinate the sender and the receiver so that the receiver knows over which chunk of spectrum, or data channel, to listen for the incoming packet. A separate signaling channel, or a control channel, is a plausible solution to enable such coordination. Second, by limiting the number of control channels in the network, the control channel algorithm enables efficient link layer broadcast which efficiently implements network layer multicast[13]. The wireless medium is a broadcast medium, and a transmission can be potentially heard by all nodes within the transmission range. However, if there are too many control channels within the radio transmission range, a node has to go to each control channel to either directly broadcast or advertise the data channel to be used for the incipient broadcast. As discussed earlier, ”cross channel communication” is important in the open spectrum environment. This is particularly true when a network is segmented by exterior interference. To be specific, consider the example in Fig.2. Originally, all the nodes use channel 0 as the control channel. Then, a jammer is up and starts jamming channel 0. Part of the network, subnet 1, is disrupted, and the rest of the network, subnet 2, which is far enough from the jammer, suffers only negligible interference. Subnet 1 moves to channel 4 and sets it as the new control channel. If there is no other jammers, nodes A and B can set channel 4 as the shared control channel. However, another jammer is up and it forces subnet 3 to change its control channel to channel 1. As a result, there is no common control channel on which all the three subnets are mutually visible. Subnet 2 keeps channel 0 as its control channel and leverages ”cross channel communication” to talk with the other two subnets, as indicated by the arrows in the figure. The details of the control channel algorithm are shown in Fig.3. When a node is powered up, it is set to a preprogrammed initial control channel C0 , which has a center frequency f0 and bandwidth b, which is set as the control channel granularity, meaning that the entire spectrum consists of contiguous non-overlapping basic control channels, each having the same bandwidth as the initial control channel. Thus, if a node moves to a new basic con-

4

Subnet 1 Subnet 2 Subnet 3

A

B

Subnet 1

3

1

0

2

4 (ctr)

Subnet 2

3

1

0 (ctr)

2

4

Subnet 3

3

1 (ctr)

0

2

4

Fig. 2. Boundary nodes use ”cross channel communication” to talk to one another.

trol channel, the new channel will have center frequency f0 + kb and bandwidth b, where k is a positive integer. If a node expands its current control channel in response to high traffic load, by stipulating the bandwidth to be a multiple of 3b, the new center frequency will be still in the form of f0 + kb. This simplifies the radio receiver design, as a radio working only on frequencies f0 + kb will suffice. Each node estimates the interference level and the traffic load using the exponential weighted moving average (EWMA) with parameters α and β which are between 0 and 1, as shown in Lines 4–5. Line 6–12 implement channel adaption that is in response to exterior interference and traffic load. If the interference level is higher than the threshold th i, a node starts to search for a new control channel and goes to Lines 14–19. Note that after switching to a new control channel, a node will come back periodically to update the interference measure for channel C0 : I0 . If I0 drops well below the threshold th i, the node will switch its control channel back to the initial control channel, which is shown in Line 17. This elasticity toward the initial channel facilitates all the nodes to quickly converge to the initial control channel if the jammer disappears. Also note that after switching to a new channel, the node sends a hello message to the initial control channel. It is feasible to do this, because the jammer prevents this node from only receiving a packet over the initial control channel, so transmission over this channel is allowed. Specifically, the control channel algorithm has the following desirable properties: The nodes that experience the same exterior interference or the same traffic load will switch to a common control channel without explicit communications. • If the jammer disappears, a node with a current estimate I0 > th i will switch back to the initial control channel i within time (blog(1−α) th 2I0 c + 1)T0 . • Similarly, if the traffic load on the initial control channel becomes negligible, a node with a current estimate L0 > th l will switch back to the initial control channel in l (blog(1−β) th 2L0 c + 1)T0 . •

The last two statements are straightforward given the updates in Fig.3.

/* control channel algorithm */ 1 Initialization: each node n is set to the initial control channel C0 with center frequency f0 and bandwidth b. Schedule the first observation timeout T0 for C0. 2 each node n periodically updates interference level and traffic load estimates: I0, L0 3 if observation timeout T0 occurs 4 I0 = (1-Į) I0 + Įi0 /* i is the new interference level observation */ 5 L0= (1-ȕ) L0 + ȕl0 /* l is the new traffic load observation */ AdaptChannel (I0, L0, T0): 6 if I0 > th_i 7 search for non-jammed basic control channel nearest to C0 8 elseif L0 > th_l 9 search for control channel with 3x current bandwidth nearest to C0 10 else 11 schedule next observation timeout T0 12 end /* also end of AdaptChannel( ) */ 13 end 14 search for non-jammed basic control channel nearest to C0: 15 if found Cn qualifies 16 switch channel, monitor In and Ln for Cn, and I0 and L0 for C0 , send hello message to C0 17 if I0 < 0.5 th_i, then switch back to C0 18 AdaptChannel (In, Ln, Tn) 19 end 20 search for control channel with 3x current bandwidth nearest to C0: 21 if found Cn qualifies 22 switch channel, monitor In and Ln for Cn, and I0 and L0 for C0 , send hello message to C0 23 if I0 < 0.5 th_l, then switch back to C0 24 AdaptChannel (In, Ln, Tn) 25 end

Fig. 3. The control channel algorithm in the DOSS protocol.

To maximize the success rate of the hello messages discussed above and other control messages, care must be taken to coordinate the transmission of these messages. The coordination methods could be any of the common ones, such as CSMA, CSMA/CA (using RTS/CTS and exponential backoff). We adopt a hybrid approach depending on whether the communication parties share the same control channel. If they share a control channel, we use an adapted non-persistent CSMA[14] algorithm, which introduces much less communication overhead than CSMA/CA since the control packets are generally small in size. To be specific, 1. If a node has a packet to send and it senses the channel idle, then it sends the packet. 2. If the channel is sensed busy, it randomly backs off and will re-sense the channel when the backoff expires. 3. If there is a collision (detected by not receiving the REQ ACK within a timeout), a node repeats the process in steps 1 and 2. On the other hand, if the communication parties do not share a control channel, CSMA does not work and we resort to explicit channel reservation to minimize potential conflicts. To illustrate, consider nodes A and B in Fig.2. Nodes A and B can communicate if node A sends packets over channel 0 to node B and node B sends packets over channel 4 to node A. However, when A sends a packet over channel 0 to node B, A’s neighbors over subnet 1 are not aware of this transmission since they cannot receive packets on channel 0 due to jamming. Therefore,

5

the CSMA approach used for the other case does not work here. To minimize potential collision with nodes over subnet 1, node A can send a transmission notification packet over channel 4, reserving channel 0 for a short period of time, during which node A transmits a hello message to node B. Likewise, node B sends a notification packet over channel 0 first and then transmits a hello packet over channel 4 to node A.

tone and send the busy tone in order to inform other neighbors not to send. g

fl

fu

The primary purpose of the data channel algorithm is to allow maximum flexibility in utilizing the radio spectrum. The flexibility can be achieved without compromising the network effectiveness by performing coordination over the control channels. The data channel algorithm can be based on the RTS/CTS mechanism. However, the RTS/CTS approach is not good enough in eliminating the hidden terminal and exposed terminal problems, which are a major cause of poor performance of wireless ad hoc networks. The use of busy tones has proven to be an effective approach to solving these problems for the single fixed channel case [11]. We extend this concept of busy tone by assigning a busy tone to each dynamic data channel. Although a node cannot monitor all the possible spectrum, we may set up the busy tones so that all the busy tones are within a short range and can be heard by any node if the channel condition permits. Doing so, we effectively decompose the data band into two bands: a busy tone band, and a diminished data band. In other words, we now have three bands: a data band, a control channel, and a busy tone band. In what follows we discuss how to build a mapping between a busy tone and a data channel. The data channel algorithm consists of the following components: (1) spectrum mapping, (2) spectrum negotiation, and (3) data transmission.

fl

We establish a one-to-one mapping between the narrow band (low bit rate) busy tones and the wide band (high bit rate) data channels. This way, by observing narrow band receiver-initiated busy tones, a node knows all the data receiving activities in its neighborhood. Denote the busy tone band as [fl , fu ] and the data band as [Fl , Fu ]. The mapping is illustrated in Fig. 4. A simple realization of the mapping is through a linear function g g(x) =

1 [(fu − fl )x + fl Fu − fu Fl ] , Fu − Fl

(1)

where x ∈ [Fl , Fu ]. With the spectrum mapping, a receiver only needs to convert the spectrum over which it is receiving to a busy

Fu frequency

(a)

C. The Data Channel Algorithm

C.1 Spectrum Mapping

fu Fl

Fl

Fu

(b) Fig. 4. Spectrum mapping in the Dynamic Open Spectrum Sharing (DOSS) protocol.

C.2 Spectrum Negotiation The next step is spectrum negotiation, by which the sender and the receiver agree on the dynamic channel for the incoming data transmission. The negotiation is done in five steps. 1. Each node monitors its own available spectrum, which for the sender is the spectrum not being used for receiving in its neighborhood, and for the receiver is the spectrum not being used by itself for receiving. 2. The sender sends a REQ packet (over the control channel) to the intended receiver. A REQ packet contains the channel parameters (center frequencies, bandwidths, etc.) for all available data channels of the sender. To avoid overfragmentation of the spectrum, those available data channels that are too narrow to be useful are excluded. By listening to busy tones and referring to the spectrum mapping, the sender has full knowledge of the spectra being used for data receiving within its neighborhood, thus being able to avoid disrupting ongoing data receiving activities. 3. The receiver compares the sender’s available channels with its own available channels, and picks up an intersection that is available to both. The receiver then replies with an acknowledgement (called REQ ACK), which contains the channel parameters of the negotiated common channel, over the control channel. Note: (1) The REQ ACK is necessary for avoiding ambiguity on the sender side. To know the negotiated common channel, the sender can not solely rely on the reception of the busy tone, since it is possible that another pair of nodes generate an identical busy tone, in which case the sender can not tell whether the received busy tone is destined for it. (2) If there are multiple dynamic channels

6

available, which one does the receiver choose? We argue that it should choose the largest one. This will not make the largest data channel overcrowded, since if the largest one is being used, its busy tone will be heard by other nodes, which will not compete for the largest one but the second largest one. 4. The receiver refers to the spectrum mapping to find and turn on the corresponding busy tone, telling its neighbors not to send over this data channel. 5. Upon receiving the REQ ACK, the senders knows the dynamic data channel over which the receiver is waiting for the data packet, and tunes its data transmitter to that channel for data transmission. Figure 5 shows an example in which node A is the sender and node B is the receiver. Node A has two available channels F1 , F2 . Node B has three available channels F3 , F4 , F5 . Node A sends its channel availability information through a REQ packet to node B. Node B realizes that channel F6 is common to both, finds f6 using the spectrum mapping, sends a REQ ACK packet, and turns on busy tone f6 .

C.4 State Description of the Data Channel Algorithm The data channel algorithm of the DOSS protocol can be formally described by a state diagram shown in Fig. 6, where the notations are: • IDLE: the state where a node is not currently involved in any MAC layer communications. • ready: a node receives data from its upper layer protocols. • REQ: a communication request and channel negotiation packet sent over the control channel by the sender. • WF REQ ACK: wait for the REQ ACK packet, which contains the channel parameters of the the negotiated data channel. • WF DATA: wait for data. • WF DATA ACK: wait for data acknowledgement. If the sender does not receive the DATA ACK for some time, a WF DATA ACK timeout (a duration greater than the maximum RTT), the sender regards the transmission failed and retransmits the data packet. WF_DATA_ACK timeout / retransmit data

Receive REQ_ACK / send data, start timer

F1

F2

WF_REQ_ACK

f

Receive DATA_ACK

F3

F4

F5 f

Negotiated Band f6

F6

f

Fig. 5. Spectrum negotiation in the DOSS protocol.

receive REQ, find channel / send REQ_ACK, BT on, start timer WF_DATA

WF_REQ_ACK timeout

Node B

ready, avai_spectr not empty / random wait, send REQ start timer

WF_DATA_ACK

Node A

IDLE receive data/send DATA_ACK, BT off

WF_DATA timer expires / BT off

C.3 Data Transfer The sender sends a data packet over the negotiated dynamic data channel to the receiver. If the packet is correctly received, the receiver replies with another acknowledgement packet (called DATA ACK) over the negotiated data channel and turns off the busy tone. Upon receiving the DATA ACK packet, the sender realizes the transmission is successful. If the sender does not receive the DATA ACK within a timeout (WF DATA ACK timeout), which is longer than the maximum RTT, the sender retransmits the data packet. Note: The sender can not rely on the turn-off of a busy tone, because it is possible that another pair of nodes turn off an identical busy tone. Thus a DATA ACK packet is necessary.

Fig. 6. State diagram for the data channel algorithm of the DOSS protocol.

D. The Pros and Cons of the DOSS Protocol The Pros of the DOSS protocol: (1) it makes the best of the precious spectrum resource by opportunistically utilizing the available spectrum, (2) supports effective multicast, (3) is robust to jamming and overcomes the control channel saturation problem, (4) is scalable, and (5) its busy tone extension eliminates the hidden and exposed problems. The first one is evident in the description of the DOSS protocol. The second one is due to the maximum homogeneity in the control channels resulted from its control channel algorithm. This allows efficient broadcast, on which multicast is based. The third one is a result

7

of the adaptivity of the control channel to both exterior interference (primary users’ activities excluded though) and traffic load. The fourth one results from the distributed nature of the DOSS protocol. Here we elaborate on the last one by comparing DOSS with the Distributed Coordination Function (DCF) scheme of the IEEE 802.11 standard. IEEE 802.11 uses RTS/CTS to relieve the hidden and exposed terminal problems. However, if the RTS or the CTS packet is not decodable (e.g., the received signal power is just below what is needed for decoding), the RTS/CTS mechanism simply does not work. Alternatively, if we take a different approach by using a busy tone as done in DOSS, only sensing (instead of decoding) is needed and that makes DOSS more robust to signal degradation. We first consider the hidden terminal problem by examining a scenario shown in Fig. 7(a), where node A is responding with a CTS packet after receiving an RTS packet from node B. If node C is located between node A’s transmission range (Rt ) and sensing range (Rs ) [15], node C can sense but can not decode the CTS packet, which includes the information about the duration during which node C should remain silent. As a result, node C may transmit and cause a collision at node A. This example shows that IEEE 802.11 cannot avoid this kind of hidden terminal problems, which, in contrast, can be eliminated in DOSS. To be specific, if node A turns on a busy tone and keeps it on while receiving, node C will be able to sense the busy tone and will not transmit until node A finishes receiving, thus avoiding the collision that would have happened in IEEE 802.11. Now consider the exposed terminal problem. IEEE 802.11 MAC can alleviate the exposed terminal problem by allowing a node to send if the node hears an RTS but not a CTS. However, if the RTS is not decodable, this mechanism will fail. This is illustrated in Fig. 7(b), where node C senses (but cannot decode) node A’s RTS and does not hear node B’s CTS. As a result, node C is not sure whether it senses an RTS packet or a CTS packet. Thus, node C has to refrain from transmitting to node D to avoid possible collision at node A. However, in DOSS, when node A is sending to node B, there will be no busy tone that node C can hear. Node C concludes that no one nearby is receiving and thus is free to transmit to node D. The cons of the DOSS protocol is the need for multiple radio transceivers. When a receiver receives a data packet, it must turn on the busy tone at the same time. Thus, we need at least two radio transceivers. In a simple design, we need three, one for each of the operational frequency bands. The need for multiple transceivers will increase the device cost. However, a modest increase in device cost is well worth the improvement of network performance.

Rt B

Rt A

C

Rs

B

A

C

D

Rs

(a)

(b)

Fig. 7. (a) Hidden terminal with IEEE 802.11: node C may cause a collision at A; and (b) exposed terminal problem with IEEE 802.11: node A may prevent C from transmitting to D.

III. T HEORETIC A NALYSIS In this section, we present analytic analysis on the DOSS algorithm to gain some insights into it. We first consider the channel capacity of a single wireless link that is achievable by DOSS and compare it with what is achievable by the corresponding static spectrum allocation protocol. Let the spectrum be B Hz wide, a γ fraction of which is taken by primary users and thus not available to DOSS nodes, where 0 ≤ γ ≤ 1. For the portions of spectrum where there are no primary users, let the signal to noise ratio (SNR) be α = Ps /Pn , where Ps is the signal power and Pn is the noise power. For analytic simplicity, we assume that the power spectrum density of the primary users equals that of the signal transmitted by a wireless node, and this gives the power of the primary users γPs . For DOSS nodes, only the portions of the spectrum that are free of primary users are chosen for the communications. These portions sum up to (1 − γ)B Hz. Thus its achievable channel capacity per unit time (in bits/sec) [4] is Cd = B(1 − γ) log2 (1 + α). (2) For the corresponding static spectrum allocation protocol, a node may take two schemes accessing the spectrum. In the first scheme, the node senses the presence of the primary users and marks the entire channel B as ”unusable” and does not access the spectrum. The achievable channel capacity then amounts to zero, i.e., Cs = 0. Since Cd ≥ 0, DOSS always outperforms this scheme. In the second scheme, the node accesses the entire spectrum B anyway, regardless of the presence of the interference from the primary users. This is the best scheme, and its achievable channel capacity is Ps ) Pn + γPs α = B log2 (1 + ). 1 + γα

Cs = B log2 (1 +

(3)

To compare the second scheme with DOSS, we consider

8

the capacity gain g = Cd − Cs ,

(4)

which is plotted in Fig. 8 as a function of α and γ. It is seen that the gain is positive except for the extreme cases where γ is close to 1 (too much spectrum is used by the primary users) and α is close to 0 (the SNR is too low). Furthermore, the higher the SNR α, the greater the gain. The reason for the gain is that if a node uses the entire spectrum B, the overall SNR over the entire spectrum B could be very low, leading to low achievable channel capacity. In contrast, DOSS only uses the portions of spectrum that are free of primary users, therefore resulting in a much higher SNR and hence higher capacity despite using less spectrum.

Capacity Gain g

4B 3B 2B 0

B 0 0.5 −B 0

50

α

100

150

γ

1

Notes: (1) To facilitate synchronization and to increase reliability, the PHY headers of all packets are sent at the fixed rate rc . (2) The propagation time τ takes into account the time spent detecting an arriving signal. We assume the attempted channel traffic over the control channel is Poisson with mean arrival rate λ packets/second. Note that this traffic takes into account both the actual transmissions, and the attempted transmissions (i.e., sensing) that do not result in actual transmissions. The attempted channel traffic over the control channel constitutes a renewal process, which includes the Poisson process as a special case. By the renewal theory[14], the PHY layer throughput is ¯ U (5) S = ¯ ¯, B+I ¯ is the average duration of a successful transwhere U ¯ is the average duration of any transmission mission, B (whether successful or not), I¯ is the average duration of ¯ + I¯ is the average duration of a any idle period, and B renewal interval. The probability that a single REQ packet is successful is e−λτ , and the probability that a REQ ACK packet is successful is also e−λτ . Since a request is successful only if both the REQ and the corresponding REQ ACK are successful. Thus, the probability that a spectrum negotiation ¯ = tr ps , we have is successful is ps = e−2λτ . Since U

Fig. 8. Gain in achievable channel capacity of DOSS over the best static spectrum allocation scheme.

Next we analyze DOSS over a fully connected wireless network. By fully connected, each node can hear the transmission of any other node in the network. This scenario allows for a tractable analysis and is yet enough to provide insights. We focus on the throughput and the interaction between the control channel and the data channel. We use the following notations: • rc : control channel raw data rate, • rd : data channel raw data rate, • tp : PHY header transmission time, • tr = lr /rc + tp : transmission time of the REQ packet, where lr is the length of the REQ packet (at the MAC layer), • td = ld /rd + tp : transmission time of the data packet (assuming all data packets have the same length), where ld is the length of the data packet, • τ : maximum one-way propagation delay, • t1 = l1 /rc + tp : transmission time of the first ACK, or REQ ACK, where l1 is the length of the REQ ACK packet, • t2 = l2 /rd + tp : transmission time of the second ACK, or DATA ACK, where l2 is the length of the DATA ACK.

¯ = tr e−2λτ . U

(6)

For analytic simplicity, we assume that tr = t1 . Let the time when a packet is sent be t. Then between time t and t + τ , any attempted transmission will result in an actual transmission and hence a collision. Let t + Y be the time when the last packet arriving between t and t + τ . Then, ¯ = tr + Y¯ + τ. B

(7)

It can be shown that[14] Y¯ = τ − λ1 (1 − e−τ λ ). The average time between attempted transmissions is 1/λ. But for any successful REQ – REQ ACK pair, the control channel will be idle for td + 2τ + t2 . This effectively extends the average idle time, and thus 1 I¯ = + ps (td + 2τ + t2 ) λ

(8)

Let Sr denote the throughput of the successful REQs whose corresponding REQ ACKs are also successful. Substituting (6), (7) and (8) into (5) and replacing S with Sr , we obtain Sr =

λtr e−2λτ λ(tr + 2τ + (td + 2τ + t2 )e−2λτ ) + e−λτ

(9)

Application layer throughput over data channel (Mbps)

9

Note that the throughput in (9) is normalized. The number of successful spectrum negotiations per second is Sr0

rc Sr , = lr + rc tp

1.8 1.6 1.4

(10)

1.2

where the denominator is the REQ packet length at the PHY layer. If we ignore such failures at the PHY layer as in synchronization and checksum, the data transmission will be successful as long as the spectrum negotiation is successful. Thus the data throughput (over the data channel) in bits per second is = =

0.8 0.6 0.4 0.2 0

0

2

4

10 10 10 Attempted transmission rate over control channel (pkts/sec)

Fig. 9. Application layer data throughput over the data channel vs. attempted transmissions over the control channel. The blue solid line is the theoretic analysis, and the green dashed line is rc ld0 λtr e−2λτ the simulation result, where the vertical hashes represent 95% lr + rc tp λ(tr + 2τ + (td + 2τ + t2 )e−2λτ ) + e−λτconfidence intervals.

Sr0 ld0

(11) where ld0 is the packet size of a data packet. If ld0 is measured at the MAC layer, i.e., ld0 = ld , then Sd0 is the MAC layer throughput. If it is measured at the application layer, then Sd0 is the application layer throughput. Now we validate (11) via QualNet simulations. The bandwidth of the control channel is 2 MHz, and the raw data rate rc = 1 Mbps. The data band is 10 MHz, and the the raw data rate is 5 Mbps. However, 1/3 of the data band is taken by primary users. Invoking the assumption that the data rate is proportional to the bandwidth, the maximum raw data rate over the data channel is rd = 3.33 Mbps. The PHY header transmission time tp = 0.192 ms. The MAC payload of the REQ, REQ ACK and DATA ACK packets are 80 bytes, 80 bytes, 40 bytes, respectively. The maximum propagation delay is 2µs. In the simulation, 50 nodes send Poisson UDP traffic to a common receiver. Figure 9 shows that the theoretical analysis agrees with the simulation result very well. The minor discrepancy is probably because the attempted traffic process is not exactly Poission. Having validated (11), we now use it to study the interaction between the data channel and the control channel. Recall the definitions of tr , t1 , td , t2 . The data throughput Sd0 is not only dependent on the raw data rate over the data channel rd but also on the raw data rate over the control channel rc . Figure 10 shows that given the attempted transmission rate λ = 4000pkts/sec over the control channel, the raw data rate over the data channel rd becomes the bottleneck of the system for rc greater than 2 Mbps. IV. S IMULATION R ESULTS AND D ISCUSSIONS Operating only over a pre-assigned narrow band is not the primary motivation of DOSS. It is desirable for DOSS

Application layer throughput over the data channel (Mbps)

Sd0

1

2.5

2

1.5

1

0.5

0 0

2 4 6 8 Raw data rate over the control channel (Mbps)

10

Fig. 10. Application layer data throughput Sd0 over the data channel vs. raw data rate over the control channel rc .

to take full advantage of the idle spectrum, acquiring as much bandwidth as possible as long as it does not disrupt primary spectrum licensees. Thus, in this section we consider DOSS for a multi-hop wireless ad hoc network over the open spectrum. We consider a network of 10 nodes evenly located in a straight line such that a node can only communicate with its immediate neighbors. One of the end nodes send Poisson UDP traffic across the line to the other end. The open spectrum is 110 MHz wide, and the use of the open spectrum is shown in Fig. 11. In the simulation, rc = 5.5 Mbps, rd = 29.7 Mbps, tp = 0.192 ms, and the data packet size is 1000 bytes. The simulation results are shown in Fig. 12. One thing that is critical in the implementation of the DOSS protocol is the radio receiver design. The motivation of a simple radio receiver for DOSS is already discussed in Section II-A. This receiver receives the RF parameters of the dynamic data channel over its control channel. Coherent modulation and demodulation schemes are

10

110 MHz Control Channel 0.9 MHz Busy Tone Band

5 MHz

Data Band (90 MHz) 2.460 GHz

2.350 GHz

Fig. 11. DOSS finds 110 MHz open spectrum available.

A PPENDIX

2

End−to−end data throughput (Mbps)

1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 0

Acknowledgements: The authors would like to thank Eric Blossom of Blossom Research, Prof. Leonard J. Cimini of the University of Delaware, and Dr. Bo Ryu and Dr. Zhuochuan Huang of San Diego Research Center for their insightful suggestions and comments.

50 100 150 200 250 300 350 Application layer Poisson traffic arrival rate (pkts/sec)

400

Fig. 12. End-to-end data throughput changes as the transmission rate of the Poisson UDP traffic increases, where the vertical hashes represent 95% confidence intervals (very small).

used. The design is thus along the traditional radio design path except that (1) the receiver is automatically tunable and (2) channel switching is quick. The first requirement is justified for two reasons: first, the channels used for transmitting the data packets are not fixed and possibly different from packet to packet, and second, the control channels may migrate slowly over time. The second requirement provides better time utilization of the data channel. Since the data channel to be agreed upon may be different from packet to packet, there is a need for the sender and the receiver to switch the channel from time to time. During the channel switching, both the sender and the receiver are idle. Therefore the channel switching should be completed quickly. V. C ONCLUSIONS We propose an MAC protocol, DOSS, for wireless ad hoc networks operating over the open spectrum. As a distributed protocol, DOSS allows for dynamic control channels and arbitrary data channels. The control channels are robust to jamming and are adaptive to traffic load, while the arbitrary data channels maximize the benefit of the available spectrum. In addition, this protocol supports efficient multicast, needs no synchronization, and provides an option that eliminates the hidden and exposed terminal problems. We conduct a theoretic analysis of the DOSS protocol and validate it through simulations, and also study its performance using simulations.

In addition to the radio design outlined in Section IV, we have tried two other designs with the motivation of allowing a receiver to detect the RF parameters (carrier frequency and bandwidth) of the incoming signal on the fly so as to eliminate the need for a control channel. This motivation is attractive from the network perspective, but they are hard to realize at reasonable costs in practice, as explained below. In the first approach, using an Fast Fourier Transform (FFT) real-time spectrum analyzer based radio, we down convert an RF signal to an intermediate frequency (IF) without knowing the actual carrier frequency of the RF signal. Following IF amplification, the IF signal is digitized and then fed into an FFT real-time spectrum analyzer, where the RF parameters are estimated. Once the RF parameters are figured out, they are used to tune the local oscillator and the various filters such as the RF amplifier and the IF amplifier for coherent demodulation, which provides better performance than non-coherent demodulation. However, there are several difficulties in this approach. First, detecting the RF parameters may be quite unreliable in practice. Since the carrier frequency and the bandwidth of the desired signal are unknown, we have to monitor the entire spectrum. As a result, any other received signal contributes to the noise, severely degrading the signal to noise ratio. Second, Analog to Digital (AD) conversion of the IF signal poses potentially heavy signal processing burden on the radio receiver. To illustrate, let the width of the operating radio bands be B. The sampling frequency must be greater than 2B in order to avoid aliasing. Let the precision of each sample be k bits. Then the AD converter has to process 2Bk bits of data per second. If B = 50MHz and k = 8, the signal processing burden is 0.8G bits/sec, which is pretty tough for stand alone radios. If we consider a larger B and the computation power needed for the FFT spectrum analyzer, things will be even worse. Another approach is to utilize noncoherent modulation/demodulation techniques [16, p. 264] for the packet preamble, which contains the RF parameters of the packet, and then use coherent modulation/demodulation techniques to process the rest of the packet. It is somehow similar to the first approach, but unlike the first approach, both coherent and nocoherent modulation is used to generate

11

the signal transmitted to the air. Also, the RF parameters of the incoming signal is estimated not by an FFT spectrum analyzer, but by decoding the noncoherently modulated packet preamble. Since the radio receiver has no knowledge to narrow down the possible range of the spectrum of the incoming signal, it has to listen to a potentially large pool of spectrum, and thus leads to poor quality in decoding the RF parameters as in the other approach. R EFERENCES [1] [2] [3]

[4] [5]

[6]

[7]

[8]

[9]

[10] [11]

[12]

[13]

[14]

[15] [16]

FCC Spectrum Policy Task Force, “Report of the spectrum efficiency group,” Nov., 2002. R. J. Berger, “Open spectrum: A path to ubiquitous connectivity,” Queue, vol. 1, pp. 60–68, May 2003. P. Gupta and P. R. Kumar, “The capacity of wireless networks,” IEEE Transactions on Information Theory, vol. IT-46, pp. 388– 404, March 2000. J. G. Proakis, Digital Communications. McGraw-Hill, Inc., 3rd ed., 1995. S.-L. Wu, C.-Y. Lin, Y.-C. Tseng, and J.-P. Sheu, “A new multichannel MAC protocol with on-demand channel assignment for multi-hop mobile ad hoc networks,” in Int’l Symposium on Parallel Architectures, Algorithms and Networks (I-SPAN), 2000. D. Raychaudhuri and X. Jing, “A spectrum etiquette protocol for efficient coordination of radio devices in unlicensed bands,” in Proceedings of the 14th IEEE International Symposium on Personal Indoor, Mobile Radio Communications, (Beijing, China), pp. 172–176, September 2003. J. Chen, S. Sheu, and C. Yang, “A new multichannel access protocol for ieee 802.11 ad hoc wireless lans,” in Proc. of PIMRC 2003, pp. 2291 – 2296, 2003. A. Tzamaloukas and J. Garcia-Luna-Aceves, “Channel-hopping multiple access,” in Proceedings of IEEE ICC 2000, (New Orleans, LA), June 2000. P. Bahl, R. Chandra, and J. Dunagan, “SSCH: Slotted seeded channel hopping for capacity improvement in IEEE 802.11 adhoc wireless networks,” in MobiCom’04, (Philadelphia, PA), Sept. 2004. F. Zhao and L. J. Guibas, Wireless Sensor Networks: An information processing approach. Morgan Kaufmann Publishers, 2004. Z. Hass and J. Deng, “Dual Busy Tone Multiple Access (DBTMA) - a multiple access control scheme for ad hoc networks,,” IEEE Trans. Communications, vol. 50, pp. 975–985, June 2002. D. Cabric, S. M. Mishra, and R. W. Brodersen, “Implementation issues in spectrum sensing for cognitive radios,” in Proceedings of the 38th Asilomar Conference on Signals, Systems and Computers, (Pacific Grove, CA), 2004. S. H. Bae, S.-J. Lee, W. Su, and M. Gerla, “The design, implementation, and performance evaluation of the on-demand multicast routing protocol in multihop wireless networks,” IEEE Network Magzine, vol. 14, pp. 70–77, Jan./Feb. 2000. L. Kleinrock and F. A. Tobagi, “Packet switching in radio channels: Part I – Carrier sense multiple-access modes and their throughput-delay characteristics,” IEEE Trans. Commun.,, vol. COM-23, pp. 1400–1416, Dec. 1975. J. Schiller, Mobile Communications. Harlow, England: AddisonWesley, 2000. L. W. Couch II, Digital and Analog Communication Systems. Prentice Hall, Inc., 5th ed., 1997.