A Distributed Three-Hop Routing Protocol to Increase the Capacity of Hybrid Wireless Networks Haiying Shen*, Senior Member, IEEE, Ze Li and Chenxi Qiu

A Distributed Three-hop Routing Protocol to Increase the Capacity of Hybrid Wireless Networks Haiying Shen*, Senior Member, IEEE, Ze Li and Chenxi Qiu

Abstract—Hybrid wireless networks combining the advantages of both mobile ad-hoc networks and infrastructure wireless networks have been receiving increased attention due to their ultra-high performance. An efﬁcient data routing protocol is important in such networks for high network capacity and scalability. However, most routing protocols for these networks simply combine the ad-hoc transmission mode with the cellular transmission mode, which inherits the drawbacks of ad-hoc transmission. This paper presents a Distributed Three-hop Routing protocol (DTR) for hybrid wireless networks. To take full advantage of the widespread base stations, DTR divides a message data stream into segments and transmits the segments in a distributed manner. It makes full spatial reuse of a system via its high speed ad-hoc interface and alleviates mobile gateway congestion via its cellular interface. Furthermore, sending segments to a number of base stations simultaneously increases throughput and makes full use of widespread base stations. In addition, DTR signiﬁcantly reduces overhead due to short path lengths and the elimination of route discovery and maintenance. DTR also has a congestion control algorithm to avoid overloading base stations. Theoretical analysis and simulation results show the superiority of DTR in comparison with other routing protocols in terms of throughput capacity, scalability and mobility resilience. The results also show the effectiveness of the congestion control algorithm in balancing the load between base stations.

Index Terms—Hybrid wireless networks, Routing algorithm, Load balancing, Congestion control !

1 INTRODUCTION by integrating all kinds of wireless devices into the net- Over the past few years, wireless networks including work. In an infrastructure network, nodes communicate infrastructure wireless networks and mobile ad-hoc net- with each other through base stations (BSes). Because works (MANETs) have attracted significant research in- of the long distance one-hop transmission between BSes terest. The growing desire to increase wireless network and mobile nodes, the infrastructure wireless networks capacity for high performance applications has stimulat- can provide higher message transmission reliability and ed the development of hybrid wireless networks [1–6]. channel access efficiency, but suffer from higher power A hybrid wireless network consists of both an infras- consumption on mobile nodes and the single point of tructure wireless network and a mobile ad-hoc network. failure problem [11]. Wireless devices such as smart-phones, tablets and lap- A hybrid wireless network synergistically combines tops, have both an infrastructure interface and an ad- an infrastructure wireless network and a mobile ad- hoc interface. As the number of such devices has been hoc network to leverage their advantages and overcome increasing sharply in recent years, a hybrid transmission their shortcomings, and finally increases the throughput structure will be widely used in the near future. Such a capacity of a wide-area wireless network. A routing structure synergistically combines the inherent advan- protocol is a critical component that affects the through- tages and overcome the disadvantages of the infrastruc- put capacity of a wireless network in data transmis- ture wireless networks and mobile ad-hoc networks. sion. Most current routing protocols in hybrid wireless In a mobile ad-hoc network, with the absence of a networks [1, 5, 6, 12–18] simply combine the cellular central control infrastructure, data is routed to its des- transmission mode (i.e. BS transmission mode) in infras- tination through the intermediate nodes in a multi-hop tructure wireless networks and the ad-hoc transmission manner. The multi-hop routing needs on-demand route mode in mobile ad-hoc networks [8, 9, 7]. That is, as discovery or route maintenance [7–10]. Since the mes- shown in Figure 1 (a), the protocols use the multi-hop sages are transmitted in wireless channels and through routing to forward a message to the mobile gateway dynamic routing paths, mobile ad-hoc networks are not nodes that are closest to the BSes or have the highest as reliable as infrastructure wireless networks. Further- bandwidth to the BSes. The bandwidth of a channel more, because of the multi-hop transmission feature, is the maximum throughput (i.e., transmission rate in mobile ad-hoc networks are only suitable for local area bits/s) that can be achieved. The mobile gateway nodes data transmission. then forward the messages to the BSes, functioning as The infrastructure wireless network (e.g. cellular net- bridges to connect the ad-hoc network and the infras- work) is the major means of wireless communication in tructure network. our daily lives. It excels at inter-cell communication (i.e., However, direct combination of the two transmission communication between nodes in different cells) and modes inherits the following problems that are rooted Internet access. It makes possible the support of uni- in the ad-hoc transmission mode. versal network connectivity and ubiquitous computing • High overhead. Route discovery and maintenance incur high overhead. The wireless random access medium • * Corresponding Author. Email: [email protected]; Phone: (864) 656 access control (MAC) required in mobile ad-hoc net- 5931; Fax: (864) 656 5910. works, which utilizes control handshaking and a back- off mechanism, further increases overhead. • The authors are with the Department of Electrical and Computer Engi- • Hot spots. The mobile gateway nodes can easily become neering, Clemson University, Clemson, SC, 29634. hot spots. The RTS-CTS random access, in which most E-mail: {shenh, zel, chenxiq}@clemson.edu traffic goes through the same gateway, and the flooding employed in mobile ad-hoc routing to discover routes 2

• High reliability. Because of its small hop path length with a short physical distance in each step, it alleviates noise and neighbor interference and avoids the adverse effect of route breakdown during data transmission. Thus, it reduces the packet drop rate and makes full use of spacial reuse, in which several source and destination nodes can communicate simultaneously without interference. The rest of this paper is organized as follows. Sec- tion 2 presents a review of representative hybrid wireless networks and multi-hop routing protocols. Section 3 details the DTR protocol, with an emphasis on its routing methods, segment structure, and BS congestion control. Section 4 theoretically analyzes the performance of the Fig. 1: Traditional and proposed routing algorithms on the DTR protocol. Section 5 shows the performance of the uplink direction. DTR protocol in comparison to other routing protocols. may exacerbate the hot spot problem. In addition, mobile Finally, Section 6 concludes the paper. nodes only use the channel resources in their route direction, which may generate hot spots while leave resources in other directions under-utilized. Hot spots lead to low 2 RELATED WORK transmission rates, severe network congestion, and high In order to increase the capacity of hybrid wireless net- data dropping rates. works, various routing methods with different features • Low reliability. Dynamic and long routing paths lead have been proposed. One group of routing methods to unreliable routing. Noise interference and neighbor integrate the ad-hoc transmission mode and the cellu- interference during the multi-hop transmission process lar transmission mode [1, 5, 6, 14, 16–18]. Dousse et cause a high data drop rate. Long routing paths increase al. [6] built a Poisson Boolean model to study how a the probability of the occurrence of path breakdown BS increases the capacity of a MANET. Lin et al. [5] due to the highly dynamic nature of wireless ad-hoc proposed a Multihop Cellular Network and derived networks. its throughput. Hsieh et al. [14] investigated a hybrid These problems become an obstacle in achieving high IEEE 802.11 network architecture with both a distributed throughput capacity and scalability in hybrid wireless coordination function and a point coordination function. networks. Considering the widespread BSes, the mobile Luo et al. [1] proposed a unified cellular and ad-hoc nodes have a high probability of encountering a BS while network architecture for wireless communication. Cho et moving. Taking advantage of this feature, we propose a al. [16] studied the impact of concurrent transmission in Distributed Three-hop Data Routing protocol (DTR). In a downlink direction (i.e. from BSes to mobile nodes) DTR, as shown in Figure 1 (b), a source node divides on the system capacity of a hybrid wireless network. a message stream into a number of segments. Each In [17, 18], a node initially communicates with other segment is sent to a neighbor mobile node. Based on nodes using an ad-hoc transmission mode, and switches the QoS requirement, these mobile relay nodes choose to a cellular transmission mode when its performance is between direct transmission or relay transmission to the better than the ad-hoc transmission. BS. In relay transmission, a segment is forwarded to The above methods are only used to assist intra-cell another mobile node with higher capacity to a BS than ad-hoc transmission rather than inter-cell transmission. the current node. In direct transmission, a segment is In inter-cell transmission [1, 5, 6], a message is forwarded directly forwarded to a BS. In the infrastructure, the via the ad-hoc interface to the gateway mobile node that segments are rearranged in their original order and sent is closest to or has the highest uplink transmission band- to the destination. The number of routing hops in DTR width to a BS. The gateway mobile node then forwards is confined to three, including at most two hops in the the message to the BS using the cellular interface. How- ad-hoc transmission mode and one hop in the cellular ever, most of these routing protocols simply combine transmission mode. To overcome the aforementioned routing schemes in ad-hoc networks and infrastructure shortcomings, DTR tries to limit the number of hops. The networks, hence inherit the drawbacks of the ad-hoc first hop forwarding distributes the segments of a mes- transmission mode as explained previously. sage in different directions to fully utilize the resources, DTR is similar to the Two-hop transmission proto- and the possible second hop forwarding ensures the high col [19] in terms of the elimination of route maintenance capacity of the forwarder. DTR also has a congestion and the limited number of hops in routing. In Two-hop, control algorithm to balance the traffic load between the when a node’s bandwidth to a BS is larger than that nearby BSes in order to avoid traffic congestion at BSes. of each neighbor, it directly sends a message to the BS. Using self-adaptive and distributed routing with high- Otherwise, it chooses a neighbor with a higher channel speed and short-path ad-hoc transmission, DTR signifi- and sends a message to it, which further forwards the cantly increases the throughput capacity and scalability message to the BS. DTR is different from Two-hop in of hybrid wireless networks by overcoming the three three aspects. First, Two-hop only considers the node shortcomings of the previous routing algorithms. It has transmission within a single cell, while DTR can also the following features: deal with inter-cell transmission, which is more challeng- • Low overhead. It eliminates overhead caused by route ing and more common than intra-cell communication in discovery and maintenance in the ad-hoc transmission the real world. Second, DTR uses distributed transmis- mode, especially in a dynamic environment. sion involving multiple cells, which makes full use of • Hot spot reduction. It alleviates traffic congestion at system resources and dynamically balances the traffic mobile gateway nodes while makes full use of channel load between neighboring cells. In contrast, Two-hop resources through a distributed multi-path relay. employs single-path transmission. 3

a hybrid wireless network. We simplify the routings in the infrastructure network for clarity. As shown in the figure, when a source node wants to transmit a message stream to a destination node, it divides the message stream into a number of partial streams called segments and transmits each segment to a neighbor node. Upon receiving a segment from the source node, a neighbor node locally decides between direct transmission and relay transmission based on the QoS requirement of the application. The neighbor nodes forward these segments in a distributed manner to nearby BSes. Relying on the infrastructure network routing, the BSes further transmit the segments to the BS where the destination node resides. The final BS rearranges the segments into the Fig. 2: Data transmission in the DTR protocol. original order and forwards the segments to the destination. It uses the cellular IP transmission method [30] There are other methods proposed to improve routing to send segments to the destination if the destination performance in hybrid wireless networks. Wu et al. [3] moves to another BS during segment transmission. proposed using ad-hoc relay stations to dynamically Our DTR algorithm avoids the shortcomings of ad- relay traffic from one cell to another in order to avoid hoc transmission in the previous routing algorithms that traffic congestion in BSes. Li et al. [20] surveyed a num- directly combine an ad-hoc transmission mode and a ber of multi-hop cellular network (MCN) architectures cellular transmission mode. Rather than using the multi- in literature, and compared and discussed methods to hop ad-hoc transmission, DTR uses two hop forwarding reduce the cost of deployment for MCNs. The work by relying on node movement and widespread base sta- in [21] investigates how to allocate the bandwidth to tions. All other aspects remain the same as those in the users to improve the performance of hybrid wireless previous routing algorithms (including the interaction networks. Thulasiraman et al. [22] further considered with the TCP layer). DTR works on the Internet layer. It the wireless interference in optimizing the resource al- receives packets from the TCP layer and routes it to the location in hybrid wireless networks. The work in [23] destination node, where DTR forwards the packet to the proposes a coalitional game theory based cooperative TCP layer. packet delivery scheme in hybrid wireless networks. The data routing process in DTR can be divided into There are also some works [24–26] that study radio two steps: uplink from a source node to the first BS and frequency allocation for direction transmission and relay downlink from the final BS to the data’s destination. transmission in hybrid wireless networks. These works Critical problems that need to be solved include how are orthogonal to our study in this paper and can be in- a source node or relay node chooses nodes for efficient corporated into DTR to further enhance its performance. segment forwarding, and how to ensure that the final BS The throughput capacity of the hybrid wireless net- sends segments in the right order so that a destination work under different settings has also been an active node receives the correct data. Also, since traffic is research topic in the hybrid wireless network. The works not evenly distributed in the network, how to avoid in [17, 27] have studied the throughput of hybrid net- overloading BSes is another problem. Below, Section 3.2 work with n nodes and m stations. Liu et al. [28] theoret- will present the details for forwarding node selection in ically studied the capacity of hybrid wireless networks uplink transmission and Section 3.3 will present the seg- under an one-dimensional network topology and a two- ment structure that helps ensure the correct final order of dimensional strip topology. Wang et al. [29] studied segments in a message, and DTR’s strategy for downlink the multicast throughput of hybrid wireless networks transmission. Section 3.4 will present the congestion and designed an optimal multicast strategy based on control algorithm for balancing a load between BSes. deduced throughput. 3.2 Uplink Data Routing 3 DISTRIBUTED THREE-HOP ROUTING PRO- A long routing path will lead to high overhead, hot spots TOCOL and low reliability. Thus, DTR tries to limit the path 3.1 Assumption and Overview length. It uses one hop to forward the segments of a Since BSes are connected with a wired backbone, we as- message in a distributed manner and uses another hop sume that there are no bandwidth and power constraints to find high-capacity forwarder for high performance on transmissions between BSes. We use intermediate n- routing. As a result, DTR limits the path length of uplink odes to denote relay nodes that function as gateways routing to two hops in order to avoid the problems connecting an infrastructure wireless network and a of long-path multi-hop routing in the ad-hoc networks. mobile ad-hoc network. We assume every mobile node is Specifically, in the uplink routing, a source node initially dual-mode; that is, it has ad-hoc network interface such divides its message stream into a number of segments, as a WLAN radio interface and infrastructure network then transmits the segments to its neighbor nodes. The interface such as a 3G cellular interface. neighbor nodes forward segments to BSes, which will DTR aims to shift the routing burden from the ad- forward the segments to the BS where the destination hoc network to the infrastructure network by taking resides.Below, we first explain how to define capacity, advantage of widespread base stations in a hybrid wire- then introduce the way for a node to collect the capacity less network. Rather than using one multi-hop path to information from its neighbors, and finally present the forward a message to one BS, DTR uses at most two hops details of the DTR routing algorithm. to relay the segments of a message to different BSes in Different applications may have different QoS require- a distributed manner, and relies on BSes to combine the ments, such as efficiency, throughput, and routing speed. segments. Figure 2 demonstrates the process of DTR in For example, delay-tolerant applications (e.g. voice mail, 4 e-mail and text messaging) do not necessarily need fast Base station Mobile node real-time transmission and may make throughput the highest consideration to ensure successful data trans- 9 mission. Some applications may take high mobility as 4 4 their priority to avoid hot spots and blank spots. Hot 7 spots are areas where BS channels are congested, while A 7 5 5 7 blank spots are areas without signals or with very weak 2 5 4 10 signals. High-mobility nodes can quickly move out of a Source B 3 3 hot spot or blank spot and enter a cell with high band- 9 6 10 width to a BS, thus providing efficient data transmission. 8 1 Throughput can be measured by bandwidth, mobility can be measured by the speed of node movement, and (a) Source node has (b) Source node has no routing speed can be measured by the speed of data higher‐capacity neighbors. forwarding. Bandwidth can be estimated using the non- higher‐capacity neighbors. intrusive technique proposed in [31]. In this work, we Fig. 3: Neighbor selection in DTR. take throughput and routing speed as examples for ment. As the neighborhood scope of a node for high- the QoS requirement. We use a metric bandwidth/queue capacity node searching grows, the probability of finding to reflect node capacity in throughput and fast data higher capacity nodes increases. Thus, a source node’s forwarding. The metric is the ratio of a node’s channel neighbors are more likely to find neighbors with higher bandwidth to its message queue size. A larger band- capacities than the source node. Therefore, transmitting width/queue value means higher throughput and message data segments to neighbors and enabling them to choose forwarding speed, and vice versa. the second relays help to find higher capacity nodes to When choosing neighbors for data forwarding, a node forward data. If a source node has the highest capacity needs the capacity information (i.e., queue size and in its region, the segments will be forwarded back to the bandwidth) of its neighbors. Also, a selected neighbor source node according to the DTR protocol. The source should have enough storage space for a segment. To node then forwards the segments to the BSes directly keep track of the capacity and storage space of its neigh- due to the three-hop limit. Though sending data back bors, each node periodically exchanges its current capac- and forth leads to latency and bandwidth wastage, this ity and storage information with its neighbors. In the ad- case occurs only when the source nodes is the highest hoc network component, every node needs to periodical- capacity node within its two-hop neighborhood. Also, ly send “hello” messages to identify its neighbors. Tak- this step is necessary for finding the highest capacity ing advantage of this policy, nodes piggyback the capac- nodes within the source’s two-hop neighborhood, and ity and storage information onto the “hello” messages in ensures that the highest capacity nodes are always s- order to reduce the overhead caused by the information elected as the message forwarders. If the source node exchanges. If a node’s capacity and storage space are does not distribute segments to its neighbors, the higher- changed after its last “hello” message sending when it capacity node searching cannot be conducted. Note receives a segment, it sends its current capacity and stor- that the data transmission rate of the ad-hoc interface age information to the segment forwarder. Then, the seg- (e.g. IEEE 802.11) is more than 10 times faster than the ment forwarder will choose the highest capacity nodes cellular interface (e.g. GSM, 3G). Thus, the transmission in its neighbors based on the most updated information. delay for sending the data back and forth in the ad-hoc When a source node sends out message segments, transmission is negligible in the total routing latency. it chooses the neighbors that have enough space for By distributing a message’s segments to different n- storing a segment, and then chooses neighbors that have odes to be forwarded in different directions, our algo- the highest capacity. In order to find higher capacity rithm reduces the congestion in the previous routing forwarders in a larger neighborhood around the source, algorithms in the hybrid wireless networks. When a each segment receiver further forwards its received seg- node selects a relay to forward a segment, it checks the ment to its neighbor with the highest capacity. That capacity of the node. Only when a node, say node mi, is, after a neighbor node mi receives a segment from has enough capacity, the node will forward a segment the source, it uses either direct transmission or relay to node mi. Therefore, even though the paths are not transmission. If the capacity of each of its neighbors is node-disjoint, there will be no congestion in the common no greater than itself, relay node mi uses direct trans- sub-paths. mission. Otherwise, it uses relay transmission. In direct Figure 3 shows examples of neighbor selection in DTR, transmission, the relay node sends the segment to a BS in which the source node is in the transmission range of if it is in a BS’s region. Otherwise, it stores the segment a BS. In the figures, the value in the node represents while moving until it enters a BS’s region. In relay its capacity. In scenario (a), there exist nodes that have transmission, relay node mi chooses its highest-capacity higher capacity than the source node within the source’s neighbor as the second relay node based on the QoS two-hop neighborhood. If a routing algorithm directly requirement. The second relay node will use direct trans- let a source node transmit a message to its BS, the mission to forward the segment directly to a BS. As a re- high routing performance cannot be guaranteed since sult, the number of transmission hops in the ad-hoc net- the source node may have very low capacity. In DTR, work component is confined to no more than two. The the source node sends segments to its neighbors, which small number of hops help to increase the capacity of the further forward the segments to nodes with higher ca- network and reduce channel contention in ad-hoc trans- pacities. In scenario (b), the source node has the highest mission. Algorithm 1 shows the pseudo-code for neigh- capacity among the nodes in its two-hop neighborhood. bor node selection and message forwarding in DTR. After receiving segments from the source node, some The purpose of the second hop selection is to find neighbors forward the segments back to the source node, a higher capacity node as the message forwarder in which sends the message to its BS. Thus, DTR always order to improve the performance of the QoS require- arranges data to be forwarded by nodes with high 5

capacity to their BSes. DTR achieves higher throughput its IP address to the home BS of mi. When a BS wants and faster data forwarding speed by taking into account to contact mi, it contacts the home BS of mi to find the node capacity in data forwarding. destination BS where mi currently resides at. However, the destination BS recorded in the home BS Algorithm 1 Pseudo-code for neighbor node selection may not be the most up-to-date destination BS since and message forwarding. destination mobile nodes switch between the coverage 1: ChooseRelay( ) { regions of different BSes during data transmission to 2: //choose neighbors with sufficient caches and bandwidth/queue (b/q) rates them. For instance, data is transmitted to BS Bi that has 3: Query storage size and QoS requirement info. from neighbors the data’s destination, but the destination has moved 4: for each neighbor n do to the range of BS Bj before the data arrives at BS 5: if n.cache.size>segment.length && n.b/q>this.b/q then Bi. To deal with this problem, we adopt the Cellular 6: Add n to R = {r1, . . . rm, ...} in a descending order of b/q IP protocol [30] for tracking node locations. With this 7: end if protocol, a BS has a home agent and a foreign agent. 8: end for 9: Return R The foreign agent keeps track of mobile nodes moving 10: } into the ranges of other BSes. The home agent intercepts 11: Transmission( ) { in-coming segments, reconstructs the original data, and 12: if it is a source node then re-routes it to the foreign agent, which then forwards the 13: //routing conducted by a source node data to the destination mobile node. 14: //choose relay nodes based on QoS requirement After the destination BS receives the segments of a 15: R=ChooseRelay( ); message, it rearranges the segments into the original 16: Send segments to {r1, . . . rm} in R message and then sends it to the destination mobile 17: else node. A vital issue is guaranteeing that the segments are 18: //routing conducted by a neighbor node 19: if this.b/q ≤ b/q of all neighbors then combined in the correct order. For this purpose, DTR 20: //direct transmission specifies the segment structure format. Each segment 21: if within the range of a BS then contains eight fields, including: (1) source node IP ad- 22: Transmit the segment directly to the BS dress (denoted by S); (2) destination node IP address 23: end if (denoted by D); (3) message sequence number (denoted 24: else by m); (4) segment sequence number (denoted by s); 25: //relay transmission (5) QoS indication number (denoted by q); (6) data; (7) 26: nodei=getHighestCapability(ChooseRelay( )) length of the data; and (8) checksum. Fields (1)-(5) are 27: Send a segment to nodei 28: end if in the segment head. 29: end if The role of the source IP address field is to inform 30: } the destination node where the message comes from. The destination IP address field indicates the destination node, and is used to locate the final BS. After sending Algorithm 2 Pseudo-code for a BS to reorder and for- out a message stream to a destination, a source node ward segments to destination nodes. may send out another message stream to the same 1: //a cache pool is built for each data stream destination node. The message sequence number differenti- 2: //there are n cache pools currently ates the different message streams initiated by the same 3: if receives a segment (S,D,m,s,q) then source node. The segment sequence number is used to find 4: if there is no cache pool with msg sequence num equals m then the correct transmission sequence of the segments for 5: Create a cache pool n + 1 for the stream m transmission to a destination node. The data is the actual 6: else information that a source node wants to transmit to a 7: //the last delivered segment of stream m has sequence num i − 1 8: if s == i then destination node. The length field specifies the length 9: Send out segment (S,D,m,s,q) to D of the DTR segment including the header in bytes. The 10: i + +; checksum is used by the receiver node to check whether 11: else the received data has errors. The QoS indication number is 12: Add segment (S,D,m,s) into cache pool m used to indicate the QoS requirement of the application. 13: end if Thus, each segment’s head includes the information 14: end if represented by (S, D, m, s, q)(m, s = 1, 2, 3, ...). When a 15: end if segment with head (S, D, m, s, q) arrives at a BS, the BS contacts D’s home BS to find the destination BS where D stays via the mobile IP protocol. It then transmits the 3.3 Downlink Data Routing and Data Reconstruction segment to the destination BS through the infrastructure As mentioned above, the message stream of a source n- network component. After arriving at the BS, the seg- ode is divided into several segments. After a BS receives ment waits in the cache for its turn to be transmitted to a segment, it needs to forward the segment to the BS, its destination node based on its message and segment where the destination node resides (i.e., the destination sequence numbers. At this time, if another segment BS). We use the mobile IP protocol [32] to enable BSes to comes with a head labelled (S, D, (m + 1), s, q), which know the destination BS. In this protocol, each mobile means that it is from the same source node but belongs node is associated with a home BS, which is the BS in the to another data stream, the BS will put it to another node’s home network, regardless of its current location stream. If the segment is labeled as (S, D, m, (s+1), q), it in the network. The home network of a node contains its means that this segment belongs to the same data stream registration information identified by its home address, of the same source node as segment (S, D, m, s, q). The which is a static IP address assigned by an ISP. In combination of the source node’s sequence number and a hybrid wireless network, each BS periodically emits segment sequence number helps to locate the stream beacon signals to locate the mobile nodes in its range. and the position of a segment in the steam. In order to When a mobile node mi moves away from its home BS, integrate the segments into their correct order to retrieve the BS where mi currently resides detects mi and sends the original data, the segments in the BS are transmitted 6 to the destination node in the order of the segments’ of b should be carefully determined based on the avail- sequence in the original message. If a segment has not able resources for routing and the reliability demand. If arrived at the final BS, its subsequent segments will wait b is set to a large value, the routing reliability is high at in the final BS until its arrival. Algorithm 2 shows the the cost of high overhead. If b is set to a small value, the pseudo-code for a BS to reorder and forward segments to routing reliability is low while the overhead is reduced. their destinations. Note that in the cache, we can set the After the neighboring BSes receive the segments, they timer based on the packet rate and storage limit. In other further forward the segments to the destination BS, words, the timer should be set as large as possible to which forwards the segments to the destination node. fully utilize the storage on BSes to ensure that a message In this way, the heavy traffic from mobile nodes to a BS has a high probability to be recovered. can be distributed among neighboring BSes quickly. Next, we discuss how to handle the case when the destination BS is congested. If a BS has not received 3.4 Congestion Control in Base Stations confirmation from the destination BS during a certain Compared to the previous routing algorithms in hybrid time period after it sends out a segment, it assumes wireless networks, DTR can distribute traffic load among that the destination BS is overloaded. Then, it sends mobile nodes more evenly. Though the distributed rout- the segment to b (b ≥ 1) lightly loaded neighboring ing in DTR can distribute traffic load among nearby B- BSes of the destination BS from its routing table. If an Ses, if the traffic load is not distributed evenly in the net- attempted neighboring BS does not respond during a work, some BSes may become overloaded while other B- certain time period, it is also considered as overloaded. Ses remain lightly loaded. We propose a congestion con- Then, the BS keeps trying other neighboring BSes until trol algorithm to avoid overloading BSes in uplink trans- finding lightly loaded BSes. Redundant neighboring mission (e.g., B1, B2 and B3 in Figure 1 (b)) and down- BSes are selected in order to increase routing reliability. link transmission (e.g., B4 in Figure 1 (b)), respectively. The constant b should be set to an appropriate value In the hybrid wireless network, BSes send beacon mes- considering factors such as the network size and the sages to identify nearby mobile nodes. Taking advantage amount of traffic in order to achieve an optimal trade-off of this beacon strategy, once the workload of a BS, say between overhead and reliability. Bi, exceeds a pre-defined threshold, Bi adds an extra After receiving the message, each lightly loaded neigh- bit in its beacon message to broadcast to all the nodes boring BS of the destination BS finds a multi-hop path in its transmission range. Then, nodes near Bi know to the destination mobile node. It broadcasts a path that Bi is overloaded and will not forward segments to query message, which includes the IDs of the destination Bi. When a node near Bi, say mi, needs to forward a BS and the destination node, to the mobile nodes in segment to a BS, it will send the segment to Bi based on its region. The path querying process is similar to the the DTR algorithm. In our congestion control algorithm, previous path querying for a lightly loaded BS. The because Bi is overloaded, rather than targeting Bi, mi nodes further forward the path query to their neighbors will forward the segment to a lightly loaded neighboring until the query reaches the destination node. Here, we BS of Bi. To this end, node mi first queries a multi-hop do not piggyback the query to beacon messages because path to a lightly loaded neighboring BS of Bi. Node mi this querying is for a specific mobile node rather than broadcasts a query message into the system. We set the any mobile node near a lightly loaded BS. Including the TTL for the path query forwarding step to a constant mobile node’s ID into beacon messages generates very (e.g., 3). The query message is forwarded along other high overhead. nodes until a node (say mj) near a lightly loaded BS In order to reduce (say Bj) is reached. Due to broadcasting, a node may the broadcasting receive multiple copies of the same queries. Each node overhead, a mobile B1 only remembers mi and the node that forwards the first node residing in s query (i.e., its preceding node), and ignores all other the the region of a BS B B same queries. In this way, a multi-hop path between the not close to the 4 2 source node and the lightly loaded base station can be destination BS drops formed. Node mj responds to the path query by adding the query. The nodes B3 a reply bit and the address of mi into its beacon message can determine their D to its preceding node in the path. This beacon receiver approximate relative B Mobile node B6 5 Source node also adds a reply bit and the address of mi into its positions to BSes by Destination node beacon message to its preceding node in the path. This sensing the signal Fig. 4: Congestion control on BSes. process repeats until mi receives the beacon. Thus, each strengths from differ- node knows its preceding node and succeeding node ent BSes. Each node adds the strength of its received in the path from mi and mj based on the address of signal into its beacon message that is periodically ex- mi. Then, mi’s message can be forwarded along the changed between neighbor nodes so that the nodes can observed path along the nodes. The observed path can identify their relative positions to each other. Only those always be used by mi for any subsequent messages to mobile nodes that stay farther than the query forwarder Bj as long as it is not broken. The neighboring BSes from the forwarder’s BS forward the queries in the of an overloaded BS may also be overloaded. As the direction of the destination BS. In this way, the query mobile nodes near an overloaded BS know that the BS can be forwarded to the destination BS faster. After the is overloaded, when they receive a query message to multi-hop path is discovered, the neighboring BS sends find a path to an underloaded BS, they do not forward the segment to the destination node along the path. the message towards their overloaded BSes. Since the destination node is in the neighboring BS’s Node mi may receive responses from a few nodes region, the overhead to identify a path to the destination near BSes. It can choose b (b ≥ 1) neighboring BSes of node is small. Note that our methods for congestion the destination to forward the segment. The redundant control in base stations involve query broadcasting. transmissions enhance the data transmission reliability However, it is used only when some base stations are while also increase the routing overhead. Thus, the value overloaded rather than in the normal DTR routing 7

algorithm in order to avoid load congestion in BSes. where |vi −vj| represents the Euclidean distance between Figure 4 shows an example of the congestion control vi and vj in the plane. on BSes when b = 2. As shown in figure, BS B1 is 2) For any other node vk that is simultaneously trans- congested. Then, the relay nodes of the source node’s mitting over the same channel, message broadcast locally by beacon piggybacking to |vk − vj | ≥ (1 + ∆)|vi − vj |. (3) find multi-hop paths which lead to B3 and B4. The Formula (3) guarantees a guard zone around the receiv- relay nodes then send segments along the paths. In this ing node to prevent a neighboring node from transmit- way, the traffic originally targeting overloaded B1 can ting on the same channel at the same time. The radius be spread out to the neighboring BSes B3 and B4. B3 of the guard zone is (1 + ∆) times the distance between and B4 further forward the segments to the destination the sender and the receiver. The parameter ∆ defines the BS B6 if B6 is not congested. If B6 is also congested, size of the guard zone and we require that ∆ > 0. B3 and B4 send the segments to the neighboring BSes We first adopt a concept called aggregate throughput of B6. Specifically, B4 sends the segment to B3. B3 does capacity introduced in [17, 33] to measure the throughput not forward the segment to another BS since it already of the network. is close to B6. B3 then finds a multi-hop path to the destination node and uses ad-hoc transmission to forward Definition (Aggregate Throughput Capacity of Hybrid the segments to the destination node. Similarly, when B2 Networks) The aggregate throughput capacity of a hy- wants to send a segment to the destination node, it also brid wireless network is of order Θ(f(σ, M)) if there are uses a multi-hop path for the segment transmission. deterministic constants α > 0, and α0 < +∞ such that lim Pr (P (σ, M) = αf(σ, M) is feasible) = 1 (4) M→∞ lim inf Pr P (σ, M) = α0f(σ, M) is feasible < 1. (5) 4 PERFORMANCE ANALYSIS OF THE DTR M→∞ PROTOCOL Since the working frequency of infrastructure net- σ M works is around 700MHz while that of ad-hoc networks Node density Number of BSes is 2.4 GHz, the communications in infrastructure mode l Segment’s length sh Area size of a cell n(S) Number of nodes in area S R Transmission range (between mobile nodes and BSes through cellular interface) would not generate interference to ad-hoc mode. Wi Bandwidth of a node vi mi Mobile node i P (σ, M)Throughput n(σ, M) Number of nodes We divide the channel for infrastructure mode transmissions into uplink and downlink parts, according to the TABLE 1: Parameter table. transmission direction relative to the BSes. Accordingly, In this section, we analyze the effectiveness of the in the DTR protocol, the traffic of each S-D pair is DTR protocol at enhancing the capacity and scalability composed of at most two intra-cell traffics, one uplink of hybrid wireless networks. In our analysis, we use the traffic and one download traffic. Since uplink traffic and same scenario in [17] for hybrid wireless networks, and downlink traffic use different sub-channels, there is also use the same scenario in [33] for the ad-hoc network no interference between these two types of traffics. For each node vi, we denote the bandwidth assigned to intra- component. We present the scenarios and some concepts int up below. We consider a large number of mobile nodes cell, uplink, and downlink sub-channels by Wi , Wi down up down uniformly and randomly deployed over a 2-D field. and Wi , respectively. We let Wi = Wi because The moving directions of the nodes are independent there are the same amount of uplink traffic and down- and identically distributed (i.i.d.). The distribution of link traffic. The transmission rates should sum to Wi, int up down mobile nodes can be modeled as a homogeneous Poisson i.e., Wi + Wi + Wi = Wi. Though no interference process with node density σ [34]. That is, given an area exists between intra-cell, uplink, and downlink traffics, of size S in the field, the number of nodes in the area, interference exists between the same type of traffic in a denoted by n(S), follows the Poisson distribution with cell and between different cells. Fortunately, there is an the parameter σS, (σS)ke−σS efficient spatial transmission schedule that can prevent Pr (n(S) = k) = , k = 0, 1, 2, ... (1) such interferences [17]. First, to avoid the interference k! in a cell, any two nodes within the cell are not allowed Besides mobile nodes, there are M BSes regularly de- to transmit with the same traffic mode at the same time. ployed in the field. The BSes divide the area into a Second, to avoid the interference between different cells, hexagon tessellation, in which each hexagon has side the cells are spatially divided into a number of groups length h. The BSes are assumed to be connected together and transmissions in the cells of the same group do not by a wired network. We assume that the link bandwidths interfere with each other. If the groups are scheduled in the wired network are large enough so that there to transmit in a round robin fashion, each cell will be are no bandwidth constraints between BSes. In single- able to transmit once every fixed amount of time without path transmission, a message is sequentially transmitted interfering with each other. in one routing path. In multi-path transmission, a message Below, we show how many groups we need to divide is divided into a number of segments that are forwarded the cells to prevent interference. We adopt the notion along multiple paths in a distributed manner. We assume of interfering neighbors introduced in [17], and give the each segment has the same length l. Table 1 lists the number of cells that can be affected by a transmission in notations used in our analysis. one cell. Two cells are defined to be interfering neighbors We assume that the transmission range of all mobile if there is a point in one cell which is within a distance nodes and all BSes is R (R > h). In this paper, we use (2+∆)R of a point in the other cell. Accordingly, if two protocol model [17, 33] to describe the interference among cells are not interfering neighbors, transmissions in one nodes; that is, a transmission from a node (here “node” cell do not interfere with transmissions in the other cell. can be either mobile node or BS) vi to another node vj is [17] has proved that (1) each cell has no more than c1 successful if the following two conditions are satisfied: interfering neighbors (Lemma 1 in [17]), where c1 is a 2 1) vj is within the transmission range of vi, i.e., constant 4 3l + 2R + ∆R c1 = , (6) |vi − vj | ≤ R (2) 3 3l 8

and (2) all cells should be divided into c1 + 1 groups P is upper bounded by MW/4, which can be achieved and the whole channel should be divided into c1 + 1 only if each cell has one node to send the message. subchannels, where each subchannel is allocated to the Hence, Formula (18) can be satisfied by setting α0 to 1/2. cells in one group. Thus, the number of group we need Then, we will show how Formula (17) can be satisfied. to divide the cells to prevent interference is c1 + 1. Since the same message has to go through an uplink and Before calculating the aggregate throughput capacity a downlink and it is counted only once in the through- of DTR, we first introduce Lemma 4.1. put capacity, calculating the throughput of the whole Lemma 4.1: The number of cells that have mobile n- network is equivalent to calculating the throughput of odes is Θ(M). uplink traffic Pup or the throughput of downlink traffic Proof: Denote the number of cells having mobile Pdown. Notice calculating intra-cell traffic throughput is nodes by M1. To prove M1 = Θ(M), we need to prove not accurate because a message may transmit twice with that there exists deterministic constants α > 0 and intra-cell mode. In this proof, we calculate Pup. α0 < +∞ such that First, we consider the throughput of the uplink traffic lim Pr (M1 = αM) = 1, (7) k M→∞ of an arbitrary cell k, denoted by Pup. Since the schedule 0 lim inf Pr M1 = α M < 1. (8) allocates 1/(c1 + 1) time slots to this cell, then M→∞ up 0 k W For Formula (8), let α = 2. Because the number of cells Pup = . (19) having mobile nodes is upper bounded by M, then c1 + 1 Then, we consider the throughput of the whole network. lim inf Pr (M1 = 2M is feasible) = 0. (9) M→∞ P = PM P i X Now, we prove that Formula (7) can also be satisfied Let up i=1 up i represent the throughput of uplink traffic, then we have for some constant α. Because the number of nodes in a cell follows a Poisson distribution√ and the size of each c2MW 2 lim Pr Pup = cell (hexagon) is sh = 3 3h , then we can derive the M→∞ 3(c1 + 1) probability that no mobile node is in a cell equals M up ! X i c2MW 0 −sh = lim Pr P X = σ e −sh up i Pr (n(Sh) = 0) = = e . (10) M→∞ c1 + 1 0! i=1 M ! Consider an arbitrary cell k, let X1,X2, ..., Xk,..., XM X X = lim Pr Xi = c2M = 1 (By Lemma 4.1) be i.i.d. random variables, where k represents whether M→∞ cell k has mobile nodes. Then, Xk is defined as follows: i=1 Accordingly, Formula (17) can be satisfied when α is set 1 cell k has mobile nodes c Xk = (11) to 2 . 0 cell k does not have mobile nodes 3(c1+1) −s −s and E(Xk) = e h . For simplicity, let c2 = 1 − e h . Corollary 4.1: With the restriction in Theorem 4.1, DTR PM can achieve Θ(W ) throughput per S-D pair. Then, M1 = Xk. By the Strong Law of Large Number k=1 Proof: Denote the throughput of per S-D pair by P , (SLLN) [34], PM ! k=1 Xk which equals P (σ, M) Pr lim = c2 = 1, (12) P = . (20) M→∞ M n W lim Pr (M = c M) = 1 Obviously, P is upper bounded by 4 because each node which implies that M→∞ 1 2 , which W indicates that when α = c2, Formula (7) can also be has at most 4 for uplink traffic (or downlink traffic), satisfied. which equals its S-D pair throughput. By Lemma 4.2 Lemma 4.2: Let n(σ, M) denote the number of mobile and Theorem 4.1, we can derive that c W nodes in the whole network. Then, lim Pr P = 2 lim Pr (n(σ, M) = shM) = 1. (13) M→∞ 3(c1 + 1)sh M→∞ P (σ, M) c W Proof: Let Z1, Z2, ..., ZM be i.i.d. random variables = lim Pr = 2 representing the number of nodes in cell 1, 2,..., M, M→∞ n(σ, M) 3(c1 + 1)sh PM respectively. Then, n(σ, M) = k=1 Zk. Because each c2WM ≥ lim Pr P (σ, M) = Pr (n(σ, M) = shM) Zk follows a Poisson distribution with parameter sh, M→∞ 3(c1 + 1) E(Zk) = sh, ∀1 ≤ k ≤ M. According to SLLN, c WM PM ! = lim Pr P (σ, M) = 2 = 1 k=1 Zk Pr lim = sh = 1, (14) M→∞ 3(c1 + 1) M→∞ M c2W which implies that limM→∞ Pr P = = 1. which implies that lim Pr PM Z = s M = 1, 3(c1+1)sh M→∞ k=1 k h Corollary 4.1 shows that DTR produces a constant and hence limM→∞ Pr (n(σ, M) = shM) = 1. throughput for each pair of nodes regardless of the Theorem 4.1: For a hybrid network of M BSes and σ number of nodes in each cell due to its spacial reuse mobile node density, where each node has the intra-cell, of the system. Theorem 4.1 and Corollary 4.1 show that uplink and downlink sub-channel bandwidth satisfying the aggregate throughput capacity and the throughput down up up int int Wi = Wi = W = W/4,Wi = W = W/2 (15) per S-D pair of DTR are Θ(MW ) and Θ(W ), respec- the aggregate throughput capacity of DTR is tively. The work in [17] proves that DHybrid achieves Θ(MW ) P (σ, M) = Θ(MW ). infrastructure aggregate throughput, and the (16) work in [33] proves that the pure ad-hoc transmission Proof: To prove P (σ, M) = Θ(MW ), we need to Θ( √ W ) prove that there exists deterministic constants α > 0 and achieves n·logn throughput per S-D pair. The results α0 < ∞ such that demonstrate that the throughput rates of DTR and DHy- brid are higher than that of the pure ad-hoc transmission. lim Pr{P (σ, M) = αMW is feasible} = 1 (17) M→∞ This is because the pure ad-hoc transmission is not effi- lim inf Pr{P (σ, M) = α0MW is feasible} < 1. (18) M→∞ cient in a large scale network [35]. A large network size Recall that any two nodes within a cell cannot transmit reduces the path utilization efficiency and increases node simultaneously in the same traffic mode, the throughput interference. Facilitated by the infrastructure network, 9

R: Transmission range of a base station and a mobile node B: Base station m : mobile node D: Distance between B and m Proposition 4.2 indicates that in a high-density network, i i a source node in DTR can find relay nodes for message forwarding with a high probability. For example, h h assume the average number of neighbor nodes of a B B source node is 10. With the daily increasing number of D D mobile devices, such an assumption is realistic. Then, m the probability of not being able to find any node in the i R mi R R P∞ k−1 10ke−10 range of a node is 1 − k=1 k k! ≈ 0.12, which (a) DHybrid (b) DTR is very small. Therefore, in a high-density network, a Fig. 5: The traffic load in DHybrid and DTR. source node can find neighbors for message forwarding DTR and DHybrid avoid long distance transmissions, with a high probability. leading to a higher transmission throughput. We use DHybrid to denote the group of routing pro- Proposition 4.1: Suppose a mobile node needs to al- tocols in hybrid wireless networks that directly combine locate totally U segments with the same length to L the ad-hoc transmission mode and the infrastructure neighboring mobile nodes m1, ..., mL, which has uplink transmission mode [1, 5, 6, 12–18]. up up Proposition 4.3: In a hybrid wireless network, the D- bandwidth W1 , ..., WL , respectively. Let Ui denote the number of segments to be allocate to mi (i = 1, 2, ..., L). Hybrid routing protocol leads to load imbalance among To minimize the average latency of these segments, the the mobile nodes in a cell. U1 UL Proof: Figure 5 (a) shows a cell with a BS and a optimal allocation should satisfy W up = ... = W up . The 1 L m Ul randomly picked mobile node i in the range of the minimized average latency equals PL up . BS. The shaded region represents all possible positions 2 i=1 Wi Proof: Recall that each segment has length l. Then, of the source nodes that choose mi as the relay node l in DHybrid. The total traffic passing through node mi for each mobile node mi it requires up time to transmit Wi th is the sum of the traffic generated by the nodes in the a segment. Therefore, the j segment that mi needs to shaded region. The area of shaded region is (j−1)l 2 transmit has to wait up slots. Hence, the total latency S = sh − πD (0 < D < h), (24) W1 of the segments that mi needs to transmit to its BS equals where D is the distance between the BS and relay node Ui 2 mi and sh is the area size of a cell. Therefore, the X (j − 1)l (0 + 1 + ... + (Ui − 1))l Ui l up = up ≈ up . (21) expected value of traffic passing through node mi is Wi Wi 2Wi j=1 W · σ · (s − πD2) (0 < D < h), (25) Hence, the average latency of transmitting all the mes- h 2 where W is the data transmission rate of a source node, PK Ui l sages should be i=1 up /U. According to Cauchy- σ 2Wi and is the density of the nodes in a region. Equa- Schwarz inequality [34], the average latency is lower tion (25) shows that the traffic passing through node mi bounded decreases as D increases. That is, the nodes closer to the L 2 L 2 L BS have a higher load than the nodes staying at the brim 1 X Ui l l X Ui X up up = up W of the cell. U 2W 2U PL W up W i i=1 i i=1 i i=1 i i=1 Proposition 4.4: In a hybrid wireless network, DTR 2 L s ! achieves more balanced load distribution among the l X U 2 q ≥ i W up mobile nodes in each cell. PL up W up i 2U i=1 Wi i=1 i Proof: The shaded region in Figure 5 (b) represents Ul all possible positions of the source and relay nodes that = L up . (22) choose node m as relay node. Suppose m neighbor 2 P W i i=1 i nodes are chosen as relay nodes, then the expected traffic s W 2 r U2 m ·σ ·πR U2 L passing through node i is m which shows that 1 up up W W L the traffic going through node mi is independent of its √ 1 √ U1 When up = ... = up , or equivalently, W up = W1 WL 1 location relative to its BS. Since every node in the cell 2 UL PL Ui l has an equal probability of generating traffic, the traffic ... = up , the average segment latency i=1 up /U WL 2Wi load is balanced among the nodes in the cell. Ul can achieve the minimum value PL up . 2 i=1 Wi Proposition 4.1 indicates that forwarding segments to the 5 PERFORMANCE EVALUATION nearby nodes with the highest capacity can minimize the This section demonstrates the properties of DTR through average latency of messages in the cell. It also balances simulations on NS-2 [36] in comparison to DHybrid [17], the transmission load of the mobile nodes within a cell. Two-hop [19] and AODV [8]. In DHybrid, a node first uses broadcasting to observe a multi-hop path to its own Proposition 4.2: A source node in DTR can find re- BS and then forwards a message in the ad-hoc transmis- lay nodes for message forwarding with probability sion mode along the path. During the routing process, if ∞ cke−cr P k−1 r , where c = πR2. the transmission rate (i.e., bandwidth) of the next hop to k=1 k k! r the BS is lower than a threshold, rather than forwarding Proof: Let m denote the number of nodes within mi’s Q the message to the neighbor, the node forwards the transmission area and define the indicator variable i by message directly to its BS. The source node will be n 1 mi is the highest capacity node notified if an established path is broken during data Qi = (23) 0 mi is not the highest capacity node transmission. If a source sends a message to the same destination next time, it uses the previously established then, Pr{mi can find relays for message forwarding} path if it is not broken. In the Two-hop protocol, a source ∞ ∞ k −c node selects the better transmission mode between direct X X k − 1 c e r = Pr (Q = 0|m = k) Pr (m = k) = r transmission and relay transmission. If the source node i k k! k=0 k=1 can find a neighbor that has higher bandwidth to the 10

600 DTR DTR 550 DTR Two-hop Two-hop DTR Two-hop 500 DHybrid DHybrid Two-hop DHybrid 9 6 500 450 AODV AODV AODV 8 DHybrid 7 5 400 400 6 AODV 4 300 350 5 (kbps) 4 pair (kbps) pair 200 300 3 3

100 per S-D Throughput 250 2 2

1 (s) pairS-D per Delay Delay Delay per S-D pair (s)

ThrougputS-D pair per 0 200 0 1 20 40 60 803456 100 20 40 60 803456 100 Network size Number of base stations Network size Number of base stations

Fig. 6: Throughput vs. networkFig. 7: Throughput vs. number Fig. 8: Delay vs. network size. Fig. 9: Delay vs. number of B- size (simulation). of BSes. Ses. BS than itself, it transmits the message to the neighbor. routing in mobile ad-hoc networks, DTR does not need Otherwise, it directly transmits the message to the BS. path query and maintenance. Also, it limits the path Unless otherwise specified, the simulated network length to three to avoid problems in long-path transmis- consists of 50 mobile nodes and 4 BSes. In the ad-hoc sion. The throughput of DHybrid and AODV decreases component of the hybrid wireless network, mobile nodes as the number of nodes in the network increases. This is are randomly deployed around the BSes in a field of mainly because when the network size increases, more 1000×1000 square meters. We used the Distributed Coor- beacon messages are generated in the network. Also, the dination Function (DCF) of the IEEE 802.11 as the MAC long transmission path also leads to high transmission layer protocol. The transmission range of the cellular interference. Then, nodes in these methods suffer from interface was set to 250 meters, and the raw physical link intense interference, leading to more transmission failure bandwidth was set to 2Mbits/s. The transmission power and degraded overall throughput. Also, the mobile node of the ad-hoc interface was set to the minimum value increase in the system leads to high network dynamism, required to keep the network connected for most times, resulting in frequent route re-establishments. even when nodes are in motion in the network. Then, The short routing paths in Two-hop reduce conges- the influence of the transmission range on different tion and signal interference, thus enabling better spatial methods’ performance is controlled. Specifically, we set reuse as in DTR. Meanwhile, Two-hop enables nodes the transmission range through the ad-hoc interface to to adaptively switch between direct transmission and 1.5 times of the average distance between neighboring relay transmission. Hence, part of the transmission load nodes, which can be obtained by measuring the simu- is transferred to relay nodes, which carry the messages lated network. We used the two-ray propagation model until meeting the BSes. As a result, gateway nodes con- for the physical layer model. Constant bit rate (CBR) necting mobile nodes and BSes are not easily overloaded. was selected as the traffic mode in the experiment with a Therefore, the throughput of Two-hop is higher than rate of 640kbps. In the experiment, we randomly chose 4 DHybrid. However, since the number of message rout- source nodes to continuously send messages to random- ing hops is confined to one, Two-hop may not find the ly chosen destination nodes. The number of channels for node with the best transmission rate to the BSes because each BS was set to 10. We set the number of redundant of the short transmission range of the ad-hoc interface. routing paths b in Section 3.4 to 1. We assumed that there Therefore, the throughput of Two-hop is lower than was no capacity degradation during transmission be- DTR, especially in a network with high node density. tween BSes. This assumption is realistic considering the The reason that AODV has the lowest throughput per advanced technologies and hardware presently used in S-D pair is its long transmission paths. wired infrastructure networks. There was no message re- Figure 7 shows the throughput per S-D pair versus transmission for failed transmissions in the experiments. the number of BSes in different routing protocols. The We employed the random way-point mobility mod- number of BSes was varied from 3 to 6. The BSes are el [37] to generate the moving direction, speed, and uniformly distributed in the network. We can see from pause duration of each node. In this model, each node the figure that as the number of BSes increases, the moves to a random position with a speed randomly cho- throughputs of DTR, Two-hop, and DHybrid increase sen from (1 − 20)m/s. The pause time of each node was while the throughput of AODV stays nearly constant. set to 0. We set the number of segments of a message to In DTR, Two-hop, and DHybrid, as the number of BSes the connection degree of the source node. The simulation increases, the total number of nodes close to the BSes warmup time was set to 100s and the simulation time increases. Then, more nodes have high transmission was set to 1000s. We conducted the experiments 5 times rates to the BSes, leading to a throughput increase. In and used the average value as the final experimental AODV, since the traffic between S-D pairs does not travel result. To make the methods comparable, we did not though BSes, the throughput between an S-D pair is not use the congestion control algorithm in DTR unless affected by the increased number of BSes in the network. otherwise indicated. The figure also shows that the throughput of DTR is constantly larger than Two-hop and the throughput 5.1 Scalability of Two-hop is constantly larger than DHybrid. AODV Figure 6 shows the average throughput measured in constantly has the lowest transmission delay due to the kbps per S-D pair of different routing protocols versus same reasons as in Figure 6. the number of mobile nodes in the system. The figure shows the throughput of DTR remains almost the same 5.2 Transmission Delay with different network sizes. This result conforms to Figure 8 shows the average transmission delay of S- Corollary 4.1. DTR uses distributed multi-path routing to D pairs for successfully delivered messages in different fully take advantage of the spatial reuse and avoid trans- routing protocols versus network size. The network size mission congestion in a single path. Unlike the multi-hop was varied from 20 to 100 with 20 increase in each step. 11

DTR DTR 120 DHybrid 500 3000 Two-hop Two-hop 450 100 2500 DHybrid AODV 400 350 80 2000 300 60 250 DTR 1500 200 40 1000 150 DHybrid Communication overhead (kpbs) overhead 100 Two-hop thorughput of Total 20 base stations (kbps) 500 50 AODV 0 throughput (kbps) Total 0 0 20 40 60 80 100 0 5 10 15 20 23456 Network size Node moving speed (m/s) Number of source nodes Fig. 10: Overhead vs. network size Fig. 11: Throughput vs. mobility. Fig. 12: Throughput of BSes vs. number of source nodes.

Transmission delay is the amount of time it takes for constant communication delay. The ﬁgure also shows a message to be transmitted from its source node to that the communication delay between S-D pairs follows its destination node. From the ﬁgure, we see that DTR DTR

1 hop 2 hops DTR DTR 600 3 hops 4 hops 800 DHybrid 700 Two-hop 700 600 DHybrid 550 600 Two-hop 500 500 500 400 400 (kbps) 300 300

pair (kbps) 450

200 station (kbps) 200 400 100 Throughput per S-D Throughput per

Throughput of a base a Throughput of 100

Average load per node node per load Average 0 0.2 0.4 0.6 0.8 1 1.2 350 0 20 40 60 80 100 The distance from a mobile node to its base 1234 Network size station Rank of base stations Fig. 13: Throughput vs. number of Fig. 14: Load distribution in a cell. Fig. 15: Load distribution among BSes. hops. the routing between the ad-hoc transmission and cellular Figure 13 shows the average throughput per S-D pair transmission, the throughput of DHybrid is much higher versus network size in DTR. As the figure shows, as than AODV’s. With no infrastructure network, AODV the network size increases, the node throughput keeps produces much lower throughput than the others. Its constant regardless of the number of forwarding hops throughput also drops as node mobility increases for the in a routing. The reason is the same as in Figure 6. We same reasons as DHybrid. can also see from the figure that the throughput of the four protocols follows 3-hop>4-hop>2-hop>1-hop. In 5.5 Effect of Workload the 1-hop routing, each node only transmits segments We measured the total throughput of BSes on the directly to a BS regardless of its current transmission messages received by BSes. Figure 12 shows the total rate. In the 2-hop routing, if the transmission rate of throughput of the BSes versus the number of source a node’s neighbor is higher than that of the node, it nodes. We can see that DTR and Two-hop have much asks its neighbor node to forward the segment to a BS. higher throughput increase rates than DHybrid. This Therefore, the 2-hop routing has higher throughput than is because in DTR and Two-hop, the number of trans- the 1-hop routing. The 3-hop routing can greatly increase mission hops from a source node to a BS is small. the number of node options for segment routing since Meanwhile, each node can adaptively switch between the number of nodes that the source node can encounter relay transmission and direct transmission based on increases from d to d2, where d is the average node the transmission rate of its neighbors. Hence, part of a degree. Thus, a node with a greater transmission rate can source node’s transmission load is transferred to a few be chosen as the forwarding node. Meanwhile, the 3-hop relay nodes, which carry the messages until meeting the routing can greatly facilitate inter-cell communication BSes. Therefore, the gateway mobile nodes are less likely because a node has a higher probability of reaching a to be congested. However, nodes in DHybrid cannot neighboring BS within a 3-hop path length than within adaptively adjust the next forwarding hop because it is a 2-hop path length. Therefore, the throughput of the predetermined in the routing path. Messages are always 3-hop routing is much higher than that of the 2-hop forwarded to the mobile gateway nodes that are closer routing. The figure also shows that the 4-hop routing to the BSes or that have higher transmission rates. There- produces lower throughput than the 3-hop routing. The fore, these mobile gateway nodes can easily become con- reason is that 3 hops are enough to find a hop with high gested as the workload of the system increases, leading transmission rate and achieve inter-cell communication to many message drops. Therefore, when the number because of widespread BSes. The 4-hop routing increases of the source nodes is larger than 4, the throughput the forwarding delay due to the greater number of of DHybrid remains nearly constant. This is also the nodes in a route; thus, it cannot increase the uploading reason that the throughput of DHybrid is constantly transmission rate of messages. lower than those of DTR and Two-hop. Additionally, the figure shows that the overall throughput of Two-hop 5.7 Load Distribution Within a Cell is lower than that of DTR. This is because most of the In this experiment, we tested the load distribution of traffic in Two-hop is confined to a single cell. When a BS mobile nodes in a randomly chosen cell in the hy- in a cell is congested, the traffic cannot be transferred brid wireless network that employs each of the DTR, to other cells. In contrast, DTR’s three-hop distributed DHybrid, and Two-hop protocols. We normalized the distance from a mobile node to its base station according forwarding mechanism enables it to distribute the traffic D to the function , where D is the actual distance and Rb among the BSes in a balance. Therefore, the BSes in Rb DTR will not become congested easily. In addition, as is the radius of its cell. We divided the space of the cell the forwarding mechanism gives nodes more flexibility into several concentric circles and measured the loads of in choosing relay nodes with higher transmission rates the nodes on each circle to show the load distribution. for message forwarding to the BSes, the overall BS Figure 14 shows the average load of a node corre- throughput in DTR is larger than in Two-hop. sponding to the normalized distance from itself to the BS in the chosen cell. The figure shows that most of 5.6 Effect of the Number of Routing Hops the traffic load of DHybrid is located at nodes near We conducted experiments to show the optimal num- the BS. The nodes far from the BS have very low load. ber of routing hops for the routing in hybrid wireless The results conform to Proposition 4.3. In DHybrid, if networks. We tested the throughput per S-D pair for x- a source node wants to access the Internet backbone hop DTR, where x was varied from 1 to 4. In the 1-hop or engage in inter-cell communication, it transmits the routing, a node directly transmits a message to the BS messages to the BSes in a multi-hop fashion. Therefore, without message division. In the other routing protocols, the nodes near the BSes will have the highest load. On the (x−1)th hop chooses the best transmission mode be- the other hand, since there is little traffic going through tween direct transmission and relay transmission. Also, the nodes at the brim of a cell, the load of these nodes in the 4-hop routing, the second relay node randomly is small. As a result, some nodes can easily become chooses the third relay node. hot spots while the resources of other nodes are not 13

600 800 450

700 400 500 Rank 1 350 600 Rank 2 400 300 Rank 1 500 Rank 3 250 Rank 2 300 Rank 1 400 Rank 4 200 300 Rank 3 200 Rank 2 station (kbps)

station (kbps) 150 200 Rank 4 Rank 3 station (kpbs) 100 100 Throughput of a base Throughputa of Throughput of a base Rank 4 100 50 Throughput of a base 0 0 0 0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000 Simulation time (s) Simulation time (s) Simulation time (s) (a) DTR (b) Two-hop (c) DHybrid Fig. 16: Base station load vs. simulation time. fully utilized. This load imbalance prevents DHybrid a congestion control algorithm. In DHybrid, if a previ- from fully utilizing system resources. The traffic load of ously established path to a destination is not broken, DTR is almost evenly distributed in the system, which a node still uses this path to transmit messages to the is in line with Proposition 4.4. In DTR, the traffic from same destination. Thus, the nodes cannot dynamically a source node is distributed among a number of relay balance load between BSes. Also, when a node finds neighbors for further data forwarding. The nodes at that its current BS is congested, it takes a long time for the brim of the cell also take responsibility for message it to find a path to a non-congested BS by re-issuing forwarding, since the neighbor nodes of the brim nodes a query message to the neighboring non-congested BS, could be located in other cells with good transmission which greatly reduces the throughput of the system. channels. In Two-hop, the source node considers direct Figure 16 further shows the throughput of the BSes transmission or one-hop relay transmission based on the versus simulation time in the three routing protocols. At channel condition. Since the node is chosen within one the beginning, the BSes with ranks 1 and 2 are congested hop, the messages will not gather close to the BS due to and those with ranks 3 and 4 do not have much traffic. the limited transmission range. However, because of its Thus, the three figures show that the BSes with ranks 1 sequential transmission, Two-hop cannot achieve load and 2 have high throughput but those with ranks 3 and balance among nodes in a cell as well as DTR. 4 have extremely low throughput at the beginning in all three protocols. Figure 16 (a) shows the throughput of 5.8 Load Balance Between Cells the BSes in DTR. As shown in the figure, since DTR can In this experiment, we tested the effectiveness of the adaptively adjust the traffic among the BSes using its congestion control algorithm in DTR. We also added a congestion control algorithm, the throughput of the two congestion control algorithm to DHybrid. In the algo- highly congested BSes is distributed to the neighboring rithm, when a node receives beacon messages from its BSes. As the traffic is forwarded from the BSes of ranks BS indicating that it is overloaded, the node broadcasts 1 and 2 to the BSes of ranks 3 and 4, the throughputs a query message to find a path to a nearby uncongested of these BSes are very similar later in the simulation. BS. We selected two BSes out of the total four BSes. In This result indicates the effectiveness of the congestion the range of each of the two selected BSes, we randomly control algorithm in DTR for load balance between cells. selected one mobile node as the source node to send Figure 16 (b) shows the throughput of the BSes in messages to a randomly selected destination node in the Two-hop. In Two-hop, since the source nodes cannot network. Once the source node moves out of the range effectively move the traffic between BSes, the BSes with of the selected BS, another mobile node in the range rank 1 and rank 2 constantly have the highest through- was selected as the source node. In order to show the put, while the BSes with rank 3 and rank 4 constantly load distribution of the BSes in different protocols, we have low throughput. The low throughput is produced ranked the BSes based on BS throughput. The BS with when the immediate neighbors of the source node are the highest throughput has a rank of 1. in the range of the neighboring BSes of the source Figure 15 shows the throughput of each BS versus the node’s BS. However, the probability of such cases is very BS rank. We can see from the figure that in Two-hop, small. Figure 16 (c) shows the throughput of the BSes in the throughput of the first two BSes is extremely high DHybrid. As the nodes in DHybrid cannot effectively while the throughput of the last two BSes is extremely balance the load between the BSes, the throughput of small. This is because the two hop routing path length in the BSes of rank 1 and rank 2 is much larger than that Two-hop is not long enough to forward messages from of the BSes of rank 3 and rank 4. Comparing Figure 16 a congested BS to a lightly loaded BS. Therefore, the (b) and Figure 16 (c), we can find that the throughput in traffic cannot be shifted to the neighboring lightly loaded DHybrid is lower than that in Two-hop. This is because BSes, leading to an unbalanced load distribution. We can the multi-hop transmission in the ad-hoc network in also see from the figure that in DTR, the variance of the DHybrid greatly reduces the throughput. Meanwhile, throughputs in different BSes is small. The reason is that the mobile gateway nodes in DHybrid easily become three forwarding hops are enough for a mobile node to congested, leading to more message drops. reach a neighboring BS and hence to balance the load between the BSes. Meanwhile, the congestion control algorithm in DTR can effectively switch the traffic from 6 CONCLUSIONS a highly loaded cell to a lightly loaded cell. Because the Hybrid wireless networks have been receiving BSes of ranks 1 and 2 in DTR are not congested, their increasing attention in recent years. A hybrid wireless throughput is less than the corresponding BSes in Two- network combining an infrastructure wireless network hop; also, the throughput of the BSes of ranks 3 and 4 in and a mobile ad-hoc network leverages their advantages DTR is much higher than that of the corresponding BSes to increase the throughput capacity of the system. How- in Two-hop. DHybrid achieves more balanced load dis- ever, current hybrid wireless networks simply combine tribution between BSes than Two-hop since it employs the routing protocols in the two types of networks for 15

[7] Y. Yao, X. Tang, and E. Lim, “In-network processing of nearest [34] “In how many ways can m balls be distributed into n boxes?” neighbor queries for wireless sensor networks,” in Proc. of DAS- http://www.fen.bilkent.edu.tr/ otekman/disc/usef.pdf. 14 FAA06, 2006. [35] G. Wang, G. Cao, T. L. Porta, and W. Zhang, “Sensor relocation [8] R. Szewczyk, J. Polastre, A. Mainwaring, and D. Culler, “Lessons in mobile sensor networks,” in Proc. of IEEE INFOCOM, 2005. data transmission,from a Sensor Network which Expedition,” prevents inthemProc. from of EWSN achieving, 2004. [18][36] H. L. Hu Y. Hsieh and D. and Evans, R. “Localization Sivakumar. for A hybrid mobile sensor network networks,” model for higher[9] C. system Intanagonwiwat, capacity. R. Govindan, In this and paper, D. Estrin, we “Directed propose Diffu- a inwirelessProc. of packet MobiCom data, 2004. networks. In Proc. of GLOBECOM, 2002. sion: A Scalable and Robust Communication Paradigm for Sensor [19][37] Y. F. Liu, Wei X. and Cheng, D. Gitlin.D. Hua, and Two-hop-relay D. Chen, “Location architecture discovery for for next- DistributedNetworks,” Three-hop in Proc. of Routing Mobicom, 2000. (DTR) data routing pro- sensorgeneration networks WWAN/WLAN with short range integration. beacons,” IJAHUCIEEE Wireless, 2009. Commu- tocol[10] that W. Zhang,integrates G. Cao, the and dual T. L. features Porta, “Data of hybrid Dissemination wireless with [38]nication R. Fonseca,, 2004. S. Ratnasamy, J. Zhao, and C. T. Ee, “Beacon vector networksRing-Based in the Index data for transmissionWireless Sensor Networks,” process. in InProc. DTR, of ICNP a , [20] routing:X. J. Li, scalable B. C. Seet, point-to-point and P. H. J. Chong.routingin Multihop wireless cellular sensornets,” networks: in source2003, node pp. divides 305–314. a message stream into segments and Proc.Technology of NSDI, and 2005. economics. Computer Networks, 2008. [11] S. Madden, M. J. Franklin, and J. M. H. W. Hong, “TAG: a Tiny [21][39] B. N. Bengfort, Bulusu, J. W. Heidemann, Zhang, and and X. Du. D. Efficient Estrin, “Gps-Less resource allocation Low-Cost in transmits them to its mobile neighbors, which further hybrid wireless networks. In Proc. of WCNC, 2011. forwardAGgregation the segments Service for to Ad-Hoc their destination Sensor Networks,” through in Proc. an of Outdoor Localization For Very Small Devices,” IEEE Personal Com- OSDI, 2002. [22] municationsP. Thulasiraman Magazine and, vol. X. Shen. 7, no. Interference 5, pp. 28–34, aware 2000. resource alloca- infrastructure network. DTR limits the routing path [40]tion “The for one hybrid simulator. hierarchical http://www.netlab.tkk.fi/.” wireless networks. Computer Networks, [12] F. Ye and G. Zhong, “GRAdient Broadcast: A Robust Data Deliv- 54(13):2271–2280, 2010. lengthery to Protocol three, for and Large always Scale Sensor arranges Networks,” for high-capacityWINET, 2005. [23][41] K. H. Akkarajitsakul, Shen, T. Li, L. E. Zhao, Hossain, andZ. and Li, D. “SDS: Niyato. Distributed Cooperative Spatial- packet nodes[13] toH. Luo, forward F. Ye, J. Cheng, data. S. Unlike Lu, and L.most Zhang, existing “Ttdd: Two-tier routing data Temporaldelivery inSimilarity hybrid wirelessData Storage mobile in Wirelessnetworks: Sensor A coalitional Networks,” game protocols,dissemination DTR produces in large-scale significantly sensor networks,” lowerWireless overhead Networks, inapproach.Proc. of ICCCNIEEE, Trans. 2009. Mob. Comput., 12(5):840–854, 2013. by eliminatingvol. 11, pp. 161–175, route 2002. discovery and maintenance. In [24] T. Liu, M. Rong, P. Li, D. Yu, Y. Xue, and E. Schulz. Radio resource addition,[14] S. Ratnasamy, its distinguishing B. Karp, S. Shenker, characteristics D. Estrin, R. of Grovindan, short allocation in two-hop cellular relaying network. In Proc. of VTC, L. Yin, and F. Yu, “Data-centric storage in sensornet with ght: A 2006. path length,geographic short-distance hash table,” in Proc. transmission, of MONET, 2003. and balanced [25] T. Liu, M. Rong, H. Shi, D. Yu, Y. Xue, and E. Schulz. Reuse load[15] distribution S. Ratnasamy, provideB. Karp, S. highShenker, routing D. Estrin, reliability and L. Yin, “Data- and partitioning in fixed two-hop cellular relaying network. In Proc. efficiency.centric DTR storage also in sensornets has a congestion with GHT, a control geographic algorithm hash table,” of WCNC, 2006. [26] L. Guan, J. Zhang,Haiying J. Li, G. Shen Liu,received and P. Zhang. the BS degree Spectral in efficientCom- to avoidMONET load, vol. 8,congestion pp. 427–442, 2003. in BSes in the case of frequency allocation scheme in multihop cellular network. In unbalanced[16] X. Li, Y. traffic J. Kim, distributions and W. Hong, “Multi-dimensional in networks. Theoretical range queries puter Science and Engineering from Tongji Uni- Proc. of VTC, 2007.versity, China in 2000, and the MS and Ph.D. in sensor networks,” in Proc. of SenSys, 2003. [27] D. M. Shila, Y. Cheng, and T. Anjali. Throughput and delay analysis and simulation results show that DTR can degrees in Computer Engineering from Wayne dramatically[17] D. Ganesan, improve “DIMENSIONS: the throughput Why do wecapacity need a new and data analysis of hybrid wireless networks with multi-hop uplinks. In handling architecture for sensor networks,” in Proc. of the ACM Proc. of INFOCOMState, 2011. University in 2004 and 2006, respectively. scalabilityHotNets of, 2002, hybrid pp. 143–148. wireless networks due to its high [28] B. Liu, P. Thiran,She and is D. currently Towsley. an Capacity Assistant of Professor a wireless in ad the hoc scalability,[18] D. Ganesan, efficiency, A. Cerpa, and reliability Y. Yu, D. Estrin, and W. low Ye, overhead. and J. Zhao, network with infrastructure.Department InofProc. Electrical of Mobihoc and Computer, 2007. Engi- “Networking issues in wireless sensor networks,” JPDC, 2004. [29] C. Wang, X. Li, C.neering, Jiang, S. and Tang, the and Director Y. Liu. of Multicastthe Pervasive throughput Com- ACKNOWLEDGEMENTS[19] B. Greenstein, D. Estrin, R. Govindan, S. Ratnasamy, and for hybrid wirelessmunications networks Laboratory under gaussian of Clemson channel University. model. S. Shenker, “Difs: A distributed index for features in sensor net- TMC, 10(6):839–852,Her 2011. research interests include distributed com- This researchworks,” in wasProc. supported of SNPA, 2003. in part by U.S. NSF grants [30] A. G. Valko. Cellularputer ip: systems A new approach and computer to internet networks, host with mobility. an IIS-1354123,[20] J. Li, J. Jannotti, CNS-1254006, D. S. J. De,CNS-1249603, C. David, R. Karger, CNS-1049947, and R. Morris, “A emphasisACM on Computer peer-to-peer Comminication and content, 1999. delivery networks, mobile com- [31] C. Sarr, C. Chaudet, G. Chelius, and I. G. Lassous. A Node- CNS-0917056scalable location and CNS-1025652, service for geographic Microsoft ad hoc routing,” Research in Proc. Fac- of puting, wireless sensor networks, and grid computing. Her research workBased has been Available published Bandwidth in top journals Evaluation and In conferences IEEE 802.11 in Ad these Hoc ulty FellowshipMobiCom, 2000. 8300751, and the United States Depart- Networks. IJPEDS, 00(00):1–21, 2005. [21] J. Newsome and D. Song, “GEM: Graph EMbedding for routing [32]areas. X. She P. Costa, was M. the T. Program Moreno, Co-Chair and H. Hartenstein. for a number A of simulation international study ment of Defense 238866. We would like to thank Mr. conferences and member of the Program Committees of many leading Kang Chenand data-centric for his help storage in in addressing sensor networks review without comments. geographic on the performance of hierarchical mobile ipv6. In Proc. of ITC, information,” in Proc. of SenSys, 2003. conferences.2003. She is a member of the IEEE and ACM. She is Microsoft [22] A. Caruso, S. Chessa, S. De, and R. Urpi, “GPS free coordinate [33]Research P. Gupta Faculty and Fellow P. R. of Kumar. 2010. The capacity of wireless networks. assignment and routing in wireless sensor networks,” in IEEE TIT, 2000. REFERENCES Proc. of [34] L. B. Koralov and Y. G. Sinai. Theory of probability and random [1] H Luo,IEEE R. INFOCOM Ramjee,, P. 2005, Sinha, pp. L. 150–160. Li, and S. Lu. Ucan: A unified cell processes. Berlin New York Springer, 2007. [23]and P. ad-hoc Desnoyers, network D. architecture. Ganesan, and In P.Proc. Shenoy, of MOBICOM “Tsar: A, 2003. two tier [35] H. Y. Hsieh and R. Sivakumar. Performance comparison of [2] P. K.sensor McKinley, storage H. architecture Xu, A. H. Esfahanian, using interval and skip L. M. graphs,” Ni. Unicast- in Proc. cellular and multi-hop wireless networks: A quantitative study. basedof SenSys05 multicast. ACM communication Press, 2005, in pp. wormhole-routed 39–50. direct net- In Proc. of SIGMETRIC, 2001. [24] C. T. Ee and S. Ratnasamy, “Practical data-centric storage,” in Proc. works. TPDS, 1992. [36] The network simulatorLianyu - Zhao ns-2.received http://www.isi.edu/nsnam/ns/. the BS and MS degrees [3] H.of Wu, NSDI C., Qiao, 2006. S. De, and O. Tonguz. Integrated cell and ad hoc [37] M. Grossglauser and D. Tse. Mobility increases the capacity of relaying systems: iCAR. J-SAC, 2001. in Computer Science from Jilin University, China. [25] M. Aly, K. Pruhs, and P. K. Chrysanthis, “KDDCS: A load- ad hoc wireless networks.He is currently In Proc. a Ph.D. of TON student, 2002. in the Depart- [4] Y. H.balanced Tam, H. in-network S. Hassanein, data-centric S. G. Akl, storage and R. scheme Benkoczi. in sensor Optimal net- multi-hop cellular architecture for wireless communications. In mentHaiying of Electrical Shen received and Computer the BS degree Engineering in Com- work,” in Proc. of CIKM, 2006, pp. 317–326. ofputer Clemson Science University. and Engineering His research from interests Tongji U- [26]Proc. F. of Bian, LCN X., 2006. Li, R. Govindan, and S. Schenker, “Using hierarchical [5] Y. D. Lin and Y. C. Hsu. Multi-hop cellular: A new architecture includeniversity, wireless China sensor in 2000, network, and the routing MS and proto- Ph.D. location names for scalable routing and rendezvous in wireless degrees in Computer Engineering from Wayne for wireless communications. In Proc. of INFOCOM, 2000. cols, applications and security issues in P2P sensor networks,” in Proc. of SenSys, 2004, pp. 305–306. State University in 2004 and 2006, respectively. [6] P. T. Oliver, Dousse, and M. Hasler. Connectivity in ad hoc and networks.She is currently an Associate Professor in the [27]hybrid J. Xu, networks. X. Tang, In andProc. W. of chien INFOCOM Lee, “A, 2002. new storage scheme for approximate location queries in object tracking sensor networks,” Department of Electrical and Computer Engi- [7] E. P. Charles and P. Bhagwat. Highly dynamic destination neering at Clemson University. Her research sequencedIEEE TPDS distance, vol. 19, vector pp. 262–275, routing (DSDV) 2008. for mobile computers. interests include distributed computer systems [28]In Proc. M. Li ofSIGCOMM and Y. Liu,, 1994. “Rendered path: range-free localization in and computer networks, with an emphasis on [8] C.anisotropic Perkins, E. sensor Belding-Royer, networks with and holes,” S. Das. in Proc. RFC of 3561: MobiCom Ad, 2007. hoc content delivery networks, mobile computing, [29]on demand H. Shen, T. distance Li, and T. vector Schweiger, (AODV) “An routing. Efficient Similarity Technical Searching report, wireless sensor networks, and cloud. She is InternetScheme Engineering Based on Locality Task Force, Sensitive 2003. Hashing,” in Proc. of ICDT, a Microsoft Faculty Fellow of 2010, a senior [9] D.2008. B. Johnson and D. A. Maltz. Dynamic source routing in ad member of the IEEE and a member of the ACM. [30]hoc D. wireless Karger, networks. E. Lehman,IEEE T. Mobile Leighton, Computing M. Levine,, 1996. D. Lewin, and Ze Lireceivedreceived the the BS degree BS degree in Electronics in Electronics and [10] V. D. Park and M. Scott Corson. A highly adaptive distributed Ze Li R. Panigrahy, “Consistent Hashing and Random Trees: Distributed Informationand Information Engineering Engineering from Huazhong from Huazhong Univer- U- routingCaching algorithm Protocols for for mobile Relieving wireless Hot Spots networks. on theWorld In Proc. Wide of niversity of Science and Technology, China, in INFOCOM, 1997. sity of Science and Technology, China, in 2007. Web,” in Proc. of STOC, 1997, pp. 654–663. 2007 and the Ph.D. degree in the Department of [11] R. S. Chang, W. Y. Chen, and Y. F. Wen. Hybrid wireless network HeElectrical is currently and aComputer Ph.D. student Engineering in the Depart- of Clem- [31]protocols. W. Nejdl,IEEE W. Transaction Siberski, M. on Wolpers, Vehicular and Technology C. Schmnitz,, 2003. “Routing mentson University, of Electrical in and 2012. Computer His research Engineering interests [12] G.and N. Aggelou clustering and in R. schema-based Tafazolli. On super the relaying peer networks,” capacity in ofProc. next- of ofinclude Clemson distributed University. networks, His research with an interests emphasis generationIPTPS, 2003. gsm cellular networks. IEEE Personal Communications includeon content distributed delivery networks, networks, with wireless an emphasis multi-hop [32]Magazine P. A. Bernstein,, 2001. F. Giunchiglia, A. Kementsietsidis, J. Mylopoulos, oncellular peer-to-peer networks, and game content theory delivery and data networks, mining. [13] T.L. Rouse, Serafini, I. Band, and I. andZaihrayeu, S. McLaughlin. “Data management Capacity for and peer-to-peer power wireless multi-hop cellular networks, game the- investigationcomputing: ofA vision,” opportunity in Proc. driven of WebDB multiple, 2002. access (ODMA) ory and data mining. He is a student member of [33]networks A. Y. Halevy, in TDD-CDMA Z. G. Ives, based P. Mork, systems. and I. In Tatarinov,Proc. ofICC “Piazza:, 2002. Data IEEE.Chenxi Qiu received the BS degree in Telecom- [14] H.management Y. Hsieh and infrastructure R. Sivakumar. for semantic On Using web the applications,” Ad-hoc Network in Proc. munication Engineering from Xidian University, Modelof WWW in Wireless, 2003. Packet Data Networks. In Proc. of MOBIHOC, China, in 2009. He currently is a Ph.D. student 2002. in the Department of Electrical and Computer [15] L. M. Feeney, B. Cetin, D. Hollos, M. Kubisch, S. Mengesha, and Engineering at Clemson University, SC, United H. Karl. Multi-rate relaying for performance improvement in ieee States. His research interests include sensor 802.11 wlans. In Proc. of WWIC, 2007. networks and wireless networks. [16] J. Cho and Z. J. Haas. On the throughput enhancement of the downstream channel in cellular radio networks through multihop relaying. IEEE JSAC, 2004. [17] B. Liu, Z. Liu, and D. Towsley. On the capacity of hybrid wireless networks. In Proc. of INFOCOM, 2003.