1 The organization isbased on the operatingfunctions, so multipleroles, such as the terminaland relay point, may beimplemented on the samephysical device.
2 Project open802.11s;http://www.open80211s.org/
AC C E P T E D F R O M OP E N CALL
YING-DAR LIN AND SHIAO-LI TSAO, SHUN-LEE CHANG, SHAU-YU CHENG, AND CHIA-YU KU,NATIONAL CHIAO TUNG UNIVERSITY
DESIGN ISSUES AND EXPERIMENTAL STUDIES OFWIRELESS LAN MESH
INTRODUCTIONInfrastructure-based wireless networks provideconvenient access to the Internet, and arebecoming increasingly popular in spite of theircostly wired deployment. On the other hand,mobile ad hoc networks (MANETs) eliminatethe need for infrastructure, decreasing deploy-ment time and alleviating network constructioncosts. However, having routing functions on allnodes in a MANET complicates the design ofnetworking devices [1]. The fact that MANETusage is typically limited to military and special-ized civilian applications also hinders its growth[2]. By combining an infrastructure-based wire-less network and a MANET, a wireless meshnetwork (WMN) presents a low-cost and fast-deployment solution compared to an infra-
structure-based wireless network, and a reliableand less complicated solution compared to aMANET. A WMN is similar to a multihop cellu-lar network (MCN) [3], which has been provedto improve aggregated throughput linearly due tospatial division.
In [1] the authors classify the WMN architec-ture into three types: infrastructure, client, andhybrid. An infrastructure WMN is organized asa hierarchical network, functionally consisting ofmesh gateways, relay points, access points, andterminals.1 A mesh gateway is a device capableof bridging the wireless mesh and wired infra-structure. A relay point implements a routingalgorithm to relay packets in a mesh. To sup-port non-mesh terminals, the mesh uses accesspoints to bridge the WMN and non-mesh termi-nals. In a client mesh there is no gateway andnon-mesh terminal because this kind of meshemphasizes flat peer-to-peer communications. Ahybrid mesh includes both infrastructure andmesh terminals that provide interfaces for endusers and mesh routing capability. Figure 1agives an example of a hybrid WMN, while Fig.1b shows the corresponding wireless local areanetwork (WLAN) mesh using IEEE 802.11s ter-minology.
Diverse mesh architectures result in varioususage scenarios [1, 2], and a considerable num-ber of challenges for designing and realizing aWMN [1, 4–6]. Industrial organizations havealso prepared standards and recommended prac-tices for existing wireless technologies, such asIEEE 802.15.5 for low-rate wireless personalarea networks (WPANs). Among these efforts,IEEE 802.11s [7], which defines a WLAN meshusing IEEE 802.11 medium access control(MAC) and physical (PHY) layers, is one of themost active standards and has increasing com-mercial opportunities.
This study is the first publicly reported workthat exploits the system design issues of an IEEE802.11s mesh system. Prior studies, such as themesh on XO-laptop for One Laptop per Child(OLPC) [8] and the open80211s project forLinux,2 evaluated the network performance ofthe IEEE 802.11s mesh. However, no studiesexamine system architectures and mesh stability.The system proposed in this study is developed
ABSTRACTWireless mesh networking, as a low-cost and
reliable technology for rapid network deploy-ment, has attracted considerable attention fromacademia and standardization in the industry.The IEEE 802.11s standard defines a wirelessLAN mesh using the IEEE 802.11 mediumaccess control and physical layers, and is one ofthe most active standards with increasing com-mercial opportunities. This study presents thedesign and development of a WLAN mesh sys-tem conforming to the latest IEEE 802.11s draftamendment. Without costly hardware modifica-tions, the proposed solution is a pure softwareextension for commercial off-the-shelf WLANchipsets. This study constructs an experimentaltestbed, and evaluates issues such as the trans-mission reliability of mesh broadcast-type con-trol messages and multichannel transmissions.Experimental results demonstrate that the deliv-ery of mesh broadcast-type control messages,such as routing construction frames, using themultiple acknowledged unicast scheme improvesmesh stability from an 86 to a 98 percent successratio in a 16-node grid. Transmitting packetsusing a single radio interface switching betweenmultiple channels reduces inter-flow interferenceand doubles the throughput in our testbed.
The authors presentthe design anddevelopment of aWLAN mesh systemconforming to thelatest IEEE 802.11sdraft amendment.
LIN LAYOUT 4/8/10 12:55 PM Page 32
Authorized licensed use limited to: National Chiao Tung University. Downloaded on June 28,2010 at 04:09:10 UTC from IEEE Xplore. Restrictions apply.
IEEE Wireless Communications • April 2010 33
based on the latest draft amendment of IEEE802.11s, and extends the functions of commercialoff-the-shelf WLAN chipsets. The proposedextension is a pure software solution that inte-grates with a WLAN driver. Considering theportability of the proposed solution and the sta-bility of the software during the system develop-ment phase, this study proposes a modularizeddesign that separates mesh functions into a driv-er and a user-space program.
Using this system prototype, this study estab-lishes a real testbed and further investigatesdesign and implementation issues that influencesystem performance. Results show that transmit-ting mesh broadcast-type control frames over amultihop wireless mesh without acknowledgmentcan cause network stability problems. In a 16-node interference-intensive grid topology, forexample, 14 percent of path request messagesare lost, and not all routing paths are found evenwhen broadcasting at the most reliable data rate(i.e., 1 Mb/s). Therefore, this study evaluates thefeasibility of replacing unacknowledged broad-cast with multiple acknowledged unicasts formesh broadcast-type control messages. Anotherissue is the serious interference between meshnodes sharing the same channel. To reduce thisinterference, the study also examines the effec-tiveness of transmitting packets using a singleradio interface switching between multiple chan-nels.
The rest of this article is organized as follows.The next section briefly presents the key featuresof IEEE 802.11s. We then discuss the design andimplementation issues of a WLAN mesh. Wethen present the experimental results, and thefinal section offers the conclusion.
IEEE 802.11SThis section overviews IEEE 802.11s. The firstpart presents the architecture of an IEEE802.11s mesh network, and the second partdescribes mesh functions, including the networklink construction, routing algorithm, data deliv-ery, and flooding control.
NETWORK ARCHITECTUREIEEE 802.11s [7] defines an IEEE 802.11-basedWMN that supports broadcast and unicast deliv-ery over a self-configured multihop link-layertopology. As Fig. 1b shows, an IEEE 802.11smesh network contains three types of nodes: themesh point (MP), mesh access point (MAP), andmesh portal (MPP). The MP is the basic meshunit that provides topology construction, routing,and data forwarding. This type of node can alsobe designed as a terminal device for end users todirectly connect with peer MPs and access themesh. Non-mesh IEEE 802.11 stations (STAs)must first associate with a MAP, which is an MPcapable of IEEE 802.11 access point (AP) func-tions, before accessing a mesh. An MPP is anMP integrated with gateway functions to inter-operate with external IEEE 802 LANs. The MP,MAP, and MPP are all logical components, andsome of them can be physically collocated.
MESH FUNCTIONSNetwork Link Construction — The IEEE 802.11sstandard specifies the boot sequence procedurefor an MP joining a mesh network based on theprocedure for an STA associating with an AP ina conventional non-mesh WLAN. First, an MPperforms an active or passive scan to obtain a
Figure 1. A logical view of a wireless mesh network and its mapping to IEEE 802.11s: a) a generic mesh network; b) an IEEE 802.11smesh network.
MPP: Mesh portalMP: Mesh pointMAP: Mesh access pointSTA: Station
Wirednetworks
Meshgateways
Meshbackbone
Mesh accesspoints
Terminaldevices
MPP
MAP MAP
STA
MPP
MPMP
MP
MP
MP
MP
LAN
Gateway Gateway
Relay point
Relay point
Accesspoint
Accesspoint
Non-meshterminal
Wired linkWireless link
(a) (b)
Meshterminal
Relay point Relay point
Relay point
LAN IEEE 802 LAN
IEEE 802.11smesh
Non-meshIEEE 802.11
IEEE 802 LAN
Wired link toexternal networksIEEE 802.11smesh link Non-mesh IEEE 802.11 wireless link
LIN LAYOUT 4/8/10 12:55 PM Page 33
Authorized licensed use limited to: National Chiao Tung University. Downloaded on June 28,2010 at 04:09:10 UTC from IEEE Xplore. Restrictions apply.
IEEE Wireless Communications • April 201034
list of existing MPs in each channel. Next, theMP uses the mesh peer link management protocolto associate with an MP matching its own prefer-ences. This protocol behaves like the conven-tional IEEE 802.11 association procedure. Theonly difference is that the mesh peer link man-agement protocol uses several new managementframes. These frames encapsulate mesh-specificinformation, such as a mesh identifier, whileremoving unrelated parameters, such as a serviceset identifier (SSID). After a successful associa-tion with a neighboring MP, the MP becomes amember of an IEEE 802.11s mesh network.
IEEE 802.11s introduces the unified channelgraph (UCG), which presents one mesh sharingthe same preferences in the same channel, tohandle several meshes spanning different chan-nels. When forming a new mesh, the initial MPrandomly decides a channel precedence valueand embeds that value in the managementframes. After the channel scan, other MPs selectthe channel with the highest precedence as theiroperating channels. This procedure forms aUCG called the simple channel unification proto-col. To resolve multiple UCGs in different chan-nels due to spatial division and the needs ofchannel switching, such as radar detection, IEEE802.11s proposes a channel graph switch proto-col. In this protocol an MP sets a waiting timerand broadcasts a mesh channel switch announce-ment. The announcement contains the waitingperiod and a precedence value of the candidatechannel. When the timer expires, the MPs receiv-ing the announcement switch to a candidatechannel with the highest precedence.
Routing — IEEE 802.11s defines a path selectionframework that flexibly allows vendors to imple-ment proprietary routing metrics and protocolsto meet special needs. The communicationbetween two MPs begins with the constructionof a routing path. Then the data frames aretransmitted along with the routing path vianeighboring MPs.
Hybrid Wireless Mesh Protocol (HWMP) isthe mandatory routing protocol recommendedby IEEE 802.11s. This protocol comprises an on-demand routing procedure to construct a pathbetween two arbitrary nodes and a proactiveextension to speed up the initial connection. Theon-demand routing procedure is derived fromAd Hoc On-Demand Vector (AODV) [9], but itworks on the link layer and adopts a radio-awaremetric called the airtime link metric. This metricconsiders the actual transmission quality interms of transmission error and data rate. Toconstruct a routing path from one node to anoth-er (e.g., from node s to node t), s broadcasts apath request message (PREQ) into the mesh.Upon receiving the PREQ, t responds to s with apath response message (PREP) through unicasttransmission. The reverse routing path from t tos is first built after t receives the PREQ, and thedesired routing path from s to t is created after sreceives the PREP. Once established, this pathcan be used before it times out.
The proactive extension of HWMP activelyforms a routing tree rooted at an MP by periodi-cally broadcasting root announcement messages.Under the HWMP, a unicast packet without a
valid on-demand routing path can be transmittedto the root first. The root then forwards thepacket along the tree to its destination. Mean-while, an on-demand routing procedure is per-formed between the source and destination tocreate a direct routing path. This help from theroot may reduce latency before the direct on-demand routing path is established [10].
Interworking — IEEE 802.11s also defines aninterworking procedure to handle the communi-cation between two terminals in which at leastone is bridged by an MPP. A terminal bridgedby the MPP is called a proxied entity. IEEE802.11s assumes that an MPP can learn all prox-ied entities it bridges. In addition, every meshnode has a proxy table that maintains the rela-tion between a proxied entity and its MPP.
To transmit a packet from a mesh node s to aproxied entity t, the interworking procedure on sfirst performs the HWMP to issue a PREQ torequest the path to t. If no MP responds to thequery (e.g., the proxied entity has not beenlearned), the packet is forwarded to one or moreMPPs and then bridged to their attached exter-nal networks. Otherwise, the corresponding MP,b, replies to s with the PREP. Then s inserts therelation, b bridges t, into its proxy table. Oncethe relation exists, the packet can be deliveredfrom s to b, and then b bridges the packet to t.
Data Frame Format — IEEE 802.11s introduces amesh header subfield in the beginning of theframe body to address multihop transmissions.When convoying packets whose source and desti-nation are both inside the mesh, the subfield indi-cates that the 4-address format in the frameheader is used. The frame header includes the MPaddresses of the next-hop receiver, transmitter,destination, and source, and is processed by MPsas it would be in a wireless distribution system(WDS). Otherwise, the subfield contains two addi-tional addresses to encapsulate the addresses ofproxied entities. When a packet enters or leaves amesh, the sender and receiver addresses areenveloped into or recovered from the subfield.
Flooding Control — The fact that an MP blindlyretransmits broadcast frames to its neighborsmay cause endless flooding due to the loopstructure of a mesh topology. To avoid infiniterebroadcasting, the source MP first tags a 32-bitincremental sequence number, called the MeshSequence Number field, on each frame beforetransmitting it. Other MPs can use this field andthe source MP address as a unique signature toavoid duplication. The source MP also transmitsthe Mesh Time to Live field, a counter decreasedper hop with each frame to limit its longevity.This field acts as a backup mechanism to detectduplication for rollover sequence numbers andlimited recording space.
DESIGN AND IMPLEMENTATION ISSUES OFAN IEEE 802.11S MESH
This section first presents the software architec-ture that considers both portability of the pro-posed solution and system stability for future
When forming a newmesh, the initial MPrandomly decides achannel precedencevalue and embedsthat value in themanagementframes. After thechannel scan, otherMPs select the channel with thehighest precedenceas their operatingchannels.
LIN LAYOUT 4/8/10 12:55 PM Page 34
Authorized licensed use limited to: National Chiao Tung University. Downloaded on June 28,2010 at 04:09:10 UTC from IEEE Xplore. Restrictions apply.
IEEE Wireless Communications • April 2010 35
extensions. Second, this section discusses thedesign and implementation issues including thetransmission reliability of mesh broadcast-typecontrol frames and multichannel transmissions.
SOFTWARE ARCHITECTURETo improve the portability of the IEEE 802.11ssoftware package, a modularized design isrequired. The proposed design separates plat-form-independent functions such as HWMP rou-tines from the kernel and implements them as aLinux daemon program, called a path selectiondaemon. This approach simplifies the develop-ment of HWMP algorithms and provides greaterflexibility in changing routing algorithms. Theuser-space daemon also improves system stabili-ty since it reduces the chance of kernel crashduring the system development stage.
Time-critical functions are implemented inthe kernel and hooked in the IEEE 802.11 driv-er. The boot sequence module performs the bootsequence procedures to associate the device withan IEEE 802.11s network. The data forwardingmodule relays multihop data and triggers thepath selection daemon to construct a routingpath when necessary. Last, the action frame han-dler module translates the MAC-layer controlframes into module-specific commands and acti-vates other modules to process these commands.Figure 2 illustrates the proposed software archi-tecture for the IEEE 802.11s nodes. Only MPPshave the IEEE 802.3 part, and only MPPs andMAPs have the bridge part.
The proposed design handles the control anddata planes separately. Figure 2 shows that whena control frame such as a routing message isreceived, it is passed to the action frame handlermodule to check if the frame comes from anassociated neighbor. If the frame is from a vali-dated neighbor, it is then forwarded to the user-space path selection daemon. The daemonupdates the routing tables, including the pathselection table and proxy table in the kernelspace, and issues the corresponding controlframes to its neighbors via the action frame han-dler module and transmission (TX) handler ifnecessary.
For the data plane, the data forwarding mod-ule either dispatches a received data frame tothe upper layers of the protocol stack while thenode is the destination, or relays it to the nexthop. To relay a data frame, the data forwardingmodule updates the next-hop MAC addresses ofthe frame by referencing the path selectiontable, and then the proxy table if the destinationis not inside the mesh. The TX handler thentransmits the data frame. If not found in bothtables, the frame is forwarded to the root andinvokes on-demand routing.
TRANSMISSION STRATEGIES FOR MESHBROADCAST-TYPE CONTROL FRAMES
In the wireless environment, a sender has diffi-culty detecting a collision on the receiver side.This results in the need for acknowledgment(ACK) and retransmission mechanisms to reducethe packet loss side-effect. However, the IEEE802.11 broadcasting scheme, called conventionalbroadcasting in this article, has no ACK. As a
result, conventional broadcasting is unreliableand packet loss is high in an interference-proneenvironment. Packet loss is cumulative in multi-hop transmissions, which brings a new challengeto the wireless mesh since some of the IEEE802.11s control frames are broadcast-type. Thus,the frequent loss of these broadcast-type controlframes may result in unstable mesh topology andnetwork.
To improve the transmission reliability ofbroadcast-type control frames, possible IEEE802.11-compatible solutions either blindly broad-cast the same frame multiple times, called themultiple-broadcast scheme, or unicast it to eachneighbor individually,3 called the multiple-uni-cast scheme. Both schemes reduce packet loss atthe expense of using more wireless resources.The multiple-broadcast scheme might use lesswireless resources than the second schemebecause it does not require ACK frames, certaininter-frame spaces (IFSs), and retransmissions.However, it must use more robust coding andmodulation (i.e., a lower transmission rate) [12]so that the broadcast frames are more likely tobe received by all neighbors. On the other hand,the multiple-unicast scheme is able to adopt ahigher transmission rate for each individualneighbor. The ACK frames used by thisapproach also improve transmission reliability.
The proposed design handles the mesh broad-cast-type control frames and other broadcast-type frames separately. The mesh broadcast-typecontrol frames (e.g., the routing request frames)are directly generated by the mesh protocols.The other broadcast-type frames come fromupper layers of the protocol stack (e.g., DynamicHost Configuration Protocol [DHCP] discovery).The latter are normally less important, and mayinherit different characteristics from applicationsand services. Therefore, this study leaves theresponse to the loss of these kinds of broadcast-type frames to upper-layer protocols and appliesthe conventional broadcasting approach.
The transmission reliability of mesh broad-cast-type control frames significantly influencesthe construction and stability of a mesh network.In addition, the number of broadcast-type con-trol frames depends on the scale of the mesh,and is relatively small compared to that of broad-cast-type data frames. Therefore, this studyapplies and evaluates the multiple-broadcast andmultiple-unicast schemes only for mesh broad-cast-type control frames.
MULTICHANNEL WITH FEWER RADIOSIn a single-channel mesh network, nodes sharingthe same channel may experience intra-flowinterference in multihop transmissions and inter-flow interference in nearby data flows. A fre-quency multiplexing approach, which usesmultiple radios to carry packets in separatechannels, can be used to avoid this interference.However, low-cost devices are usually equippedwith a limited number of radio interfaces, somultiple-radio schemes might not be adopted.Therefore, the mesh system in this study appliesa low-cost solution that uses fewer radios switch-ing between multiple channels for data transmis-sion.
Several channel switching problems must be
The daemon updatesthe routing tables,including the path
selection table andproxy table in thekernel space, and
issues the corresponding control
frames to its neighbors via the
action frame handlermodule and
transmission handlerif necessary.
LIN LAYOUT 4/8/10 12:55 PM Page 35
Authorized licensed use limited to: National Chiao Tung University. Downloaded on June 28,2010 at 04:09:10 UTC from IEEE Xplore. Restrictions apply.
IEEE Wireless Communications • April 201036
solved. The first problem is that switchingbetween channels without coordination canresult in packet loss. To avoid this, a notificationmechanism that notifies neighbors to buffer databefore channel switching is necessary. Second,frequent switches cause a heavy switch overhead,while infrequent switches may result in pro-longed latency and buffer overflow. Hence, apolicy that manipulates radios to switch betweenmultiple channels at suitable switch intervalsmust be decided.
Notification mechanisms compatible with theWLAN standards can be employed to avoidpacket loss and link disruption caused by chan-nel switching. Compatible mechanisms include,but are not limited to, the power-saving mecha-nism and the channel switch announcement inIEEE 802.11v. For example, an MP can use thepower-saving mechanism to announce its sleepmode. When the MP enters sleep mode, itsneighbors must buffer data for that MP. Duringthe sleep period, the MP can switch to differentchannels for data transmissions and receptions.
In the proposed approach, with fewer radiosand multiple channels, a mesh node where dataflows merge, such as a gateway, could performchannel switching to serve separate data flows.The separation can reduce the inter-flow inter-ference and improve the throughput. The switchnode first uses a notification mechanism toannounce its channel departure or return. Whilethe node switches to a channel, how long thenode stays in the channel should be determined.A suitable staying period is the time sufficientfor transmitting queued packets on each chan-nel. Therefore, it is important to know the chan-nel resources each channel requires fortransmitting queued packets. The desired chan-
nel resources are derived from the number ofpackets to exchange between mesh nodes, thechannel quality (i.e., transmission rate) of a link,the average packet size, and so on. An unsuit-able staying period might produce a high switch-ing overhead or buffer overflow that degradesthe throughput. Thus, the length of staying peri-od should be investigated.
EXPERIMENTAL RESULTS
This study conducts a number of experiments toevaluate the proposed approaches for the designand implementation issues presented in the pre-vious section. This section first describes theexperimental testbed, and then the experimentalresults.
EXPERIMENTAL TESTBEDThis study implemented a WLAN mesh systemon the Realtek RTL8186 platform, which is acommercial system-on-a-chip embedded with anEthernet, a single-radio 802.11b/g controller, anda 180 MHz 32-bit MIPS processor. The platformruns an embedded Linux (v. 2.4.18) and inte-grates with an open source link-layer bridgemodule (called Ethernet Bridge) that bridges theEthernet and IEEE 802.11 networks. By combin-ing the Ethernet, bridge, and WLAN functions,this platform supports MP, MAP, and MPPfunctions. Figure 3a shows that the integratedtestbed is powered by a battery, where a fixedattenuator is attached to the antenna to regulateits transmission power. This greatly reduces theactual space required for the experimentaltestbed, and makes it possible to conduct small-scale mesh experiments in the laboratory.4 Thetraffic generator injects the desired experimental
Figure 2. The proposed software architecture and the control/data plane flows. The modules in bold text arethe IEEE 802.11s extensions.
RX ISR TX handler
IEEE 802.3 NICdriver
IEEE 802.3 NIC
Flow of IEEE 802.11scontrol-plane
IEEE 802.11 NIC
Action framehandler
Bootsequencemodule
Data forwardingmodule
IEEE 802.11sNIC driver
Path selectiontable
Bridge
User space
Kernel space
AppPath selectiondaemon
Flow of IEEE802.11s data-plane Other flows
Proxy table
The transmission reliability of meshbroadcast-type control frames significantly influencesthe construction andstability of a meshnetwork. In addition,the number of broadcast-type controlframes depends onthe scale of themesh, and is relativelysmall compared withthat of broadcast-type data frames.
3 Those two strategiesappeared in the early ver-sion of IEEE 802.11s.However, the latest draftremoves the part andstates that the problem isbeyond their scope.Besides, some literature,such as [11], also pro-vides solutions that needto extend the IEEE802.11 control frames.
LIN LAYOUT 4/8/10 12:55 PM Page 36
Authorized licensed use limited to: National Chiao Tung University. Downloaded on June 28,2010 at 04:09:10 UTC from IEEE Xplore. Restrictions apply.
IEEE Wireless Communications • April 2010 37
patterns, and the data rate and testbed topologyvary in different experiments. Figure 3b showsthe testbed configuration.
EVALUATION OF BROADCASTING STRATEGIESBoth the multiple-broadcast and multiple-unicastschemes produce more reliable transmissionsthan conventional broadcasting. Unlike the multi-ple-broadcast scheme, which has to broadcastpackets using the minimal transmission rate, themultiple-unicast transmission may apply higherand different data rates for individual nodes. Theexperiments in this section evaluate these twoschemes in terms of reliability, routing construc-tion success ratio, latency, and channel utilization.
Figure 4a depicts the packet error rate (PER)of a one-hop transmission under different IEEE802.11b/g data rates. The results present the linkquality baseline for the following experiments.The results are based on broadcasting 100-bytepackets.5 Both the sender and receiver are con-nected by a 20 dB attenuator, and the distancebetween each unit is fixed at 70 cm. The receivedsignal strength of each MP ranges from –70 to–80 dBm, which is a common reference value forWi-Fi network deployment.6 During these tests, aspectrum analyzer detected two APs in the neigh-boring channels with received signal strengths of–71 and –75 dBm. The signal causes adjacentchannel interference (ACI) and increases thePER in wireless transmissions.7 Under the samechannel quality, including signal strength andACI, the 11b PHY delivers a slower data rate andalso lower PER, while the 11g PHY gives higherones, as Fig. 4a indicates. Thus, a common prac-tice is to use 11b PHY when the channel qualitydrops and 11g PHY when it clears up. Experi-mental results also show that the PER increasessignificantly when the data rate exceeds 36 Mb/s.To help the multiple-broadcast and multiple-uni-cast schemes utilize wireless resources more effi-ciently, the following experiments manually selecta maximum of 36 Mb/s for the data rate.
Figure 4b illustrates the degradation of broad-cast reliability for increasing hop counts in achain topology. Three schemes (i.e., convention-al broadcast, multiple-broadcast, and multiple-unicast) are evaluated with various data rates.The packet arrival ratio measured in this experi-ment represents the percentage of broadcast
packets successfully received by the measuringnode. The conventional broadcast scheme at 36Mb/s has the worst performance in all six config-urations. The two-broadcast scheme significantlyimproves reliability when using the same datarate. The result of the multiple-broadcast schemeat 36 Mb/s is comparable to the conventionalbroadcast scheme at lower data rates (1 Mb/sand 11 Mb/s) when the broadcast repetitionincreases to three. Naturally, the multiple-uni-cast scheme presents the best reliability in allconfigurations, even at 36 Mb/s.
To determine the feasibility of these threeschemes, this study investigates the routing con-struction success ratio in an n × n grid topology,as Fig. 5a shows. The routing construction is atypical function that uses mesh broadcast-typecontrol frames in the IEEE 802.11s path selec-tion framework. The success ratio is measured bycounting successful routing trials, that is, whenthe source node (SRC) correctly receives a PREPfrom the destination node (DST). To reflect uni-casting behavior in the real world, this study alsomeasures the multiple-unicast scheme using theRTL8186 auto-rate mode, automatic rate fall-back (ARF) [15]. Figure 5b shows these experi-mental results. The success ratio of conventionalbroadcast with the most reliable data rate (i.e., 1Mb/s) lowers to 86 percent in the largest experi-mental grid. However, the multiple-unicastscheme at 36 Mb/s keeps above 90 percent suc-cess ratio in all network sizes. Furthermore, thesame scheme using the auto data rate is superiorto all other schemes. Even in the 16-node grid, itshows a 98 percent success ratio for routing con-struction. Examining the distribution of auto datarates reveals that only 40 percent successfullyreceived frames are delivered with 11g data rates(12~54 Mb/s), and 60 percent of the frames aresent at more reliable 11b data rates (1, 2, 5.5,and 11 Mb/s). In other words, the high reliabilityof using auto data rate comes from transmittingmany frames with more robust modulations.
To investigate the side-effects caused by theseschemes, this study measures latency and chan-nel utilization during the experiments above.This study defines latency as the time intervalstarting when SRC issues a broadcast-typePREQ and ending when it receives the corre-sponding PREP. The experimental results in Fig.
Figure 3. Illustration of the testbeds: a) experimental mesh platforms where the Realtek RTL8186 is insidethe ZyXEL P-330W product; b) an experimental deployment.
4 When using attenuators,some effects, such as thenear-far problem [13],can be simulated, whilethe characteristics of mul-tipath fading [14] are stilldifferent from the resultobtained in a field test.
5 Most of the mesh broad-cast-type control messagesare small but extensible,so we chose a small-sizepacket to represent them.
7 IEEE 802.11 regulatesthat the PER must be lessthan 10 percent whentransmitting 1000 octets ina 10 dB noise environ-ment, while the noisecaused by ACI in ourtestbed is much higherthan that.
Unlike the multiple-broadcast
scheme, which hasto broadcast packets
using the minimaltransmission rate,
the multiple-unicasttransmission mayapply higher and
different data ratesfor individual nodes.
LIN LAYOUT 4/8/10 12:56 PM Page 37
Authorized licensed use limited to: National Chiao Tung University. Downloaded on June 28,2010 at 04:09:10 UTC from IEEE Xplore. Restrictions apply.
IEEE Wireless Communications • April 201038
5c indicate that the latency curves of all configu-rations are similar, with only small differences.Even for the largest grid, the slowest routingconstruction time (the multiple-unicast schemeusing the auto data rate) is only 8 ms slowerthan the fastest time, the conventional broadcastat 11 Mb/s, at 40 ms. These results demonstratethat an increased delay is acceptable for routingconstruction.
This study defines the channel utilization asthe percentage of wireless media time occupiedby all nodes on flooding the mesh broadcast-typecontrol frames. This study investigates the rootannouncement messages, which are periodicalmesh broadcast-type control frames maintaininga mesh routing tree, to assess the average chan-nel utilization of mesh broadcast-type controlframes in an IEEE 802.11s network. The totaloccupied media time includes the time spent onPHY layer headers, MAC layer headers, pay-load, and IFSs. Figure 5d depicts the calculatedchannel utilization. Due to the large proportionof low-speed 11b frames, the multiple-unicastscheme using auto data rate consumes morewireless resources than other configurations.However, since this is a small fraction (less than3 percent), this overhead helps maintain therouting structure of a WLAN mesh.
The results above reveal that the multiple-unicast scheme is the most suitable scheme forthe transmission of mesh broadcast-type controlmessages in our testbed. In real-world deploy-ment, however, the PER can change unpre-dictably, and a larger mesh scale may lead tounacceptable latency. Therefore, a sophisticatedbroadcasting strategy considering actual runtimeconditions deserves further study.
EVALUATION OF MULTICHANNEL WITHFEWER RADIOS
Utilizing multichannel with fewer radiosenhances the total throughput of a mesh net-work. This study implemented three cases,including one channel with a single radio
(1C1R), two channels with dual radios (2C2R),and two channels with a single switching radio(2C1R). Figure 6a shows the experimental topol-ogy, which consists of two paths beginning at twotraffic generators and merging at a switch node.During the experiments, two saturated nonblock-ing UDP data streams were simultaneously gen-erated and destined to the traffic sink. Theseexperiments reveal the upper bound throughputof the examined topologies. In the case of 2C1R,the switch node is equipped with one radio toround-robin serve two paths which operate innon-overlapping channels. In the other cases, theswitch node is equipped with a single radio(1C1R) and dual radios (2C2R) to serve twopaths without switching.
Figure 6b depicts the experimental results.The 2C1R throughput is nearly twice as high asthe 1C1R for a staying period of 240 ms (notincluding the 6 ms switch overhead). Thisdemonstrates the improvement of using multi-channel transmissions with fewer radios. Theimprovement is achieved by reducing the inter-flow interference between two paths. Althoughusing 2C1R with a staying period of 240 ms max-imizes the utilization of a single radio, the 2C2Rresults reveal the difference in capacity (i.e., theradio capacity served by single or dual radios)and the side-effect caused by the switching. Theswitch overhead, which is about 6 ms on theRealtek RTL8186, represents the time wastedduring channel switching. Shorter staying periodsincur higher switch overheads. However, longstaying periods produce prolonged latency andpossible buffer overflow, which deterioratestransmission quality. For example, buffer over-flow can decrease throughput when the stayingperiod is greater than 240 ms.
Obtaining an optimal staying period is moredifficult in the real world than under the experi-mental environment above. A channel switchingalgorithm should be based on the switch over-head, queue size, traffic requirement, data rate,and channel status of each channel. The switchoverhead is a burden for switching, but reducing
Figure 4. a) The relation between PER and data rate during experiments: for each data rate, 5000 tests are conducted and the error barsindicate the maximum and minimum measured values; b) packet arrival ratio for different broadcasting strategies in chain topologies:for each data rate, 3000 tests are conducted.
Authorized licensed use limited to: National Chiao Tung University. Downloaded on June 28,2010 at 04:09:10 UTC from IEEE Xplore. Restrictions apply.
IEEE Wireless Communications • April 2010 39
it using a longer staying period causes prolongedlatency. To prevent buffer overflow, the queuesize constrains the maximum staying period.Switching strategies, such as adjusting the switch-ing frequency or allocating different staying peri-ods for each requirement, can be applied to dealwith the different traffic requirements of eachchannel. The switching algorithm should alsotake the data rate and channel status intoaccount, because the time spent on transmittingthe same amount of data with different datarates and different channel status is distinct.Finally, an adaptive channel switching algorithmcombining the above factors in the runtime tooptimize the multichannel transmissions withfewer radios is worthy of development.
CONCLUSIONS AND FUTURE WORKS
This study presents the design and developmentof a WLAN mesh system, and discusses issuesincluding the system design, reliable transmis-sion of mesh broadcast-type control frames, andmulti-channel transmissions. Experimentalresults show that the multiple-unicast schemeachieves a 98 percent routing construction ratio,acceptable latency, and channel utilization in ourtestbed. This study also demonstrates that multi-channel transmissions using a single switchingradio improve effectively throughput withoutextra hardware costs.
Future studies should further investigate theconditions that affect the multiple-unicastscheme. An adaptive channel switching algo-rithm to optimize the multichannel transmissionswith fewer radios is currently being developedand evaluated. Much larger-scale indoor andoutdoor deployments are also planned, with thegoal of finding the optimal number of meshpoints, meshes, and channels in a specific deploy-ment scenario.
ACKNOWLEDGMENTSWe would like to acknowledge the help fromour colleagues, Jui-Hung Yeh, Cho-Hao Lee,and Yung-Sheng Chen. We also thank theanonymous reviewers for their fruitful com-ments.
REFERENCES[1] I. F. Akyildiz, X. Wang, and W. Wang, “Wireless Mesh
[2] R. Bruno, M. Conti, and E. Gregori, “Mesh Networks:Commodity Multihop Ad Hoc Networks,” IEEE Com-mun. Mag., vol. 43, no. 3, Mar. 2005, pp. 123–31.
[3] Y-D. Lin and Y-C. Hsu, “Multihop Cellular: A New Archi-tecture for Wireless Communications,” Proc. IEEE INFO-COM, 2000.
[4] J. Bicket et al., “Architecture and Evaluation of anUnplanned 802.11b Mesh Network,” Proc. ACM Mobi-Com, 2005.
[5] A. Raniwala and T. Chiueh, “Architecture and Algo-rithms for an IEEE 802.11-based Multi-Channel WirelessMesh Network,” Proc. IEEE INFOCOM, 2005.
Figure 5. Experiments for routing establishment: a) the n × n grid topology; b) the success ratio, which is 98 percent when adopting themultiple-unicast scheme using auto data rate during 600 trials; c) the latency to establish a routing path; d) the channel utilization tomaintain the routing tree.
Network size
(b)
2x2
75
70
Succ
ess
rati
o (%
)
80
85
90
95
100
3x3 4x4
Network size
(d)(c)
2x20
Cha
nnel
uti
lizat
ion
(%)
2
4
6
8
3x3 4x4
SRC
(a)
Neighbor
DST
Network size2x2
16
8
Late
ncy
(ms)
3x3 4x4
12
20
24
28
32
36
40
44Conventional broadcast, 1Mb/sConventional broadcast, 11Mb/sTwo-broadcast, 36Mb/sThree-broadcast, 36Mb/sMultiple-unicast, 36Mb/sMultiple-unicast, auto data rate
Conventional broadcast, 1Mb/sConventional broadcast, 11Mb/sTwo-broadcast, 36Mb/sThree-broadcast, 36Mb/sMultiple-unicast, 36Mb/sMultiple-unicast, auto data rate
Conventional broadcast, 1Mb/sConventional broadcast, 11Mb/sTwo-broadcast, 36Mb/sThree-broadcast, 36Mb/sMultiple-unicast, 36Mb/sMultiple-unicast, auto data rate
LIN LAYOUT 4/8/10 12:56 PM Page 39
Authorized licensed use limited to: National Chiao Tung University. Downloaded on June 28,2010 at 04:09:10 UTC from IEEE Xplore. Restrictions apply.
IEEE Wireless Communications • April 201040
[6] K. Ramachandran et al., “On the Design and Implemen-tation of Infrastructure Mesh Networks,” Proc. IEEEWksp. Wireless Mesh Net., 2005.
[7] IEEE P802.11s/D2.03, “Wireless LAN Medium AccessControl (MAC) and Physical Layer (PHY) Specifications— Amendment: Mesh Networking,” Nov. 2008.
[8] L. Hideki et al., “Evaluating the Impact of RTS-CTS inOLPC’s XOs’ Mesh Networks,” XXV Simpósio Brasileirode Telecomunicações, 2007.
[9] C. Perkins and E. Royer, “Ad-Hoc On-Demand DistanceVector Routing,” Proc. 2nd IEEE Wksp. Mobile Comp.Sys. Apps., 1999.
[10] M. Bahr, “Proposed Routing for IEEE 802.11s WLANMesh Networks,” Proc. 2nd Annual Int’l. Wksp. Wire-less Internet, 2006.
[11] K. Tang and M. Gerla, “MAC Reliable Broadcast in AdHoc Networks,” Proc. MILCOM, 2001.
[12] A. Doufexi et al., “A Comparison of the HIPERLAN/2and IEEE 802.11a Wireless LAN Standards,” IEEE Com-mun. Mag., vol. 40, no. 5, May 2002, pp. 172–80.
[13] C. Ware et al., “Unfairness and Capture Behavior in802.11 Ad Hoc Networks,” Proc. IEEE ICC, 2000.
[14] A. Sheth et al., “Packet Loss Characterization in WiFi-BasedLong Distance Networks,” Proc. IEEE INFOCOM, 2007.
[15] A. Kamerman and L. Monteban, “WaveLAN-II: A High-Performance Wireless LAN for the Unlicensed Band,”Bell Labs. Tech. J., 1997, pp. 118–33.
BIOGRAPHIESYING-DAR LIN [SM] ([email protected]) received his Bach-elor’s degree in computer science from National TaiwanUniversity in 1988, and his M.S. and Ph.D. degrees in com-puter science from the University of California, Los Angeles(UCLA) in 1990 and 1993, respectively. He joined the facul-ty of the Department of Computer Science at NationalChiao Tung University (NCTU) in August 1993 and hasbeen a professor since 1999. He spent his sabbatical yearas a visiting scholar at Cisco, San Jose, California, in2007–2008. He is also the founder and director of the Net-work Benchmarking Laboratory (NBL), co-hosted by Indus-trial Technology Research Institute (ITRI) and NCTU since2002, which reviews the functionality, performance, con-formance, and interoperability of networking productsranging from switches, routers, and WLAN to security andVoIP. In 2002 he co-founded L7 Networks Inc., whichaddresses the content networking markets with the tech-nologies of deep packet inspection. His research interestsinclude quality of service, network security, deep packetinspection, embedded hardware software co-design, WLANmesh, and P2P networking. He has been on the editorialboards of IEEE Communications Magazine, IEEE Communi-
SHIAO-LI TSAO [M] ([email protected]) received his B.S.,M.S., and Ph.D. degrees in engineering science fromNational Cheng Kung University, Taiwan, in 1995, 1996,and 1999, respectively. His research interests includemobile communication, wireless networks, and embeddedsoftware. He is currently an assistant professor with theDepartment of Computer Science at National Chiao TungUniversity. He has published more than 60 internationaljournal and conference papers, and holds or has appliedfor 16 U.S. patents. He received the Research AchievementAward of the Industrial Technology Research Institute in2000 and 2004, the Outstanding Project Award of theMinistry of Economic Affairs of Taiwan in 2003, YoungResearchers Award of the Pan Wen-Yuan Foundation in2007, and Young Engineer Award of the Chinese Instituteof Electronic Engineering in 2008.
SHUN-LEE CHANG [M] ([email protected]) received hisB.S. and M.S. degrees in computer science from NationalChiao Tung University, Taiwan, in 1997 and 1999, respec-tively. Currently, he is working toward his Ph.D. He workedfor Formosoft Inc. and developed PKI products during1999–2004. In 2006–2009 he participated in IEEE 802.11s-related research projects at the Realtek-NCTU joint researchcenter. His research interests are the design, analysis, andimplementation of wireless mesh networks.
SHAU-YU CHENG ([email protected]) received hisB.S. degree from the Department of Computer Science andEngineering, Tatung University, Taipei, Taiwan, in 1995,and his M.S. degree from the Institute of Computer Sci-ence, National Chiao Tung University, Hsinchu, Taiwan, in1999. In 1999 he joined the army for substitute service inFormosoft Inc. After retirement from the army, he is pursu-ing his Ph.D. degree at the Institute of Computer Scienceof National Chiao Tung University. He was the team leaderin the Realtek-NCTU joint research center. His researchmainly involves channel estimation and synchronization inwireless communications, multi-spec receiver designs, wire-less mesh networks, and associated VLSI architectures.
CHIA-YU KU [M] ([email protected]) received his M.S.degree from the Department of Computer Science, Nation-al Chiao Tung University, Taiwan, in 2008. He has beenearning his Ph.D. degree since September 2007. Hisresearch interests include wireless mesh networking, multi-channel multiradio, mobility management, integration ofwireless networks, and performance evaluation.
Figure 6. Results of multi-channel transmissions: a) the experimental topology; b) impact of staying period for throughput. The perioddoes not include the 6 ms switch overhead.
Staying period (ms)30
5
0
Thro
ughp
ut (
Mb/
s)
10
15
20
25
30
20 60 120 240 480 960
2C1R2C2R1C1R
Traffic sink Traffic generator
Traffic generator
CH 1CH 11LAN
(a) (b)
MP MPP
MP Switch mode MPP
LIN LAYOUT 4/8/10 12:56 PM Page 40
Authorized licensed use limited to: National Chiao Tung University. Downloaded on June 28,2010 at 04:09:10 UTC from IEEE Xplore. Restrictions apply.
