sticky csma/ca: implicit synchronization and real-time qos in … · 2009-02-19 · sticky csma/ca:...

Ad Hoc Networks 5 (2007) 744–768

www.elsevier.com/locate/adhoc

Sticky CSMA/CA: Implicit synchronization and real-timeQoS in mesh networks

Sumit Singh a,*, Prashanth Aravinda Kumar Acharya b, Upamanyu Madhow a,Elizabeth M. Belding-Royer b

a Department of Electrical and Computer Engineering, University of California, Santa Barbara, CA 93106, United Statesb Department of Computer Science, University of California, Santa Barbara, CA 93106, United States

Received 31 March 2006; received in revised form 24 October 2006; accepted 9 December 2006Available online 14 January 2007

Abstract

We propose a novel approach to QoS for real-time traffic over wireless mesh networks, in which application layercharacteristics are exploited or shaped in the design of medium access control. Specifically, we consider the problem ofefficiently supporting a mix of Voice over IP (VoIP) and delay-insensitive traffic, assuming a narrowband physical layerwith CSMA/CA capabilities. The VoIP call carrying capacity of wireless mesh networks based on classical CSMA/CA(e.g., the IEEE 802.11 standard) is low compared to the raw available bandwidth, due to lack of bandwidth and delayguarantees. Time Division Multiplexing (TDM) could potentially provide such guarantees, but it requires fine-grained net-work-wide synchronization and scheduling, which are difficult to implement. In this paper, we introduce Sticky CSMA/CA, a new medium access mechanism that provides TDM-like performance to real-time flows without requiring explicitsynchronization. We exploit the natural periodicity of VoIP flows to obtain implicit synchronization and multiplexinggains. Nodes monitor the medium using the standard CSMA/CA mechanism, except that they remember the recent historyof activity in the medium. A newly arriving VoIP flow uses this information to grab the medium at the first availableopportunity, and then sticks to a periodic schedule, providing delay and bandwidth guarantees. Delay-insensitive trafficfills the gaps left by the real-time flows using novel contention mechanisms to ensure efficient use of the leftover bandwidth.Large gains over IEEE 802.11 networks are demonstrated in terms of increased voice call carrying capacity (more than100% in some cases). We briefly discuss extensions of these ideas to a broader class of real-time applications, in which arti-ficially imposing periodicity (or some other form of regularity) at the application layer can lead to significant enhancementsof QoS due to improved medium access.� 2007 Elsevier B.V. All rights reserved.

Keywords: Mesh networks; Quality of service; Medium access control; Voice over IP; Time division multiplexing

1570-8705/$ - see front matter � 2007 Elsevier B.V. All rights reserved

doi:10.1016/j.adhoc.2006.12.008

* Corresponding author. Tel.: +1 805 893 7520.E-mail addresses: [email protected] (S. Singh), acharya@

cs.ucsb.edu (P.A.K. Acharya), [email protected] (U. Mad-how), [email protected] (E.M. Belding-Royer).

1. Introduction

Wireless mesh networks offer an attractive solu-tion for providing voice, multimedia, and best effortdata services in areas where deployment of a wiredinfrastructure is not viable or is economically

.

mailto:[email protected]

mailto:acharya@ cs.ucsb.edu

mailto:acharya@ cs.ucsb.edu



S. Singh et al. / Ad Hoc Networks 5 (2007) 744–768 745

unattractive. Such networks can extend the reach ofthe wired backbone through a backhaul mesh ofwireless routers, not only providing communicationservices to remote areas but also offering a flexibleand economical means of scaling network capacityto serve increasing demand in enterprises and cam-pus communities. Mesh networks can also providea self-contained communication infrastructure thatcan be deployed in a plug-and-play fashion forsearch, rescue and disaster recovery operations.However, a key drawback of current wireless meshtechnology based on IEEE 802.11 standards is thatit is not possible to provide bandwidth and delayguarantees to real-time applications such as voiceand video [1]. The primary reason is the variabilityin delay and loss performance that results fromthe requirement that each packet must contendafresh for the medium. These drawbacks of IEEE802.11 are only partly alleviated by its QoS exten-sion IEEE 802.11e. Given that medium accesscontrol is the bottleneck for providing QoS, analternative approach is to employ Time DivisionMultiplexing (TDM) to obtain a collision-freetransmission schedule. However, this requires fine-grained network-wide synchronization and schedul-ing, which are difficult to implement. In this paper,we introduce Sticky CSMA/CA,1 a new mediumaccess control mechanism that provides TDM-likeperformance to real-time flows without requiringexplicit synchronization.

Our starting point is the observation that thetraffic generated by real-time flows exhibits a signif-icant regularity. Can this regularity be exploited formore effective medium access control? After all,contention between nodes occurs because a nodecannot predict the actions of its neighbors. If thetransmission schedule for a node follows a regularpattern, its neighbors should be able to learn thispattern using the CSMA mechanism, and thusavoid contention. While these ideas are potentiallyapplicable to a broad class of real-time flows, in thispaper, we focus specifically on illustrating how theyenable efficient support of Voice over IP (VoIP)over wireless mesh networks. The success of VoIPover the wireline Internet leads us to expect that itwill be a key application driving the wider accep-

1 The term sticky routing has been used for routing protocols intelephone networks [2] that grab and hold resources whilefeasible. Sticky CSMA/CA also grabs and holds resources, butfor channel access: the context and protocol details are com-pletely different.

tance and commercial success of wireless mesh net-works. A broad range of wireless VoIP solutions arecommercially available and are being deployed inenterprises and campuses [3,4]. However, theseproducts typically optimize VoIP performance overa single wireless hop from a tetherless node to anAccess Point connected to the wired network. Ourgoal is to support VoIP in wireless mesh networks,thus significantly increasing the flexibility of deploy-ment. The challenge is to provide an adequate levelof QoS, in terms of assured bandwidth, andbounded delay and delay jitter.

Sticky CSMA/CA is based on the assumptionthat all the real-time flows in the network are eithernaturally periodic (e.g., VoIP) or have been shapedby the higher layers as periodic (constant bit rate)streams with the same period. A node that has totransmit packets from a real-time flow monitorsthe medium using the standard CSMA mechanism.When it detects an opportunity to transmit, itattempts packet transmission, and on being success-ful, sticks to a periodic schedule. Neighbors detectsuch periodic schedules using the CSMA mecha-nism and avoid interfering with them. This mediumaccess approach leads to a TDM-like sharing of themedium among the real-time flows, with the period-icity of the application and transmission schedulebeing used to obtain an implicit form of synchroni-zation. Delay-insensitive traffic has a lower prioritythan real-time traffic and utilizes the bandwidth leftover from the real-time flows.

From the point of view of implementation,Sticky CSMA/CA requires that the nodes maintaincarrier sense tables that record the periodic flowswhose existence it can infer using the CSMA mech-anism. Upon arrival of a new real-time flow, a nodelooks at its carrier sense table to determine a peri-odic schedule over which the medium is free, andcontends for and grabs the medium at the first suchopportunity (i.e., it starts sending in a time intervalthat does not interfere with the existing real-timeflows in its carrier sense table). If the flow setup issuccessful, the node locks onto a periodic transmis-sion schedule. All the other nodes in the networkthat overhear the control message exchange updatetheir carrier sense tables immediately, and the nodesthat sense the new periodic transmission infer that anew flow has been set up. Delay-insensitive traffichas lower priority than real-time traffic and fills inthe gaps remaining after the real-time flows havebeen accommodated. We develop new contentionresolution mechanisms for delay-insensitive traffic

746 S. Singh et al. / Ad Hoc Networks 5 (2007) 744–768

that enhance the efficiency with which they utilizethe left-over bandwidth.

Our performance evaluations focus on a mix ofVoIP and data traffic, for which large gains in VoIPcall carrying capacity are demonstrated relative toboth IEEE 802.11b and IEEE 802.11e. We showthrough analysis and extensive simulations thatSticky CSMA/CA achieves more efficient mediumutilization than IEEE 802.11e and can be used toprovide the required QoS in terms of bandwidth,delay and delay jitter to real-time flows. The under-lying reason is that this scheme obviates the need formedium contention for every frame that a nodetransmits, exploiting the predictable and periodicnature of the high priority traffic. This approach sig-nificantly reduces the probability of collision andthe overhead due to the backoff required for colli-sion avoidance for every frame in IEEE 802.11.We also show that, along with efficient support ofsignificantly more real-time flows, the performancein terms of throughput achieved for the low prioritydelay-insensitive traffic is comparable to that ofIEEE 802.11. These encouraging results indicatethat even for applications that do not exhibit peri-odicity and constant data rates naturally (e.g.,video, or even high-speed data pipes in a wirelessbackhaul), it may be useful to artificially imposeperiodicity and predictability to improve the effi-ciency of medium access. Detailed investigation ofcross-layer architectures that shape variable ratereal-time traffic to take advantage of StickyCSMA/CA is beyond the scope of this paper. How-ever, we do provide some discussion as to how thismight be done for applications such as video inSection 6.

The rest of the paper is organized as follows. Sec-tion 2 describes related work. The design of StickyCSMA/CA is described in Section 3. In Section 4,we present an approximate analysis to obtaininsight into the performance of IEEE 802.11 andSticky CSMA/CA. Section 5 presents the simulationresults obtained in mesh networks with line and gridtopologies. We conclude with a discussion of vari-ous design and application aspects of StickyCSMA/CA in Section 6. Note that we use the termsVoIP and voice interchangeably throughout thispaper.

2. Related work

In the past few years, a considerable amount ofwork has addressed the area of MAC layer design

for providing the required QoS to voice, multimediaand data applications over wireless networks. Wediscuss some of the proposed approaches next.

The IEEE 802.11 standard MAC sublayer pro-vides two different access mechanisms: the Distrib-uted Coordination Function (DCF), and the PointCoordination Function (PCF) [5]. PCF is a central-ized approach to perform bandwidth reservations,which splits time into a contention-free period(CFP) and a contention period (CP). A pollingscheme controlled by a Point Coordinator thatresides in the Access Point arbitrates channel accessduring the CFP. PCF is not suited for multihopwireless networks due to the requirement of central-ized control and the Point Coordinator at theAccess Point. The Distributed Coordination Func-tion (DCF) is based on CSMA/CA for frame trans-missions and a random backoff mechanism to avoidpacket collisions. We describe this scheme in moredetail in Section 3. DCFs drawback is that it doesnot provide service differentiation. Due to lack ofboth medium access priorities and QoS supportfor different applications, DCF is not suited for sup-porting real-time applications.

The IEEE 802.11e standard extends the IEEE802.11 framework by including mechanisms for ser-vice differentiation [6]. Four access categories (AC),mapped to eight traffic priorities, are defined forIEEE 802.11e. These ACs correspond to differentvalues of the following parameters: minimum con-tention window size (CWmin), maximum contentionwindow size (CWmax), and the Arbitration Inter-frame Space (AIFS) (time-interval between idletransition of the medium and the start of channelaccess or backoff). A lower value of these parame-ters causes shorter backoff intervals or wait periodsbefore medium access, thereby prioritizing traffic.IEEE 802.11e has different queues for differentACs. It also provides an optional token-like mecha-nism called TXOP (Transmission Opportunity) tosupport packet bursts, where a node that hasobtained access to the medium can send multipleframes for an AC, up to a maximum time-limit,without further contention. This contention-freebursting mechanism aids in transmission of burstytraffic and improves medium utilization. One ofthe drawbacks of IEEE 802.11e is the lack ofassured QoS for individual flows. IEEE 802.11ecan only provide service differentiation amongdifferent ACs, but QoS degrades in the presence ofmultiple contenders in the same AC because of med-ium access contention with equal priority for every


packet belonging to the same AC. The underlyingreason for lack of support of a deterministic QoSin IEEE 802.11e and IEEE 802.11 is that everypacket requires contention for the medium in theseschemes, and no attempt is made to exploit the (nat-ural or possibly imposed) predictability of packettransmissions to minimize contention losses.

Blackburst [7] is a scheme based on modifiedIEEE 802.11 CSMA/CA designed to provide QoSto real-time applications. In this scheme, a nodewith real-time data contends for medium access bysending an energy burst (called a blackburst) fortime proportional to the wait time of the node.The node with the longest burst wins the contentionand transmits a frame. Blackburst does not addressthe hidden terminal problem and requires that thenodes have the capability to jam the medium.Another distributed MAC scheme derived fromIEEE 802.11 DCF is Distributed Fair Scheduling(DFS) [8], which is based on Self-Clocked FairQueuing (SCFQ) [9]. DFS achieves service differen-tiation by assigning different backoff intervals topackets that belong to different flows based on theweights assigned to the flows. This feature ensureshigher throughput to flows with higher weights,and fairer overall allocation of bandwidth as com-pared to IEEE 802.11 DCF. DFS does not considerthe delay bounds of real-time packets, nor does itaddress mechanisms for determining appropriateweights for different flows, both of which greatlyinfluence the performance.

Multiple Access Collision Avoidance with Piggy-back Reservations (MACA/PR) [10] is a scheme thatsupports bandwidth reservations for real-time traf-fic. In this scheme, each node maintains the real-timescheduling information of the neighboring nodes inits reservation tables (RT) by overhearing the datapackets or ACKs. The global reserved slot informa-tion is obtained by periodic RT exchange amongneighbors. The RT exchange takes care of the hiddenterminal problem. The drawbacks of MACA/PR arethe piggybacking-based reservation approach andthe overhead of periodic RT exchange.

A few recent proposals attempt to take advan-tage of TDMA schedules for reducing interferenceand for fair resource allocation over CSMA basednetworks [11,12]. The Overlay MAC Layer (OML)[11] is an access control and scheduling scheme thatpartitions time into equal size slots and allocatesthese slots among loosely synchronized contendingnodes. Though OML implements temporal fairnessand reduces interference by improving the predict-

ability of medium access among nodes, the coarsetime-slot based resource allocation does not addressthe QoS requirements of different applications.Another interesting MAC framework proposed inthe context of sensor networks is ZMAC [12].ZMAC exploits TDMA schedules among locallysynchronized neighbors to improve medium utiliza-tion and reduce packet collisions among two hopneighbors in a CSMA network operating under highcontention. Under low contention, all the nodescontend for a slot but the slot-owner nodes have ahigher priority over their neighbors by virtue of ashorter contention window. This scheme incurs ini-tial schedule allocation overhead in the networkdeployment phase and, because it was designed fora sensor network, it does not consider service differ-entiation among different applications for QoSsupport.

To the best of our knowledge, the proposedSticky CSMA/CA scheme is the first approach thatexploits the characteristics of real-time applicationsto devise medium access control strategies thatreduce delay and delay jitter. Exploiting the naturalperiodicity of VoIP is shown (in the succeedingsections) to lead to large capacity gains relative toIEEE 802.11. This motivates future work (see thediscussion in Section 6) in taking this approachone step further, by imposing artificial regularityat the application layer in order to obtain theenhanced QoS obtained using Sticky CSMA/CAmedium access.

3. The sticky CSMA/CA protocol

In this section, we describe the Sticky CSMA/CAprotocol. We briefly outline conventional CSMA/CA and introduce history considerations intoCSMA/CA. We then describe the carrier sensingmechanism and discuss the operation of StickyCSMA/CA for real-time and delay-insensitivetraffic.

3.1. Conventional CSMA/CA

The IEEE 802.11 DCF uses CSMA/CA for datapacket transmission, along with a random backoffmechanism to avoid collisions. A node that has aframe to transmit senses the transmission activityin the medium. If the medium is free for a distributedinterframe space (DIFS) interval, the frame is trans-mitted. If the medium is sensed as busy, the nodewaits until the medium is free for a fixed time interval

DIFS

Node B has packet to send

Node A has packet to send

Transmit Busy Transmit

Backoff Slots

TransmitBusySIFS

DATA ACK

DATA

Busy

DATA

DIFS

ACK ACKNode A

Node B

Node A has packet to send

DIFS

DIFS

Backoff for Node B

Backoff for node A

Fig. 1. CSMA/CA: Backoff mechanism.

ABA BABA B

Cycle Time

Activity Map of BActivity Map of AA B

Activity Map of C

BusyBusyC

Fig. 2. Activity map: nodes A and B have a bidirectional real-time flow between them. Node C is a neighbor of nodes A and B.The activity maps of nodes A, B and C contain reservations forthe real-time flow between nodes A and B. Note that the cyclesneed not start at the same time.


(DIFS if the last frame was received without anytransmission errors, or extended interframe space(EIFS) otherwise). After the medium is sensed idlefor a DIFS/EIFS interval, the node generates abackoff counter bi, which is a pseudo-random integerwith a uniform distribution in the interval [0, CWi],where CWi = min(CWmax,2i�1(CWmin + 1) � 1) forthe ith transmission attempt for a frame. The nodethen decrements the backoff counter bi by one forevery slot width interval if the medium remains idle.If the medium becomes busy before bi reaches zero,the backoff procedure is frozen and is continuedafter the medium is sensed idle again for a DIFSinterval. When the backoff counter reaches zero,the node transmits the frame. If the transmissionfails and the ACK is not received from the receiver,then the node moves to the next backoff stage (i + 1)and repeats the procedure. Fig. 1 describes the back-off procedure for two nodes, A and B, that can heareach other.

3.2. Introducing history into CSMA/CA

CSMA/CA with history incorporates the historyof the recent transmission activity in the mediuminto the conventional CSMA/CA mechanism forpacket transmissions. The primary difference fromconventional CSMA/CA is that the medium activityhistory information is used to infer the time win-dows that are occupied by periodic real-time flows.Time windows with periodic transmissions are con-sidered reserved by the existing real-time flows, andhence there is no contention for these time windows.We assume that all the realtime flows are periodicwith a common period. The most relevant exampleof such flows is voice applications, which generateperiodic data based on the codec used (e.g., ITU-T G.711 codecs periodically generate fixed sizedpackets at a data rate of 64 Kbps).

We define a cycle as the time period between twoconsecutive packet transmissions of any ongoingreal-time session. Therefore, the transmission activ-ity over the last few cycles can be used to infer thereal-time flow reservations of the medium. All thenodes in the network maintain a medium activity

map containing the medium busy/free state infor-mation over a fixed number of past cycles. A nodeinvokes conventional CSMA/CA to find a transmitopportunity, taking into account its medium activ-ity map to ensure that it does not interfere withany existing real-time flow. A backoff mechanismsimilar to IEEE 802.11e, but with smaller conten-tion window parameters, is followed to avoid colli-sions. As an example, consider a three nodenetwork with a bidirectional real-time flow betweennodes A and B. Fig. 2 shows the activity maps of themedium at nodes A, B and C, assuming that node Ccan hear nodes A and B.

3.3. Sticky CSMA/CA

We view continuous time in terms of time-slots.As mentioned before, we assume that the cycle timeis the same for all the nodes in the network. Notethat since we assume an asynchronous framework,the cycles can start at different times at different

Busy

Idle

Busy

Idle

Busy

CYCLE TIME

Carrier Sense History Table

Medium

Fig. 3. Nodes use CSMA to monitor the medium and mark theircurrent carrier sense history table according to the sensedmedium busy and idle times.

Carrier Sense Table

Carrier Sense History Tables

Cycle T–1Cycle T–2

Cycle T–3Cycle T–4

Cycle T

Fig. 4. At the beginning of each cycle, nodes derive their carriersense table as a summary of their carrier sense history tables. Thecarrier sense table helps to identify periodic real-time flows.


nodes. A node with a new voice session practices agreedy algorithm to grab a free time window. A freetime window is defined as a set of contiguous time-slots during which the medium is free. If the attemptto establish a flow is successful (we describe themechanism of establishing a call later in this sec-tion), the node sticks to the time window. Thismeans that the node sends packets periodically dur-ing the same time window in every cycle, unless thesession performance, measured in terms of packetloss, degrades beyond a certain threshold. All newvoice connection setup attempts and other packettransmissions honor the time windows occupied bythe existing voice connections and do not contendfor those slots. We describe the carrier sensing andcontention mechanisms of Sticky CSMA/CA, anddiscuss various implementation considerations next.

3.3.1. Carrier sensing mechanism

All nodes in the network maintain a set of Car-rier Sense History Tables (CSH Tables). A singleCSH Table is a log of medium transmissions sensedby a node over a cycle. The CSH table contains oneentry for each time-slot of the cycle. Each entry inthe CSH Table is marked busy or free, based onthe transmissions sensed in the medium during thecorresponding time-slot. A node marks a slot asbusy if it senses a transmission in its neighborhoodduring that time-slot, or if it transmits or receives inthat time-slot. The CSH Table is updated at everymedium transition among the four possible physicallayer states: Idle, Sensing, Receiving, and Transmit-ting. It is important to note that the MAC layerdoes not need to be explicitly informed about thestate of each time slot. MAC layer can deduce theslot states from the medium state transition infor-mation passed from the PHY layer without anyadditional overhead. Medium state transition infor-mation is required even in conventional CSMAMAC schemes. Fig. 3 shows an example of CSHTable maintenance. The number of CSH Tablesrequired in order to maintain the history of mediumactivity over the past N cycles is N.

In addition to the CSH Tables, the nodes main-tain another table called the Carrier Sense Table

(CS Table). The CS Table is an activity map ofthe medium that represents the predicted busy andfree time-slots over the current cycle. The CS Tablehas a format similar to a single CSH Table. The CSTable is derived at the beginning of every cycle usingthe CSH Tables over the previous N cycles. A sim-ple approach to obtain a CS Table is to do a major-

ity decision over each slot in a cycle, e.g., if a slot isfound busy in 75% of the CSH Tables over the lastN cycles, it is marked as busy in the CS Table. Fig. 4shows an example of CSH and CS Tables.

The CS Table is maintained only on the basis ofperiodicity of the sensed transmissions, and theinformation about the flows for which the nodeitself is the sender or receiver. This approach signif-icantly reduces the control overhead as compared tothe other techniques involving message exchanges,piggybacking control information over data pack-ets, or high power transmissions. There is no extramessage exchange required in this technique.

At high medium utilization, a few large sizeddata packets (e.g., FTP packets of size 1500 bytes)transmitted in the carrier sensing range (CSR) of anode over contiguous cycles might cause the nodeto falsely interpret these transmissions as voice con-nections. This can cause the node to mark the corre-sponding slots in its CS Table as busy, resulting inunderutilization of bandwidth. We observe that ahigh percentage of data packets in the network aresignificantly larger than voice packets. We use thisproperty to distinguish such data packets from voicepackets. Whenever a node senses a transmission fora time interval significantly longer than that possiblefor a voice packet (e.g., more than double the timerequired for a voice packet), it concludes that ithas sensed a low-priority data packet transmissionand does not mark its CSH Table. Also, a node does


not mark its CSH Table whenever it receives oroverhears low-priority data packet transmissions.In addition, simple sanity checks of the CS Tabledata can also help in removing irrelevant historyinformation from the CS Table. The underlying rea-son for not accounting for low priority data trans-missions in the CSH Table is to ensure that anode considers only relevant high priority periodicdata transmission information while preparing itsCS Table, thus foregoing transmission opportuni-ties only for the periodic high priority traffic in thenetwork. These mechanisms improve medium utili-zation and aid in prioritizing voice flows in theentire CSR of a node.

3.3.2. Contention mechanism for real-time flows

To set up a new voice connection, a node checksits CS Table to find a set of contiguous free time-slots to accommodate the new flow. If a free timewindow is found, the node waits until the beginningof the free time window. It then employs the IEEE802.11 CSMA/CA mechanism to transmit a Real-time RTS (R-RTS) message to the intended recei-ver. The R-RTS message contains the requiredpacket transmission duration information for thenew flow and a flow-identifier. The flow-identifier,along with the source node address, uniquely identi-fies a flow in the network. On receiving an R-RTSmessage, the receiver checks its CS Table (activitymap) to determine whether it can support the flowfrom the time instant when the R-RTS transmissionstarted (the propagation delay is negligible ascompared to the transmission duration, and thusis not considered). If the corresponding time-slotsare free, the node sends a Real-time CTS (R-CTS)message to the sender. The R-CTS message containsthe flow-identifier and the duration information

Transmit R–RTS

R–CTS Received

DIFS

Backoff Slots

Busy

Medium Activity

Busy

DIFS

Backoff Slots

Backoff = 0, but cannot transmitdue to upcoming busy period

Cycle T

Sender’s Activity map at cycle T

Cyc

Se

Fig. 5. Establishing a new flow: the sender uses its activity map to selectThe node transmits an R-RTS message at the chosen slot and receives

obtained from the R-RTS message. The senderand the receiver mark their CS Table and overwriteall the CSH Tables to indicate the presence of thenew flow. The neighboring nodes of the senderand the receiver overhear the R-RTS and R-CTSmessages and use the duration information toupdate their CSH Tables in a similar manner. Thenodes in the carrier sense range (CSR) of the senderand the receiver sense this periodic transmission andupdate their CS Tables after a few cycles. Fig. 5shows the flow setup procedure for a node.

To ensure that real-time flows win over lowpriority datagrams while contending for the med-ium during flow setup, we use an approach similarto IEEE 802.11e. IEEE 802.11e uses the contentionwindow parameters (CWmin,CWmax) and the AIFSto prioritize different Traffic Classes (TC). SmallerCWmin and CWmax values and a shorter AIFS per-iod ensure that a packet that belongs to a high pri-ority TC accesses the medium ahead of a packetbelonging to a low priority TC, thereby prioritizingdifferent TCs.

After a successful flow setup, the sender locks onto the slots; i.e., it transmits voice packets during thesame time-slots in every cycle. The receiver does notsend ACKs for the voice packets. ACKs result in asignificant overhead for real-time flows with shortpackets such as voice because the size of the ACKpackets is comparable to the size of the data packets[13]. Also, the physical layer preamble and header(PLCP preamble and header), which are transmittedat the lowest data rate, consume most of the trans-mission time required for these short packets.

Packet loss information summaries are sent bythe receiving node to the sender through a feedback

mechanism. The feedback mechanism is required forthe sender to evaluate the condition of a real-time

BusyBusy Busy

le T+1

nder’s Activity map at cycle T+1

a suitable time window when it can establish a new real-time flow.an R-CTS message in response, thus setting up the flow.


flow in terms of packet loss in the medium. This isnecessary because the ACK mechanism of IEEE802.11 is not used for real-time packets. Feedbackmessages are used to inform the sender of the cumu-lative packet losses seen by the receiver over the lastM cycles. A sender node requests a feedback mes-sage by setting the Feedback Requested bit in theMAC frame header of the voice packet. The senderrequests a feedback message after every M/2 cycleswithout feedback and continues sending the requestwith every packet of the flow until it receives a feed-back message. If the maximum number of cycles Mwithout the feedback message is reached, the senderpauses the flow and reattempts flow setup at a differ-ent time-slot through an R-RTS/R-CTS dialog. Wecall this mechanism slot reconfiguration. The senderalso utilizes the cumulative packet loss informationsent in the feedback messages to decide whether slotreconfiguration is required.

Sticky CSMA/CA addresses the hidden terminalproblem in the following ways. The control messageexchange for slot reservation (R-RTS and R-CTS)between the sender and the receiver incorporatesthe receiver’s view of the network before choosingslots for data transfer. A receiver does not respondto the sender’s R-RTS message if it is aware ofperiodic transmissions over a portion of the timewindow requested, or if it senses the medium to bebusy during these slots. In case a sender-receiverpair observes persistent packet losses over thereserved slots because of interference (e.g., from ahidden terminal), the node pair attempts to shiftto another free time window that has a higher prob-ability of being interference free. Thus, nodes avoidbecoming trapped in lossy transmission patterns byattempting slot reconfiguration after the observedpacket loss exceeds a certain threshold. Our perfor-mance evaluation shows that slot reconfiguration isan effective technique to avoid lossy transmissionpatterns.

Another effective solution to counter packetlosses due to interference is out-of-stream packetretransmission. When the receiver of a periodic flowdoes not receive an expected packet, it sends aretransmit request to the sender. On receiving aretransmit request, the sender retransmits the corre-sponding packet as a high priority datagram. Therequested packet should be sent out-of-streambecause inclusion of this packet in the periodicstream will delay the subsequent packets in thestream. Our performance evaluation indicates thatout-of-stream retransmissions effectively counter

losses of high priority voice packets. Anotherapproach for minimizing packet losses due to inter-ference is to provide redundancy in the transmittedpayload by including multiple voice payloads in oneframe, so that the loss of a few frames does notaffect the performance [13].

In general, the hidden terminal problem can bepartially addressed by using sensitive antennas(i.e., with low carrier sense threshold) at the nodesso that any node that wishes to transmit can sensewhether it will interfere with its neighbors’ ongoingpacket receptions, and thereby avoid such interfer-ence. This solution has an associated tradeoff ofreduced spatial reuse because of the aggravatedexposed terminal problem. Nevertheless, thisapproach has been proposed in literature for reduc-ing packet collisions in systems where the requiredantenna sensitivity is not a design constraint [14].We have observed this tradeoff for Sticky CSMA/CA in our performance evaluation.

3.3.3. Contention mechanism for delay-insensitive

flows

Delay-insensitive flows are assigned a lower pri-ority than real-time flows. MAC layer operationfor delay-insensitive data flows is based onCSMA/CA with history considerations. The datapacket transmissions honor the reservations madeby real-time flows. Nodes do not attempt to trans-mit data packets over slots in which high priorityperiodic transmission is expected. As mentionedbefore, data packets have higher backoff parametersand a higher AIFS than that of real-time flows toensure that real-time flow setup requests receive ahigher priority. However, the CWmin and CWmax

parameters are set to lower values as compared toIEEE 802.11. This is because the time gaps left bythe real-time flows might not be sufficiently largeto incorporate the probable long delays due to theexponential backoff mechanism used in IEEE802.11. Moreover, with reduced contention in avoice/data network, large contention windows arenot required for collision avoidance. Our perfor-mance evaluation shows that even with smaller con-tention window parameters, the throughputachieved by Sticky CSMA/CA for delay-insensitivebest effort traffic is comparable to that for IEEE802.11.

With a large number of voice calls in the net-work, the time gaps left might not be sufficient totransmit data packets, which can lead to starvationof data flows. In systems where starvation of data


flows is not acceptable, this problem can be allevi-ated by setting a threshold on the number of voicecalls admitted. Admission control schemes toenforce such thresholds are higher layer functional-ities that are beyond the scope of this paper. Wetherefore do not impose any limits on the numberof voice calls in our performance evaluation.

3.3.4. Implementation considerations

We discuss some important considerations forthe implementation of Sticky CSMA/CA in thissection.

Time-slots: Time-slots are the time units in termsof which we define the entries (marked busy or free)in the CS Table and the CSH Tables. We definetime-slots of a fine granularity. We mark a time-slotas busy even when the medium is busy for only afraction of the time-slot duration. Thus, a time-slotshould be small enough so that the portion of timefor which the medium is not utilized is insignificant.Also, a small slot width increases the memoryrequirements for the tables. Based on these consid-erations, we chose the time-slot duration for StickyCSMA/CA as 20 ls, which is the default slot dura-tion (used for backoff) of IEEE 802.11b. In our sim-ulations, we assume the cycle time as 20 ms (thepacketization interval for voice). To allow for devi-ations of clocks of different nodes, extra slots (calledLeeway slots) at the beginning and the end of thetransmission are marked as busy in the CSH Table.

CS Table Maintenance: The CS and the CSHTables are implemented as bitmaps, with each bitcorresponding to a time-slot. The CSH Table forthe current cycle is updated only on medium transi-tions in order to avoid the overhead of updating thetable at every time-slot. Given a fixed cycle time, thesize of the CS Table and the CSH Tables dependson the width of the time-slots. Also, the total sizeof the CSH Tables depends on the number of previ-ous cycles (N) needed to derive the CS Table. Forexample, with the above mentioned slot width andcycle time, the memory requirement for a systemwith CSH Tables spanning six cycles is 875 bytes.

Reattempt Timer: It is necessary that the CSTable represent an accurate prediction of the trans-missions in the current cycle. A node that senses themedium as free might not be allowed to contend forthe channel because its CS Table indicates animpending transmission. An example scenariowhere this can happen is when a real-time flow endsin the carrier sensing range (CSR) of a node; thenode will have the corresponding slots in its CS

Table marked as busy in the current (and possiblythe next) cycle, depending on the majority decisionrule used to decide the busy/free status of each slot.In order to improve medium utilization in suchcases, we use a Reattempt timer. A node starts areattempt timer when it is not allowed to set up anew flow or transmit a data packet because the med-ium is expected to be busy as per its CS Table. Thetimer is set to expire at one of the intermediate slotsof the expected busy period. When the reattempttimer expires, the node determines whether the med-ium is actually busy. If the medium is found to befree, the node initiates the Sticky CSMA/CA packettransmission procedure again. This mechanismincreases medium utilization even in the less proba-ble case of an inconsistent CS Table.

Dummy Reservation Packets: When there is apacket loss at an upstream node during an ongoingmultihop voice flow, the downstream nodes of theflow do not have packets to send to their next hopneighbors. In this situation, a downstream nodesends dummy packets at its scheduled transmissionslot to the next hop node. This transmission helpsin avoiding contention for the slot by the CSRneighbors, in case of loss of a few packets (less thanthe loss threshold M, at which we delete the flow).Dummy reservation packets can also be used tomaintain slot reservations during short silence peri-ods (up to a maximum time limit) when VoIP pack-ets are not generated because of the silencesuppression methods being used in the codec.

Probabilistic CSH Table Updates: A node thatreceives a data packet from its neighbor cannotsimultaneously sense the real-time transmissions ofanother non-interfering node in its CSR. Therepeated occurrence of this event (due to high datatraffic) can cause the nodes to lose the periodic flowinformation in their CS Tables. This in turn allowsthem to transmit in the slots reserved by their CSRneighbors and cause interference. For this reason,we increase the probability of a node’s retentionof the real-time flow information in its CS Tableby using probabilistic updates to the CSH Table.Whenever a node loses the periodic transmissioninformation of its CSR neighbors from its CS Tabledue to receiving a long data packet, it copies theinformation from its CS Table to the current CSHTable with a high probability, assuming that theCS Table has more reliable information for thoseslots. Probabilistic CSH Table updates reduce thechances of losing the slot reservation informationand result in improved performance for voice flows


in the CSR of a node. We incorporate probabilisticupdates in our simulations with a high probabilityvalue (p = 0.95).

4. Performance analysis for a clique

We now provide an approximate analytical com-parison of the number of VoIP calls supported bySticky CSMA/CA against those supported by IEEE802.11b DCF [15] and IEEE 802.11e EDCF [6]. Forthis purpose, we consider single hop flows over aclique, which is a network where all nodes can heareach other. We assume that there are no hidden ter-minals and the radio link is error free. Despite thesesimplifications, the clique-based analysis can beused to predict performance for multihop flows overmore complex networks, which can be roughlyviewed as multiple cliques stitched together. Becauseof the shared nature of the wireless medium, thebandwidth required by a multihop flow can be esti-mated as the number of hops traveled times thebandwidth required over one link. Such computa-tions can then be used to identify bottleneck links,which limit the maximum throughput in a multihopwireless network [16]. Applying the clique-basedestimates to the nodes surrounding the bottlenecklink then provides capacity estimates that matchclosely with the simulated performance.

4.1. IEEE 802.11b DCF performance

The performance of IEEE 802.11 DCF has beenanalyzed in detail in [17,18] for single hop scenarios.Refs. [19,20] incorporate the considerations for mul-tihop flows. Our objective is to obtain the total voicecall carrying capacity for an IEEE 802.11b network.We estimate the voice carrying capacity of IEEE802.11b using the saturation throughput analysisin [17], where saturation throughput is defined asthe maximum stable throughput of the network.This analysis assumes ideal channel conditions (no

DIFSBusy Busy

t+2 t+4t+1t t+3

Backoff Counte

Backoff

Medium State

Fig. 6. Variable time-intervals between two consecutive b

hidden terminals, error free radio link, and back-logged packet queues at every node) for a networkof n nodes. Through Eqs. (1)–(3), we briefly outlinethe derivation found in [17] for convenience (refer to[17] for details).

Continuous time is divided into variable lengthtime intervals, defined as epochs, at the beginningof which the backoff counter of each node decreasesby one. These epochs can be the IEEE 802.11b slot

width (hence forth called d) of 20 ls if the medium isfree, or can also include a medium busy time inter-val, since the backoff counter is frozen when themedium is sensed busy. In the following analysisof IEEE 802.11b DCF, by the term epoch, we willrefer to this definition of variable length time inter-vals. Fig. 6 shows the epochs as the time intervalsbetween two consecutive backoff counter decre-ments. A key assumption in this model is that ateach transmission attempt, irrespective of the back-off state of the node, the packet transmitted collideswith a constant and independent probability p,called the conditional collision probability. Underthese assumptions, the stochastic processes repre-senting the backoff stage i 2 (0,m); and the backoffcounter k 2 (0, Wi � 1), where Wi = 2iWmin; cantogether be modeled as a bidimensional discrete-time Markov chain. Using this Markov chainmodel, the probability that a node transmits in achosen epoch is obtained as

pt ¼2ð1� 2pÞ

ð1� 2pÞðW min þ 1Þ þ pW minð1� ð2pÞmÞ ; ð1Þ

where Wmin is the minimum contention window,and m is the maximum allowed backoff state (i.e.,Wmax = 2mWmin). In order to evaluate the condi-tional collision probability p, due to the indepen-dence assumption, p can be defined as

p ¼ 1� ð1� ptÞn�1: ð2Þ

This means that the probability of collision for atransmitted frame is the probability that at least

DIFS Busy

t+7t+6t+5

r Decrement Intervals

Backoff

ackoff counter decrements at an IEEE 802.11 node.

Table 1Overhead in terms of time, with average data rate R

Overhead Bytes Time (ls)

RTP 12 12 * 8/RUDP 8 8 * 8/RIP 20 20 * 8/RIEEE 802.11 MAC 28 28 * 8/RPHY preamble 72/1PHY header 48/2SIFS 20DIFS 50ACK 14 14 * 8/R + PHY overheadFeedback 20 20 * 8/R + PHY overheadPropagation (D) 1


one of the other n � 1 nodes also transmits in thesame epoch. Eqs. (1) and (2) can be solved for differ-ent values of n, m, and Wmin to obtain p and pt forthese cases.

Define the probability that there is a transmissionin the considered epoch as Ptx, and the probabilitythat a transmission that occurs in an epoch is suc-cessful as Ps. The normalized saturation throughputS is defined as the fraction of time the channel isused to successfully transmit payload bits, and canbe evaluated as the ratio of the average time spentin the transmission of useful payload informationover an average epoch duration. Define Tpayload asthe time taken to transmit a payload of a fixed size,Ttx as the time taken for successful transmission of acomplete packet including headers and other over-heads, and Tcollision as the epoch duration whenthe medium is sensed busy by each node during acollision. Therefore,

S ¼ P txP sT payload

ð1� P txÞdþ P txP sT tx þ P txð1� P sÞT collision

: ð3Þ

The normalized saturation throughput S can beevaluated for different sets of n, m, and Wmin. Thecapacity of a bottleneck link in a multihop networkcan be approximated through S evaluated for thetwo node case using (3), where both nodes alwayshave packets to transmit. The bottleneck link capac-ity constrains the maximum number of multihopflows in the network where all the nodes can heareach other. For example, if the bottleneck link sup-ports a maximum of k voice calls, the network cansupport a maximum of k/2 two hop voice calls, ork/3 three hop voice calls, etc. Note that in reality,this approximation might be imprecise because allthe transmitters might not transmit with the samepower; the nodes might not be spaced equally apart;and the channel conditions are not the same at allthe nodes. Nevertheless, it helps to obtain an esti-mate of the number of calls that can be supported.

4.2. IEEE 802.11b VoIP call carrying capacity

Table 1 shows the overhead associated with eachvoice frame sent, due to different protocol layerheaders and the MAC protocol overhead. Weassume that an ITU-T G.711 codec packetizingaudio data every 20 ms is used, which results in pay-loads of size 160 bytes (data rate = 64 Kbps). Weassume that IEEE 802.11b operates at the highestdata rate of 11Mbps using the basic access mecha-nism (no RTS-CTS). We also assume that IEEE

802.11b uses the short physical layer preamble(PLCP preamble), which reduces the synchroniza-tion time from 192 ls (when the long physical layerpreamble is used) to 96 ls. The values mentionedabove correspond to the state-of-the-art IEEE802.11b VoIP equipment commercially available[4]. Define Tframe as the time required to transmita frame (payload and headers), and D as the propa-gation time. From Table 1, we have

T payload ¼160 � 8

11ls;

T frame ¼ T PHY þ T MAC þ T RTPþUDPþIP þ T payload

¼ 262 ls;

T tx ¼ T frame þ SIFSþACKþDIFSþ 2 � D ¼ 430 ls;

T collision ¼ T frame þDIFSþ D ¼ 313 ls:

Solving Eqs. (1) and (2) numerically for the two nodecase n = 2, Wmin = 32, and m = 5, (as (CWmin,CWmax) for IEEE 802.11b are (31, 1023)) [15] andplugging in the required probability and transmis-sion time values in (3), we get the normalized satura-tion throughput S = 0.1924. This gives the saturationthroughput = S * 11 Mbps = 2.1164 Mbps. Eachvoice call is equivalent to two unidirectional con-stant bit rate (CBR) flows of data rate 64 Kbps.Therefore, the required data rate for a single VoIPflow = 128 Kbps. Thus, we obtain the maximumnumber of calls supported as 16. This numbermatches with the capacity we obtain from simula-tions, as shown in Fig. 7. The simulation parametersare summarized in Section 5. Note that if the longpreamble is used, the maximum number of calls sup-ported by IEEE 802.11b network decreases to 12.

The IEEE 802.11e EDCF network for a voice-only scenario is similar to IEEE 802.11b DCF, withdifferent contention window parameters (i.e., (7, 31)

0

5

10

15

20

25

0 5 10 15 20 25

Num

ber

of S

uppo

rted

Cal

ls

Number of Incoming Calls

Sticky CSMA/CAIEEE 802.11eIEEE 802.11b

Fig. 7. The total number of voice calls supported as a function ofthe incoming calls for a two node network, with nodes 150 mapart.


instead of (31,1023) for voice packets). Note thatprioritization of different access categories doesnot play a role in this scenario because we are onlyconsidering voice calls. The total number of voicecalls obtained in this case is 19, which matches withthe capacity obtained via simulations (see Fig. 7). Itis important to note that optional IEEE 802.11eprotocol features (such as transmission opportunity(TXOP) that facilitates contention free bursting)when enabled, result in a different overall networkbehavior. The effect of TXOP has been accountedfor in the simulation based performance evaluationin Section 5.

4.3. Sticky CSMA/CA VoIP call carrying capacity

Sticky CSMA/CA requires the nodes with real-time flows to stick to a periodic transmission sche-dule. The other nodes in the carrier sensing range(CSR) neighborhood that hear the periodic trans-missions do not attempt to contend for the mediumduring the slots in which they expect a periodictransmission. By sticking to a periodic transmissionschedule and using contention only to reserve slotsfor a real-time flow, Sticky CSMA/CA reduces thenumber of contending nodes in the medium drasti-cally (in contrast with IEEE 802.11, where the num-ber of contenders at any instant equals the numberof nodes that have packets to transmit). This leadsto substantial gains in terms of reducing the over-head involved with backoff for every packet trans-mitted in the medium. Based on this fact, wereason that in a network with only voice flows, the

total number of contenders at any time instant ina Sticky CSMA/CA network equals the number ofnew arriving calls (and the calls being reconfiguredto a different time window), which is significantlyless than the total number of nodes with packetsto transmit at any instant. With this observation,we argue that it is reasonable to assume that, onaverage, the first stage of backoff yields a winnerwith a high probability, even with a very smallCWmin. Thus, we use contention window parame-ters smaller than IEEE 802.11b for real-time flowsetup requests in a Sticky CSMA/CA network:CWmin = 3 and CWmax = 7. These contentionwindow parameters are used in a manner similarto IEEE 802.11e for service differentiation. We donot need to account for the slots wasted due to col-lisions in the calculation of the maximum number ofVoIP connections because the time-slots wasted inthe current cycle due to a collision (and the conse-quent DIFS interval and backoff) can effectivelybe used by a new VoIP call in the coming cycles.

Sticky CSMA/CA uses feedback messages sentperiodically or on-demand to the sender in placeof ACK messages for voice packets. In our simula-tions, the sender demands a feedback message everysixth cycle. Thus, the average additional mediumaccess time required per cycle to transmit feedbackmessages for a real-time flow is one-sixth of the timerequired for transmitting a single feedback message.Note that the slots reserved for a voice packet alsoinclude leeway slots (L = 1) on both sides of thereserved slots.

We express all the time requirements in terms oftime-slots (with slot size d = 20 ls). Because a cyclespans 20 ms (the rate at which the 160 bytes voicepackets are generated), T s

cycle ¼ 1000. The minimumaverage gap between the slots reserved for a newvoice call placed next to an already existing reserva-tion in the cycle is DIFS + E[x] Æ d, where E[x] isdefined as the expected value of a random variablex that models the portion of the time gap (expressedin terms of d) between an established voice call anda new call placed next to it that is contributed by thebackoff duration associated with every new call. Thelast reserved slot of a new call need not have anytime gap from the first slot of an already existingcall reservation placed next to it. This means thattwo adjacent calls might or might not have any timegap between them, depending on the order of theirarrival. We assign equal probabilities to the twopossible call arrival orders. Thus, we define the aver-age minimum time gap associated with each call as

3 41 2 5

150m

Gateway Gateway

Fig. 8. Line topology: the nodes ace 150 m apart. The edge nodes1 and 5 act as gateway nodes, with all the calls being routedthrough them.


T sempty ¼ 0:5 DIFS

d þ E½x��

, where E½x� ¼ CWmin

2¼ 1:5

slots. Define the total slots required for transmittinga frame, including the overheads as T s

tx; total slotsrequired for feedback messages (of size 20 bytes)as T s

feedback; and the total number of slots requiredas Slotsreqd. Then,

T sempty ¼

ðE½x� � dþDIFSÞ2d

¼ 2;

T stx ¼ðT frame þ DÞ

dþ T s

empty ¼ 16;

T sfeedback ¼

T PHY þ 20�811þ DþDIFS

dþ E½x� ¼ 10;

Slotsreqd ¼ T stx þ

T sfeedback

6

� �þ 2 � L ¼ 20:

Therefore, the number of slots required per voiceconnection comprising two unidirectional flows isT s

voip ¼ 2 � Slotsreqd ¼ 40. Thus, the maximum num-ber of calls that can be supported is estimated asT s

cycle=T svoip ¼ 25. This analysis shows that Sticky

CSMA/CA has more than 50% additional call car-rying capacity as compared to IEEE 802.11b, andmore than 25% additional call carrying capacity ascompared to IEEE 802.11e. The maximum numberof calls supported for the two node scenario ob-tained by simulations is 23 (see Fig. 7). The errorbars at 21 or more incoming calls indicate that thenumber of supported calls varies depending on callarrival times in the different simulation runs. Theloss at high number of incoming calls can be attrib-uted to the packing inefficiency while placing calls ina cycle. We do not employ any mechanism to tightlypack the calls in the cycle where a greater number ofcalls could be supported.

Thus, Sticky CSMA/CA achieves almost 45%more VoIP call capacity as compared to IEEE802.11b, and 20% more VoIP call capacity as com-pared to IEEE 802.11e over the two node scenario.The underlying reasons are the gains due to periodicscheduling and reduced contention as discussedearlier.

While evaluating the voice call carrying capacity,we have considered only voice traffic in the network.When data is introduced, Sticky CSMA/CA per-forms much better than IEEE 802.11b/e in termsof the number of supported calls because of the pri-oritization of voice traffic induced by grabbing aperiodic transmission schedule. Moreover, the lowercontention window and AIFS parameters help inachieving a higher priority for voice flow setuprequests, which aids the overall voice throughput.

In Section 5 we demonstrate that even with a higherpriority to voice calls and no packing optimization,the data throughput achieved by Sticky CSMA/CAis comparable to IEEE 802.11b/e.

We now evaluate the performance of StickyCSMA/CA and compare against both IEEE802.11b and IEEE 802.11e via simulations.

5. Performance evaluation

In this section, we evaluate the performance ofSticky CSMA/CA in terms of VoIP calls supportedalong with the low priority data traffic, over differ-ent network scenarios. Through simulations, wedemonstrate that Sticky CSMA/CA achieves amuch higher VoIP call carrying capacity than IEEE802.11b/e while providing the required QoS to theVoIP flows.

5.1. Simulation setup

We use the QualNet Network Simulator [21] toevaluate the performance of Sticky CSMA/CA.We consider three different network topologies: asingle line of five nodes, spaced 150 m apart; a reg-ular grid of 25 nodes, spaced 150 m apart; and anirregular grid of 25 nodes with an average distanceof 150 m between neighboring nodes. Figs. 8 and9 illustrate the line and the regular grid topologiesalong with the gateway nodes that connect the meshnetwork to the wired backbone. We discuss theirregular grid topology later in this section. Thesetopologies model different degrees of medium con-tention and interference among nodes in mesh net-works and aid in understanding the performanceof Sticky CSMA/CA and IEEE 802.11b/e over realmesh network deployments. Because all the (out-going/incoming) voice calls and data traffic arerouted to their destinations through the wired back-bone, the gateway nodes act as terminal points forthe voice and data traffic in a mesh network. Thisarrangement of nodes models the current singleradio mesh network deployments, where the closestgateway for each mesh node is usually no more than

150m

150m

Fig. 9. Grid topology: 25 nodes placed 150 m apart in a squaregrid. The nodes colored black act as the gateway nodes for thenetwork.


two hops away [22]. Because Sticky CSMA/CA isdesigned for mesh networks, all the nodes in the net-work are assumed to be stationary.

Table 2Simulation parameters

(a) Physical layer and MAC layer

MAC layer parameters Values PHY

Slot time 20 ls PropSIFS 10 ls AnteDIFS 50 ls PHYIEEE 802.11b CWmin 31 TranIEEE 802.11b CWmax 1023 Data

PLCSyncPropCon

(b) Sticky CSMA/CA

Sticky CSMA/CA parameters

AIFS (Voice)AIFS (Data)(CWmin,CWmax) voice(CWmin,CWmax) high priority data(CWmin,CWmax) background dataR-RTSR-CTSMAC data headerCycle timeMissed feedback threshold

(c) IEEE 802.11e

IEEE 802.11e parameters

AIFS (Voice)AIFS (Background data)(CWmin,CWmax) voice(CWmin,CWmax) background dataTXOP limit for voice

We use the two-ray propagation model with con-stant shadowing as our physical layer model. Toevaluate Sticky CSMA/CA using a simple and trac-table physical layer model, we consider a flat terrainand neglect the time-varying effects of multipathfading. Table 2(a) lists the simulation parametersused in our simulations. The physical layer parame-ters used for Sticky CSMA/CA result in a transmis-sion range (TR) of 200 m, and a carrier sensingrange (CSR) of approximately 450 m. For nodesspaced 150 m apart, the interference range (IR) isapproximately 420 m. We use the short physicallayer preamble provided by the IEEE 802.11b stan-dard. IEEE 802.11b uses the basic access mecha-nism (RTS-CTS turned off). Table 2(c) lists theparameters used for service differentiation in IEEE802.11e. The simulation parameters used for IEEE802.11b/e simulations are known default parametersthat are recommended by the respective standards

layer parameters Values

agation model Two ray modelnna Omnidirectionalnoise factor 10

smission power 11.0 dBmrate 11.0 Mbps

P preamble Short preamble (72 bits)hronization time 96 lsagation delay 1 ls

stant shadowing mean 4 dB

Values

50 ls70 ls(3,7)(15,31)(31,63)26 bytes20 bytes30 bytes20 ms10

Values

50 ls150 ls(7,15)(31,1023)3008 ls

0

2

4

6

8

10

12

14

0 2 4 6 8 10 12 14

Num

ber

of S

uppo

rted

Cal

ls


IEEE 802.11e (TxOP):Voice CallsIEEE 802.11e:Voice Calls

Fig. 10. The effect of TXOP on the total number of voice callssupported over a five node line topology.


or are commonly used by the state-of-the-art equip-ment available in the market [15,6,4].

To simulate the bi-directional VoIP calls, we usetwo simultaneous constant bit rate (CBR) flowsbetween the source and the destination nodes, onein each direction. The network layer is assumed tosupport priority queuing of voice packets over thebackground data traffic. We consider an ITU-TG.711 codec packetizing 20 ms of audio data in160 bytes of payload (data rate 64 Kbps). We modelcall arrival as a Poisson process. Table 2(b) listsparameter values specific to Sticky CSMA/CA usedin the simulations. In order to prioritize real-timetraffic over data packets, we set the contention win-dow parameters (CWmin,CWmax) for voice setuprequests as (3, 7), for high priority data as (15,31),and for background data as (31, 63). The Arbitra-tion Interframe Space (AIFS) is 50 ls for voiceand high priority data, and 70 ls for backgrounddata traffic. These parameters ensure a higher prior-ity to the voice traffic than the data traffic. The car-rier sense history information is maintained for thepast six cycles. A sender node that does not receivea feedback message from the receiver for 10 consec-utive cycles (Missed Feedback Threshold) attemptsslot reconfiguration. Slot reconfiguration is alsoattempted when the feedback message sent by thereceiver indicates more losses than a certain thresh-old (10–20% in our simulations).

We evaluate the performance of Sticky CSMA/CA and compare it with IEEE 802.11b and IEEE802.11e in terms of the total number of VoIP callssupported with the required QoS, i.e., boundedpacket loss, average delay, and delay jitter. Weexplore the impact of background traffic (bothTCP and UDP) on the call carrying capacity andestimate the achievable throughput for these flows.Note that we do not address the starvation of back-ground data flows due to voice traffic in our simula-tions, since the objective is to understand the effectof these flows on the voice call traffic and vice-versa.Our results can be used in practical systems to deter-mine the maximum number of voice calls that canbe admitted, while sustaining a given data through-put. Call admission control, resource constrainedrouting, and end-to-end signaling for call establish-ment are higher layer functionalities that are beyondthe scope of this paper.

We impose a delay constraint of 50 ms for voicepackets, since this constraint allows ample time forthe transfer of voice packets over the backbone net-work to ensure a total end-to-end delay of 150 ms.

We consider a VoIP call as useful only if 95% of thepackets sent in both directions are delivered withinthe delay bound. We use these constraints to modelthe QoS requirements of VoIP calls [23]. The simula-tion results have been averaged over 10 runs.

We generate the irregular grid topology by uni-formly varying the placement of nodes within asquare of side 40 m around the corresponding posi-tions in the regular grid topology. This arrangementof nodes ensures that, with a high probability,neighboring nodes are within the TR of each otherso that links between them still exist, but withdifferent link qualities. We use a different uniformrandom grid topology with each simulation run.

5.2. Line topology

Consider a mesh network that consists of a singleline of five nodes as shown in Fig. 8. Nodes 1 and 5act as gateways for the network. All the calls arerouted out of the network through these gateways.Thus, one of the two gateways in the network isalways the terminal point of the CBR flows corre-sponding to a VoIP call in the network.

As mentioned in Section 2, IEEE 802.11e has anoptional feature, called Transmission Opportunity(TXOP), that allows contention-free bursting forthe corresponding access category (AC) after a nodehas won medium contention. TXOP means that anode that has won medium contention can continu-ously send data up to a maximum allowed transmis-sion duration for that AC without having tocontend again. As shown in Fig. 10, over the linetopology, the total number of calls supported by


IEEE 802.11e improves from 9 to 11 with TXOP.We incorporate TXOP support in our IEEE802.11e simulations to increase the IEEE 802.11evoice call carrying capacity.

Fig. 11a shows the average number of voice callssupported as a function of the number of incomingcalls. We generate voice calls by randomly pickingthe nodes (other than gateways) in the networkand assigning the nearest gateway as the terminalnode for each call. Hence, calls traverse either a sin-gle hop or multiple hops to the gateway dependingon the source node. Each data point in Fig. 11a cor-responds to the average number of calls supportedfor a fixed number of incoming calls. Observe fromFig. 8 that Node 3 can carrier sense all the nodes inthe network. Since the links from Node 3 are effec-tively shared by all the other nodes in the network,Node 3 is a bottleneck node. From Fig. 11a, weobserve that the total number of calls supportedby Sticky CSMA/CA is 16, whereas IEEE 802.11e

0

3

6

9

12

15

18

0 2 4 6 8

Num

ber

of S

uppo

rted

Cal

ls

Number of Inco


0

3

6

9

12

15

18

0 2 4 6 8 10 12 14 16 180

0.5

1

1.5

2

2.5

3

3.5

Num

ber

of S

uppo

rted

Cal

ls

TC

P T

hrou

ghpu

t (M

bps)


Sticky CSMA/CA:Voice CallsIEEE 802.11e:Voice CallsIEEE 802.11b:Voice Calls

Sticky CSMA/CA:TCP TputIEEE 802.11e:TCP TputIEEE 802.11b:TCP Tput

a

b c

Fig. 11. Sticky CSMA/CA call carrying capacity for the line topology: (voice and UDP background traffic. All figures show 95% confidence in

and IEEE 802.11b support a maximum of 11 and9 calls respectively. The gains in the VoIP call carry-ing capacity for Sticky CSMA/CA result fromreduced contention overhead, reduced averagebackoff time, and the elimination of per-packetACKs. Notice that in the absence of a call admis-sion control mechanism, the performance degradeswhen the number of incoming calls exceeds the totalcapacity of the network. Lack of resource reserva-tion and medium contention for every packet trans-mission result in reduced total call carrying capacityfor IEEE 802.11b/e.

Fig. 11b shows the average number of supportedvoice calls when there is background TCP trafficconsisting of payloads of varying sizes (maximumsegment size of 1460 bytes). In this evaluation sce-nario, we have two TCP flows in the network, orig-inating from gateway nodes 1 and 5, and terminatingat nodes 2 and 3, respectively. This TCP traffic sce-nario models the long-lived FTP and other TCP

10 12 14 16 18

ming Calls

0

3

6

9

12

15

18

0 2 4 6 8 10 12 14 16 180

100

200

300

400

500

600

700

800

Num

ber

of S

uppo

rted

Cal

ls

UD

P T

hrou

ghpu

t (K

bps)



Sticky CSMA/CA:UDP TputIEEE 802.11e:UDP TputIEEE 802.11b:UDP Tput

a) voice traffic only, (b) voice and TCP background traffic and (c)tervals for the average number of supported calls.

0

1

2

3

4

5

6

0 50 100 150 200 250 300 350

Thr

ough

put (

Mbp

s)

Time (s)

First Call Arrives

Last Call Arrives

Last Call Ends

First CallEnds

Fig. 12. Instantaneous throughput of a FTP flow in a represen-tative simulation for sticky CSMA/CA.


based content download applications from the gate-way nodes that connect to the Internet. The datatraffic workload ensures that all the nodes in the net-work have backlogged low priority data queues, andthat the nodes are directly affected by the back-ground TCP traffic dynamics. The capacity of StickyCSMA/CA, in terms of the total number of sup-ported calls, remains the same even in the presenceof background TCP traffic of varying payload sizes.This is because of the effective priority given to voicepackets over data transmissions in the entire carriersensing range (CSR) of a node. IEEE 802.11e (withTXOP) supports a higher number of voice calls ascompared to IEEE 802.11b because of service differ-entiation, but the total number of calls supported byIEEE 802.11e is still significantly less than that sup-ported by Sticky CSMA/CA. Due to interferenceand lack of prioritization in IEEE 802.11b, the num-ber of supported voice calls reduces drastically withTCP traffic, since different nodes with voice or datapackets contend with each other for medium access.For data traffic in Sticky CSMA/CA, the bandwidthleft over by the real-time flows is sufficient to main-tain average data throughput slightly lower thanIEEE 802.11b but relatively higher than IEEE802.11e.

With fewer voice calls in the network, the effectof interference caused by heavy data traffic in termsof loss of many calls for IEEE 802.11b is apparentin Fig. 11b. IEEE 802.11e is able to recover fromthese losses by multiple fast retransmissions of voicepackets aided by shorter contention windows forvoice. The TXOP feature helps in recovering fromdelays due to retransmissions by transmitting theaccumulated voice packets together in bursts. StickyCSMA/CA effectively copes with the interferencecaused by data traffic with the help of techniquesdiscussed in Section 3.3.2, which results in approxi-mately 200% additional call carrying capacity ascompared to that of IEEE 802.11b, and 65% addi-tional call carrying capacity as compared to thatof IEEE 802.11e.

Fig. 11c shows the average number of supportedvoice calls in the presence of background UDP traf-fic. We use three UDP flows, each with a data rateof 256 Kbps and a payload size of 1460 bytes. TheseUDP flows originate from the three non-gatewaynodes in the network, and are routed to the nearestgateway node. As in the TCP background traffic sce-nario, this data traffic workload ensures that all thenodes in the network have low priority data packetsto send, and that they have to deal with the back-

ground UDP traffic dynamics to sustain voice callswithout performance degradation. We observe fromFig. 11c that the voice call performance is not affectedby the background traffic, and the capacity in termsof the total number of calls supported remains thesame. UDP throughput for Sticky CSMA/CA startsto decrease after the number of incoming callsincreases to an extent that the leftover gaps betweenthe voice calls are not sufficient to sustain the UDPflows. Sticky CSMA/CA UDP throughput is compa-rable to that of IEEE 802.11b/e.

Fig. 12 shows the instantaneous TCP throughputfor Sticky CSMA/CA with increasing number ofincoming calls over the line topology. Observe thatTCP throughput decreases as the number of callsincreases, demonstrating that the priority mecha-nism provides effective service differentiationbetween background data traffic and voice. Dataflows adapt to the voice traffic in the networkdepending on the leftover capacity.

5.3. Grid topologies

Consider a 25 node regular grid topology asshown in Fig. 9. All VoIP calls in the system arerouted out of the network through the gateways.The gateways are also assumed to be the sourcesof long-lived TCP or UDP flows. Though the gridtopology provides more opportunities for spatialreuse, it is also characterized by increased conten-tion and interference among nodes. We comparethe performance of Sticky CSMA/CA with IEEE802.11b and IEEE 802.11e in terms of the totalnumber of VoIP calls supported with the required


QoS over the grid topology. We explore the impactof background traffic (both TCP and UDP) on thecall carrying capacity.

Fig. 13a shows the average number of voice callssupported as a function of the number of incomingcalls. We generate calls by picking nodes in the net-work at random and assigning the nearest gatewayas the terminal node for each call. Ties betweenthe gateways are broken arbitrarily. For UDP orTCP flows, the gateway nodes are chosen as thesources, and the destination nodes are randomlypicked among the nodes serviced by the gateway.We see that the total number of calls supported bySticky CSMA/CA is 25, whereas IEEE 802.11e sup-ports a maximum of 18 calls and IEEE 802.11b sup-ports only 15 calls. The increase in the number ofcalls supported by Sticky CSMA/CA and IEEE802.11b/e over the grid topology is because ofincreased spatial reuse of the medium.

0

3

6

9

12

15

18

21

24

27

0 5 10 15 20 25

Num

ber

of S

uppo

rted

Cal

ls



Num

ber

of S

uppo

rted

Cal

ls

0

3

6

9

12

15

18

21

24

27

0 5 10 15

Num

ber

of S

uppo

rted

Cal

ls

Number of Incoming



a

c

b

Fig. 13. Sticky CSMA/CA call carrying capacity for the regular grid topand (c) voice and UDP background traffic. All figures show 95% confi

Fig. 13b shows the average number of supportedvoice calls as a function of the number of incomingcalls in the presence of background TCP traffic withvarying payload sizes (maximum segment size of1460 bytes). In this scenario, we generate the back-ground TCP traffic by initiating one TCP flow fromeach gateway node and choosing a random non-gateway node from its service area as the destina-tion, routing packets via preconfigured shortestpath routes. Thus, the background TCP trafficdynamics differ over different simulation runs withdifferent seeds. We observe that the maximum callcarrying capacity for Sticky CSMA/CA remainsthe same as that without the background TCP traf-fic. The effect of high interference in the grid topol-ogy is significantly more pronounced for IEEE802.11b/e in terms of majority of incoming callsbeing dropped. Sticky CSMA/CA is able to copewith this increased interference except for a small

0

3

6

9

12

15

18

21

24

27

0 5 10 15 20 250

1

2

3

4

5

6

7

8

TC

P T

hrou

ghpu

t (M

bps)




20 250

100

200

300

400

500

600

700

800

900

1000

1100

UD

P T

hrou

ghpu

t (K

bps)

Calls

ology: (a) voice traffic only, (b) voice and TCP background trafficdence intervals for the average number of supported calls.


probability of call drop in some cases with fewerincoming calls. As the number of VoIP flowsincreases, the interference problem is alleviated asthe time gaps between VoIP flows decrease, reduc-ing the overall data throughput. The voice call car-rying capacity of Sticky CSMA/CA in this scenariois many times higher than that of IEEE 802.11b/e.

The increased data throughput observed forIEEE 802.11b can be attributed to the lack of prior-itization of voice over background data because ofwhich data packets achieve much higher throughputat the expense of voice calls. This lack of service dif-ferentiation adds to the effect of increased interfer-ence in the grid topology and further degrades thevoice call performance for IEEE 802.11b. Withfewer calls in the network, TCP throughput forSticky CSMA/CA is slightly less than that for IEEE802.11e, but this increased throughput for IEEE802.11e is at the expense of most of the incomingvoice calls being dropped. Note that the voice call

0

3

6

9

12

15

18

21

24

27

0 5 10 15

Num

ber

of S

uppo

rted

Cal

ls

Number of Inco


0

3

6

9

12

15

18

21

24

27

0 5 10 15 20 25 0

1

2

3

4

5

6

7

8

Num

ber

of S

uppo

rted

Cal

ls

TC

P T

hrou

ghpu

t (M

bps)




a

b c

Fig. 14. Sticky CSMA/CA call carrying capacity for the irregular gridtraffic and (c) voice and UDP background traffic. All figures show 95%

performance of Sticky CSMA/CA does not changesignificantly in the presence of TCP packet trans-missions of varying sizes. This demonstrates thatthe CS Tables maintained by the nodes in the net-work contain the relevant periodic high priorityflow information and are not corrupted by thedata packet transmissions even under heavy TCPtraffic.

Fig. 13c shows the average number of supportedcalls when there is background UDP traffic in thenetwork. The background UDP traffic is generatedin a similar manner as the background TCP traffic,with each UDP flow having a data rate of 256 Kbpsand a payload size of 1460 bytes. Again, a capacitygain is observed. The sharp fall in data throughputas compared to IEEE 802.11b/e after 10 voice callscan be attributed to the time gaps between the voicetransmissions being too small for data packets to betransmitted. Note that the data throughput isapproximately equal for Sticky CSMA/CA and

20 25

ming Calls

0

3

6

9

12

15

18

21

24

27

0 5 10 15 20 25 0

100

200

300

400

500

600

700

800

900

1000

1100

Num

ber

of S

uppo

rted

Cal

ls

UD

P T

hrou

ghpu

t (K

bps)




topology: (a) voice traffic only, (b) voice and TCP backgroundconfidence intervals for the average number of supported calls.


IEEE 802.11b/e until the number of voice calls inthe network reaches the maximum that can be sup-ported by IEEE 802.11b/e. The loss of backgrounddata for more incoming calls is the cost of support-ing significantly more high priority voice calls.

Fig. 14 illustrates performance comparison interms of voice call carrying capacity of the networkin voice only case, and in presence of backgroundTCP/UDP traffic over the irregular grid topology.We observe that the results obtained for the irregu-lar grid topology and the regular grid topologyshow very similar trends.

We make the following observations regardingthe supported calls results for different simulationruns for all the test topologies: the variation ofthe results for different simulation seeds is minimaluntil the number of calls approaches the system’scall-carrying capacity for Sticky CSMA/CA andIEEE 802.11b/e for the voice only case. A similartrend is observed for voice call results with UDPbackground traffic, except that the results start tovary earlier (by approximately two calls) than thevoice only case. For TCP background traffic, thevariation of the number of supported calls is higherthan that observed for the voice only and the UDPbackground traffic cases for all the protocols. Inter-estingly, for both IEEE 802.11b and 802.11e overthe regular and irregular grid topologies, the mostdominant result for all data points (over mostseeds) is zero calls, and the deviations are non-zeronumbers that account for the very low average callcarrying capacity for both IEEE 802.11b/e. This isbecause of high interference and contention thatresult in heavy packet losses for both IEEE

0.1

1

10

100

1000

0 2 4 6 8 10 12 14 16 18

Del

ay (

ms)



a

Fig. 15. Average end-to-end delay and delay jitter for voice calls ov

802.11b/e. Also, Sticky CSMA/CA voice callresults vary comparatively more than the voice onlycase for low number of incoming calls (up to five)because of increased interference effects from back-ground data traffic.

5.4. QoS performance: delay and delay jitter

We compare the QoS performance of StickyCSMA/CA and IEEE 802.11b/e in terms of theaverage end-to-end delay and delay jitter for voicecalls. Figs. 15–17 illustrate the average end-to-enddelay and delay jitter for Sticky CSMA/CA andIEEE 802.11b/e for the line topology, the regulargrid topology and the irregular grid topologies,respectively. The traffic in the network consists ofonly voice calls. The delay performance for StickyCSMA/CA remains fairly constant with increasingnumber of voice calls. We notice the steep rise indelay for IEEE 802.11b and IEEE 802.11e whenthe number of incoming calls exceeds the networkcapacity. Because of the TDM-like synchronizationachieved by Sticky CSMA/CA, the average delayjitter is orders of magnitude less compared to thatobtained for IEEE 802.11b/e.

Performance trends of Sticky CSMA/CA interms of the voice call carrying capacity and theoffered QoS are very similar over the line topology,the regular grid topology and the irregular gridtopologies. This observation implies that the perfor-mance of Sticky CSMA/CA scales over moregeneral, irregular mesh network topologies. Inorder to gain more insight into Sticky CSMA/CA

1e-04

0.001

0.01

0.1

1

10

100

0 2 4 6 8 10 12 14 16 18

Jitte

r (m

s)



b

er the line topology: (a) end-to-end delay and (b) delay jitter.

0.01

0.1

1

10

100

1000

0 5 10 15 20 25

Del

ay (

ms)



1e-05

1e-04

0.001

0.01

0.1

1

10

100

1000

0 5 10 15 20 25

Jitte

r (m

s)



a b

Fig. 17. Average end-to-end delay and delay jitter for voice calls over the irregular grid topology: (a) end-to-end delay and (b) delay jitter.

1 2 3 4

Gateway

150m

65 7 8 9

Gateway

Fig. 18. Nine node line topology: the edge nodes 1 and 9 act asgateway nodes.

0.01

0.1

1

10

100

1000

0 5 10 15 20 25

Del

ay (

ms)



1e-05

1e-04

0.001

0.01

0.1

1

10

100

1000

0 5 10 15 20 25

Jitte

r (m

s)



a b

Fig. 16. Average end-to-end delay and delay jitter for voice calls over the regular grid topology: (a) end-to-end delay and (b) delay jitter.


performance over different practical mesh net-work scenarios, we next study the effect of increas-ing number of hops on Sticky CSMA/CA and802.11b/e voice call performance. We then look atthe voice call performance over networks employingsilence suppression to optimize bandwidth con-sumption of voice calls.

5.5. Effect of increasing number of hops

In order to evaluate the effect of paths with largenumber of hops on Sticky CSMA/CA voice call per-formance, we consider a mesh network with ninenodes arranged in a single-line topology (seeFig. 18). We include background TCP traffic tomodel realistic network traffic loads. The voice calland data traffic load generation methodology is sim-ilar to that used for the previous evaluations in this

section. Thus, there are voice calls that traverse two,three or four hops over the network depending onthe nodes chosen as sources of voice calls for eachsimulation run. Note that for calls traversing fourhops, the gateway node is out of the carrier senserange of the source node, and thus it does not havecomplete carrier sensing information of the VoIPcall in its CS Table. There are two TCP flows inthe network which traverse three hops (with gate-way nodes 1 and 9 as sources and nodes 4 and 6as destinations). Fig. 19 compares the total number

0

3

6

9

12

15

18

0 2 4 6 8 10 12 14 16 0

0.5

1

1.5

2

2.5

3

3.5

Num

ber

of S

uppo

rted

Cal

ls

TC

P T

hrou

ghpu

t (M

bps)




Fig. 19. Voice call support for nine node single line topologynetwork.


of supported calls and aggregate TCP throughputfor Sticky CSMA/CA and IEEE 802.11b/e. Weobserve that Sticky CSMA/CA has a higher call car-rying capacity in this scenario. This is consistentwith the results reported for other network topolo-gies, and shows that Sticky CSMA/CA performsbetter than IEEE 802.11b/e even for paths withmore hops.

5.6. Silence suppression

Silence suppression is an important configurablefeature that is used to reduce the bandwidth usageof VoIP flows. End nodes that employ silence sup-pression do not send packets during the silence peri-ods in the VoIP call thereby saving bandwidth.Many common audio codec standards along withthe Realtime Transport Protocol (RTP) support

0 3 6 9

12 15 18 21 24 27 30 33 36 39 42

0 5 10 15 20 25 30 35 40 0

1

2

3

4

5

6

7

Num

ber

of S

uppo

rted

Cal

ls

TC

P T

hrou

ghpu

t (M

bps)




Fig. 20. Voice call support for regular grid topology networkwith silence suppression enabled.

silence suppression [24]. We compare the perfor-mance of Sticky CSMA/CA with IEEE 802.11b/eprotocols over an example scenario of the regulargrid topology with background TCP traffic. Weuse Brady’s On–Off Markov Modulated Fluidmodel [25] to model voice calls with talk spurtsand silence periods. Brady’s model has been com-monly used in the VoIP literature because of its sim-plicity and ability to capture the dynamics oftelephonic conversations with reasonable accuracy.In this model, the silence periods and the talk spurtperiods of a voice source are exponentially distrib-uted random variables with means of 1.35S andIS, respectively. The probability p that the sendingend is active is given by 0.43.

Fig. 20 compares the voice call carrying capacityof Sticky CSMA/CA with IEEE 802.11b/e whensilence suppression is used. We observe that StickyCSMA/CA supports a higher number of voice calls.The underlying reason for this performance gain isthat Sticky CSMA/CA is able to exploit the predict-ability of packet transmissions over talk spurt dura-tions, thereby reducing medium contention andpacket loss due to collisions.

6. Discussion and conclusions

In this paper, we have introduced Sticky CSMA/CA, a MAC framework that can exploit regularityin real-time traffic to achieve implicit synchroniza-tion over wireless mesh networks, thereby enablingTDM-like QoS for such applications. For the mixof voice and data traffic considered in our numericalresults, Sticky CSMA/CA locks onto a periodictransmission schedule that matches the period ofthe VoIP flows. This periodicity in transmissionenables each node to accurately predict the real-timetraffic generated by nodes that are capable of inter-fering with it, thereby drastically reducing thecontention seen by real-time flows. Thus, StickyCSMA/CA mechanism provides quasi-deterministicQoS to a real-time flow once it is established, whileits slot reconfiguration mechanism enables robustrecovery from losses due to unexpected interferencepatterns resulting from hidden terminals. Throughanalysis and simulations, we have shown that StickyCSMA/CA significantly increases the total voicecall carrying capacity of mesh networks as com-pared to the conventional CSMA/CA basedschemes such as IEEE 802.11b/e. While we imposea strict priority for real-time flows over delay-insen-sitive data traffic, the new contention mechanisms


introduced for data packet transmissions ensurethat the data traffic effectively fills the bandwidthleft over after serving the VoIP flows. Our simula-tion results for different scenarios show that thetotal voice call carrying capacity of Sticky CSMA/CA is much higher than that of IEEE 802.11b/e,while its data throughput remains comparable tothat of IEEE 802.11b/e.

It is worth pointing out that, while IEEE 802.11eis intended to provide QoS to real-time flows by giv-ing them priority over delay-insensitive traffic, thisservice differentiation has little effect when a numberof real-time flows of equal priority contend for themedium. In particular, our performance evaluationsover the grid topology (characterized by high inter-ference and contention for the medium) demon-strate that IEEE 802.11e yields much lower voicecall carrying capacity and higher delay/delay jitterthan Sticky CSMA/CA. In contrast, StickyCSMA/CA achieves fine-grained resource alloca-tion among nodes that have real-time packets totransmit, while enforcing strict service differentia-tion among real-time and delay-insensitive traffic.

While our performance evaluation was restrictedto VoIP flows that naturally exhibit periodicity thatcan be exploited by Sticky CSMA/CA, the promis-ing results that we obtain motivate research onextending these ideas to real-time flows that maynot be naturally periodic. In particular, our ongoingresearch focuses on efficient support of variable bitrate (VBR) traffic such as video using StickyCSMA/CA. We propose to impose a basic unit ofresource allocation in the network as a periodic con-nection with packets of a fixed size, thus satisfyingthe basic requirements for Sticky CSMA/CA: impli-cit synchronization and prediction of network activ-ity via Carrier Sense Tables. Nodes initiate andterminate such connections based on their band-width needs, analogous to the arrival and departureof voice connections in the scenario considered inthis paper. The specific criteria for opening and clos-ing connections depends on the applications beingsupported. For VBR video, for example, it may bepossible to use an approach similar to RenegotiatedCBR, which was proposed in [30] for supportingvideo over ATM networks. This approach is basedon splitting a VBR flow into a base CBR streamwith minimal bandwidth, and renegotiating foradditional bandwidth in terms of additional CBRstreams whenever required. The buffer occupancylevels at the application layer can be used to deducethe need for additional bandwidth and to estimate

the bandwidth required. Significant cross-layerinteraction is required to be able to provide a deter-ministic QoS. This interaction includes admissioncontrol to decide whether or not a new VBR flowshould be admitted while ensuring that bandwidthrenegotiations for existing flows are successful,end-to-end signaling for flow establishment andbandwidth renegotiations, and buffering at theapplication layer along with an intelligent algorithmto determine if bandwidth renegotiation is requiredfor an existing VBR flow. The efficacy of such tech-niques for supporting VBR video over wireless meshnetworks is the subject of our current research.

More broadly, Sticky CSMA/CA could poten-tially be used as a fundamental building block of across-layer architecture to provide comprehensiveQoS solutions for wireless mesh networks. Specifi-cally, Sticky CSMA/CA can be exploited by intelli-gent routing, transport and application layerschemes to provide deterministic QoS guaranteesto applications with different QoS requirements.For example, a wireless mesh routing protocol canleverage the medium utilization information main-tained in the CS Tables of nodes to search forQoS capable routes that can help in efficientresource utilization. Admission control algorithmscan utilize the CS Table maintained at each nodeto obtain an accurate estimate of the availableresources in the network, to be able to decidewhether a new flow should be admitted, and to pro-vide deterministic QoS guarantees to existing flows.In order to enforce the desired resource allocationamong different traffic classes and to prevent starva-tion, thresholds on maximum allocated bandwidthfor each traffic class can be decided by the networkdesigner and can be enforced through admissioncontrol schemes. By enabling local (per-hop)resource reservations, supporting service differentia-tion, and providing an estimate of availableresources in the network, Sticky CSMA/CA sup-ports QoS constrained routing and admission con-trol, which are essential components of any QoSarchitecture.

Acknowledgments

This work was supported by the National SceinceFoundation under grants ANI-0220118 and NSFNeTS Award CNS-0435527, and by the Institutefor Collaborative Biotechnologies under grantDAAD19-03-D-0004 from the US Army ResearchOffice.


References

[1] Y. Sun, I. Sheriff, E.M. Belding-Royer, K.C. Almeroth,An experimental study of multimedia traffic performancein mesh networks, in: WiTMeMo’05: Workshop on Wire-less Traffic Measurements and Modeling, Seattle, WA,2005.

[2] F. Kelly, Modelling communication networks, present andfuture, Philosophical Transactions of the Royal Society ofLondon A354 (1707) (1996) 437–463.

[3] T. Henderson, D. Kotz, I. Abyzov, The changing usage of amature campus-wide wireless network, in: Proc. 10th AnnualInternational Conference on Mobile Computing and Net-working (MobiCom’04), Philadelphia, PA, 2004.

[4] Cisco 7900 Series IP Phones (2006). URL http://www.cis-co.com/en/U8/products/hw/phones/ps379/.

[5] Wireless LAN Medium Access Control (MAC) and PhysicalLayer (PHY) Specifications, IEEE Std. 802.11, January1997.

[6] Wireless LAN Medium Access Control (MAC) and PhysicalLayer (PHY) specifications Amendment 8: Medium AccessControl (MAC) Quality of Service Enhancements, IEEE Std.802.11e-2005, November 2005.

[7] J.L. Sobrinho, A.S. Krishnakumar, Real-time traffic over theIEEE 802.11 medium access control layer, Bell Labs Tech-nical Journal 1 (2) (1996) 172–187.

[8] N.H. Vaidya, P. Bahl, S. Gupta, Distributed fair schedulingin a wireless LAN, in: Proc. 6th Annual InternationalConference on Mobile Computing and Networking (Mobi-Com’00), Boston, MA, 2000.

[9] S.J. Golestani, A self-clocked fair queueing scheme forbroadband applications, in: Proc. IEEE INFOCOM’94,Toronto, Canada, 1994.

[10] C.R. Lin, M. Gerla, Asynchronous multimedia multihopwireless networks, in: Proc. IEEE INFOCOM’97, Kobe,Japan, 1997.

[11] A. Rao, I. Stoica, An overlay MAC layer for 802.11networks, in: MobiSys’05: Proc. of the 3rd InternationalConference on Mobile Systems, Applications, and Services,Seattle, WA, 2005.

[12] I. Rhee, A. Warrier, M. Aia, J. Min, Z-MAC: a hybrid MACfor wireless sensor networks, in: SenSys’05: Proc. of the 3rdInternational Conference on Embedded Networked SensorSystems, San Diego, CA, 2005.

[13] C.-H. Lin, H. Dong, U. Madhow, A. Gersho, Supportingreal-time speech on wireless ad hoc networks: inter-packetredundancy, path diversity, and multiple description coding,in: Proc. 2nd ACM International Workshop on WirelessMobile Applications and Services on WLAN Hotspots(WMASH’04), Philadelphia, PA, 2004.

[14] A. Vasan, R. Ramjee, T. Woo, ECHOS enhanced capacity802.11 Hotspots, in: Proc. IEEE INFOCOM’05, Miami, FL,2005.

[15] Wireless LAN Medium Access Control (MAC) and PhysicalLayer (PHY) Specifications: Higher Speed Physical Layer(PHY) Extension in the 2.4GHz Band, IEEE Std. 802.11b/D5.0, April 1999.

[16] K. Sanzgiri, I.D. Chakeres, E.M. Belding-Royer, Determin-ing intra-flow contention along multihop paths in wirelessnetworks, in: Proc. 1st Annual International Conference onBroadband Networks (BROADNETS’04), San Jose, CA,2004.

[17] G. Bianchi, Performance analysis of the IEEE 802.11distributed coordination function, IEEE Journal on SelectedAreas in Communications 18 (3) (2000) 535–547.

[18] Y.C. Tay, K.C. Chua, A capacity analysis for the IEEE802.11 MAC protocol, Wireless Networks 7 (2) (2001) 159–171.

[19] H.S. Chhaya, S. Gupta, Performance modeling of asynchro-nous data-transfer methods of IEEE 802.11 MAC protocol,Wireless Networks 3 (3) (1997) 217–234.

[20] M.M. Carvalho, J.J. Garcia-Luna-Aceves, A scalable modelfor channel access protocols in multihop ad hoc networks, in:Proc. 10th Annual International Conference on MobileComputing and Networking (MobiCom’04), Philadelphia,PA, 2004.

[21] Qualnet Network Simulator, version 3.9 (2005). URL http://www.scalable-networks.com.

[22] MeshDynamics-High Performance Metro Mesh Networkswith Structured Mesh (TM) (2006). URL http://www.meshdynamics.com.

[23] A.P. Markopoulou, F.A. Tobagi, M.J. Karam, Assessing thequality of voice communications over internet backbones,IEEE/ACM Transactions on Networking 11 (5) (2003) 747–760.

[24] R. Zopf, Real-time Transport Protocol (RTP) Payload forComfort Noise (CN), RFC 3389 (September 2002).

[25] P.T. Brady, A model for generating on–off speech patterns intwo-way conversation, The Bell System Technical Journal 48(9) (1969) 2445–2472.

[30] M. Grossglauser, S. Keshav, D.N.C. Tse, RCBR: a simpleand efficient service for multiple time-scale traffic, IEEE/ACM Transactions on Networking 5 (6) (1997) 741–755..

Sumit Singh received the bachelors degreein Electrical Engineering from the IndianInstitute of Technology, Bombay, in2002. He worked as a software engineer inthe Network Systems Division at Sam-sung Electronics, Bangalore from 2002 to2004. He is currently pursuing the Ph.D.degree in Electrical and Computer Engi-neering at the University of California,Santa Barbara, where he is working withProf. Upamanyu Madhow and Prof.

Elizabeth M. Belding-Royer. His research interests include qualityof service and scalability issues in wireless networks.

Prashanth Aravinda Kumar Acharya

received his B.E. degree in ComputerEngineering from National Institute ofTechnology, Suratkal, India in 2002.From 2002 to 2004, he was with TexasInstruments, Bangalore, India as aSoftware Design Engineer. He is cur-rently pursuing his Ph.D. in ComputerScience at the University of California,Santa Barbara. He is a member of theMoment Research Lab directed by Prof.

Elizabeth M. Belding-Royer. His research interests includeQuality of Service, multi-media streaming and multi-radio wire-
less mesh networks.
http://www.cisco.com/en/U8/products/hw/phones/ps379/

http://www.cisco.com/en/U8/products/hw/phones/ps379/

http://www.scalable-networks.com

http://www.scalable-networks.com

http://www.meshdynamics.com

http://www.meshdynamics.com

Networks 5 (2007) 744–768

Upamanyu Madhow received his bache-lor’s degree in electrical engineering fromthe Indian Institute of Technology,Kanpur, in 1985. He received the M.S.and Ph.D. degrees in electrical engineer-ing from the University of Illinois,Urbana-Champaign in 1987 and 1990,respectively.

From 1990 to 1991, he was a VisitingAssistant Professor at the University ofIllinois. From 1991 to 1994, he was a

research scientist at Bell Communications Research, Morristown,NJ. From 1994 to 1999, he was on the faculty of the Department

768 S. Singh et al. / Ad Hoc

of Electrical and Computer Engineering at the University ofIllinois, Urbana-Champaign. Since December 1999, he has beenwith the Department of Electrical and Computer Engineering atthe University of California, Santa Barbara, where he is currentlya Professor. His research interests are in communication systemsand networking, with current emphasis on wireless communica-tion, sensor networks and multimedia security.

He is a Fellow of the IEEE, and recipient of the NSFCAREER award. He has served as an Associate Editor for

Spread Spectrum for the IEEE Transactions on Communica-tions, and as an Associate Editor for Detection and Estimationfor the IEEE Transactions on Information Theory.

Elizabeth M. Belding-Royer is an Asso-ciate Professor in the Department ofComputer Science at the University ofCalifornia, Santa Barbara. Her researchfocuses on mobile networking, specifi-cally ad hoc and mesh networks, multi-media, monitoring, and advanced servicesupport. She is the author of over 50papers related to mobile networking andhas served on over 40 program commit-tees for networking conferences. She was

the TPC Co-Chair of ACM MobiCom 2005 and IEEE SECON2005, and is currently on the editorial board for the IEEE
Transactions on Mobile Computing. She is the recipient of anNSF CAREER award, and a 2002 Technology Review 100award, awarded to the world’s top young investigators. Seehttp://www.cs.ucsb.edu/~ebelding for further details.
http://www.cs.ucsb.edu/~ebelding

sticky csma/ca: implicit synchronization and real-time qos in … · 2009-02-19 · sticky csma/ca:...

Documents