Performance of TCP/IP Over Next Generation Broadband Wireless Access Networks

Ivana Stojanovic, Manish Airy, David Gesbert1, Huzur Saran2

Iospan Wireless Inc., 3099.North First Street , San Jose , CA 95134 [ivana, mairy, gesbert, hsaran]@iospanwireless.com

1 Also with Dept. of Informatics, University of Oslo, Norway.2 Also with Dept. of Computer Science, Indian Institute of Technology, Delhi.

Abstract

This paper presents performance evaluation results forInternet access over the next generation of broadbandwireless access networks (2G-BWA). The challenge forthese, mostly fixed access technologies stems from nonline-of-sight deployment with high coverage andspectrum efficiency combined with link performancecomparable to that of competing wired technologies(DSL, Cable modem). We present an overview of thefeatures at the physical (PHY) and medium access control(MAC) layers which are key to a successful 2G-BWAdesign and demonstrate the impact of such features onend-to-end link performance with typical TCP/IPapplications and wireless channel models.

Keywords

Broadband wireless access, MIMO, adaptive modulation,link layer ARQ protocol, TCP/IP.

1. Introduction

The demand for broadband Internet access is booming.Announced delays in the deployment of third generationhigh speed wireless networks (WCDMA and cdma2000are unlikely to be widely available before 2003-2004), aswell as slow progress in satisfying demand for wiredsolutions such as x-DSL and Cable modems place highexpectations on alternative access technologies such asfixed wireless. Fixed wireless technologies of focus hereaddress the lower frequency bands (MMDS, 3.5GHz)dedicated to mass market – residential, small officeshome office (SoHo) use. Current internet broadbandwireless access technologies are based on the existence ofa line of sight link between the subscriber’s unit and theaccess point (base station). This assumption allows toavoid multipath fading and equalization, however it putsvery stringent limits on the scalability and ubiquity of thetechnology, preventing low cost mass deployment andalso preventing any future evolution to support mobility.Figure 1 illustrates the main network components ofBWA systems.

��� ���������� �������������������������� �!�"�#��$&%'��(*)+%�,.-�%����������#�

In order to successfully respond to the demandand carry their future share of access traffic, BWAsystems must evolve into a second generation technologythat can give high TCP/IP performance over severelyfading channels caused by non line-of-sight deployment.The performance needs to be at least comparable to wiredaccess: DSL and Cable modem. BWA systems need,above all, to make an efficient use of spectrum, while atthe same time maintaining user friendliness especiallyduring installation and low cost of operation. Typicaloverall figure-of-merit requirements for these upcomingsystems are shown in Table 1.

Figure of Merit Requirement

Aggregate Rates 4-7 Mb/s

Spectrum Efficiency 2 Bits/Hz/BTS

Coverage 4-6 miles (90% area)

Latency Comparable to DSL

Link reliability 0.999

Table 1: Requirements for 2G BWA.

Non line-of-sight broadband wireless channels areprone to frequency selective and time selective fading,[1]. In addition, because of the quasi-stationary nature ofthe subscriber unit, the Doppler spread in the channel is

low (less than 1Hz) causing fades to typically last longerthan wireless mobile applications (up to a few seconds).Fades are due to moving reflecting objects in theenvironment such as tree leaves, cars and pedestrians,rather than the movements of the access unit itself. Toprovide good link performance at TCP/IP level over suchchannels, fade mitigation at either the PHY or MAClayers becomes absolutely critical. A number of emergingtechnologies that satisfy this need have been investigatedfor use in the AirBurst (a proprietary system beingdeveloped by Iospan Wireless Inc.) BWA system. Notethat only generic or standard versions of the systemfeatures are mentioned here. Actual implementationdetails are system proprietary however. This paper firstpresents a tutorial background on the windowing and lossrecovery mechanisms of TCP and address difficultiesfaced by TCP over unreliable and lossy wireless links.Then, a selection of technologies, both at the MAC andPHY, that are suited for non line-of-sight BWA arepresented. Next, end-to-end link performance (includingTCP/IP) for a typical 2G-BWA system exposed totime/frequency fading channels is evaluated. In particular,the focus is on the respective impact of key PHY andMAC layer features on end-to-end TCP/IP performanceover these channels. We conclude with designrecommendations for 2G-BWA.

2. TCP Slow Start and Congestion Avoidance

In this section, we recall the TCP windowing and lossrecovery mechanisms and emphasize differences betweenseveral widely used versions: Tahoe, Reno and NewReno.

TCP is a dynamic window based flow controlprotocol. The window size, denoted by W(t), varies inresponse to packet acknowledgments (ACKs) and lossdetection. While the receive processes are the same for allTCP versions, their transmit processes and windowadjustment algorithm in response to loss detection aredifferent [2].

At any given time, t, the TCP transmittermaintains several variables for each connection: lowerwindow edge, congestion window and slow startthreshold. Lower window edge X(t), denotes that all datapackets numbered up to X(t) – 1 have been transmittedand acknowledged. Next eligible packet the transmittercan send is numbered X(t). Each ACK receipt causes X(t)to advance by the amount of acknowledged data and,thus, lower edge window evolution is monotonically non-decreasing time function. The transmitter’s congestionwindow, W(t), defines the maximum amount of datapackets the transmitter is permitted to send, starting fromX(t). Control mechanism upon ACK reception allows W(t)to increase or decrease within Wmax boundaries. Duringconnection set up, the receiver advertises a maximumwindow size, Wmax, so that the transmitter does notadvance instantaneous window W(t) beyond this value

during whole connection period. Slow start thresholdWth(t) controls the increments in W(t). If W(t) < Wth(t)transmission is assumed to be in slow start phase, whereeach ACK causes W(t) to be incremented by 1. This rapidgrowth switches to slow increase, once the windowcrosses slow start threshold, that is, when W(t)>=Wth(t).This phase, called congestion avoidance phase, aims tocautiously probe for extra bandwidth. Each ACKreception increments window size W(t) by 1/W(t).

On the receiver side, packets can be accepted outof order, but will be delivered only in sequence tointended TCP user. The destination returns cumulativeacknowledgement for every correctly received packet.Each ACK carries next in-order packet sequence numberexpected at the receiver, which is the first among thepackets required to complete the in-sequence delivery.Thus, a single packet loss can be detected if the same“next expected” number is received in all consecutiveacknowledgements of packets successfully transmittedafter the lost packet. Reaction on packet loss detectiondepends upon the TCP version. In the simplest case, TCPOldTahoe, the TCP transmitter continuously sendspackets till the congestion window is exhausted and thenwaits for timer expiry. The sender maintains a timer onlyfor the last transmitted packet, that is, each time a newpacket is sent, the transmitter resets the already runningtransmission timer. The timer is set for a round triptimeout (rto) value that is derived from a smoothedround-trip time (rtt) estimator applying the low pass filterover the last estimate and the current measurement of thertt [3]. The timeout values are set only in multiples of atimer granularity: in a typical implementation the timergranularity is 500ms. According to the basic recoveryalgorithm, upon timeout window parameters are adjusted- Wth(t+) is set to W(t)/2 and W(t) is reset to 1, andretransmission is initiated from the first missing packet.

More sophisticated TCP protocols such as Tahoeand Reno, implement fast retransmit and loss recoveryprocedures. The Fast retransmit procedure takesadvantage of additional information that duplicateacknowledgements carry. The transmitter waits forseveral duplicate acks to arrive (typically three) toaccount for possible network delay and packet reordering.Then, the missing packet is retransmitted before the timerexpiry. In subsequent loss recovery, TCP Tahoetransmitter behaves as if a timeout has occurred andresorts to the basic recovery algorithm.

In the case of Reno, only the lost packet is fastretransmitted and loss recovery procedure is handleddifferently. Congestion window recorded at this time isreferred as loss window. After Nth duplicate ACK attimet, sender adjusts slow start threshold Wth(t

+) to W(t)/2and W(t+) to Wth(t

+) + N. Congestion window W(t) isadditionally incremented by N to account for packets thathave successfully left the network and produced duplicateACKs. While waiting for the ACK for the first

retransmission, any successfully received outstandingpacket produces duplicate ACK, which further incrementsW(t) by 1. If single packet were lost, the packetretransmission would complete loss recovery. At thistime, congestion window W(t) is set to Wth(t), and thetransmission resumes according to the normal windowcontrol algorithm. However, the algorithm can experiencea possible stall time if multiple packet losses occur withinthe loss window. In this case, the congestion windowmight close before sufficient number of duplicate ACKshas been generated to trigger multiple fast retransmits. Inthat case algorithm waits for timer expiry. The Renorecovery algorithm outperforms Tahoe version, but ispessimistic and suffers performance degradation in caseof multiple losses within the loss window [2].

New Reno attempts to fix multiple lossdrawback by setting fast retransmit and congestionwindow adaptation as in Reno and implementing the lossrecovery mechanism as in Tahoe. In this paper, we focusmainly on TCP Reno.

2.1. Performance of TCP on wireless links

From the discussion above we see that TCP is notdesigned for lossy links. The TCP protocol has beendeveloped for wired networks to deal with networkcongestion, rather than non-negligible random losses. Onthe contrary, wireless link impairments raise the packeterror probability several orders of magnitude. The TCPsending rate drops upon loss detection. Afterwards, TCPloss recovery kicks in and offered throughput increasesgradually. Congestion window increase, clocked byarrival of acknowledgements, largely depends on roundtrip time. In other words, if round trip time is large andthe congestion window decreased frequently due tomultiple losses, the TCP connection may not be able tofully utilize available channel bandwidth. Although someof the previous studies show that different versions ofTCP protocol perform better or worse in response towireless lossy link with correlated errors, none of themgives satisfactory performance. This calls for newtechniques that will allow reliable transmission inheterogeneous networks comprising both wired andwireless links. Comparative analysis of several higherlayer schemes proposed to alleviate the effects of non-congestion related losses has been conducted in [4]. Theauthors classify the schemes in two categories accordingto the approach they take to improve TCP performance inlossy systems. The first category includes link-layerprotocols, split connection approaches and TCP awarelink layer schemes. These protocols try to locally solvethe problem and hide wireless losses from the TCPsender. As a result, wireless hops appear as high qualitylinks with reduced effective bandwidth. The second classof techniques attempts to make the sender aware of theexistence of non-congestion related losses. The idea is

that sender restricts itself from invoking congestionavoidance algorithm in response to such events. However,second class schemes require undesirable change in TCPend-to-end algorithm. Comparison analysis shows thatlocal protocols outperform the second class techniques,[4]. Moreover, reliable link layer schemes offer higherthroughput than split connection approach, while at thesame time preserve end-to-end protocol semantics.

3. Design features for 2G-BWA networks

In this section, the focus is on BWA system designfeatures that attempt to overcome unreliability introducedby time and frequency selective fading and interference-prone channels. Both PHY and MAC layer features arepresented and TCP fixes that have been proposed to dealwith wireless channel behavior are ignored. The focushere is on a set of selected features that the authors viewas particularly critical or relevant. The MAC layer ensuresa low resultant error rate seen by TCP. The PHY layerimproves link quality and enables higher data rates.

3.1. MAC Layer

Automatic Retransmission/Fragmentation (ARQ/F)

ARQ/F is widely considered a key tool to deal with burstyerrors occurring in wireless channels. A low-latencyacknowledgment and retransmission mechanism isimplemented at the MAC layer between the subscriberunit (or CPE) and the BTS. The MAC layer fragments IPpackets into ’atomic’ data units (ADU). Only lost ADUsare retransmitted. A technique complementary to coding,ARQ/F only introduces redundancy during the fraction oftime when data gets corrupted. A finite buffer makesARQ/F efficient in dealing with short as well as longfades. ARQ/F removes packet errors at the cost of onlymoderate additional link latency. ARQ/F-based systemscan be designed to operate at high BER levels while stillproviding satisfactory TCP/IP performance, as shownlater in the results section.

At a transmitter, PDUs received for transmissionfrom higher protocol layers are fragmented into ADUsand stored in a transmit buffer. These ADUs are normallytransmitted sequentially, with each outgoing ADU beinguniquely identified by its sequence number. Each ADU iskept stored in the transmit buffer even after it has beentransmitted. At a receiver, uncorrupted incoming ADUsare stored in a reassembly buffer. Whenever all of theADUs comprising a PDU are received, the reassembledPDU is passed up to the higher layer protocols. However,there may be situations where one or more ADUscomprising a PDU is received corrupted and needs to beretransmitted. In order to notify the transmitter aboutthese corrupted ADUs, a receiver periodically generatesan Acknowledgement message (ACK). This messageenumerates the sequence numbers of ADUs that are in the

receiver’s reassembly buffers. When a transmitterreceives an ACK, it retransmits the “missing” ADUs fromits transmit buffers. These retransmitted ADUs are givenhigher priority over regular ADU transmissions. ADUsthat have been successfully received at the receiver arefreed from the transmit buffer.

The ARQ protocol is based on a sliding windowmechanism, that is, each flow is allowed to transmit onlyif it has no more than a window size (W) worth ofunacknowledged ADUs. More formally, the followinginequality must always hold:

T(N) – T(U) <= W if T(N) >= T(U) (1)MAX_SEQ_NUM -T(U)+T(N) <= W if T(N) < T(U) (2)

where MAX_SEQ_NUM denotes the size of the sequencenumber space, T(N) denotes the sequence number of thenext in-sequence RLC ADU eligible to be transmitted at atransmitter and T(U) denotes the sequence number of the“oldest” RLC ADU that has not been positivelyacknowledged by the receiver. The window size W shallbe configurable at the time of link initialization but it canbe no larger than MAX_SEQ_NUM/2 ADUs. This limitarises from the size of the sequence number space.

Weighted Round Robin Scheduling

At the beginning of every frame, the BTS transmits aMAP message that allows each NAU to determinewhether it is supposed to receive or transmit ADUs duringthat frame.

For a downlink or uplink transmission, eachMAP allows a particular NAU to determine which timeslots it should decode or transmit in, respectively, andwhat modulation/coding is currently being used for theADUs.

The scheduling policy used to compute the uplinkand downlink transmission schedules at the BTS has tomeet the following requirements and constraints:

• Ensure that CBR service flows get allocated aconstant bandwidth (with acceptable jitter)

• Ensure that UBR service flows share availablebandwidth in a fair manner

• Support spatial multiplexing

• Support channel dependent scheduling. That is,compute transmission schedule based on theknowledge of modulation/coding being used overcurrent channel state

• Provide ability to temporarily shut off transmissionto a NAU with a bad link

Adaptive Modulation (AM)

Adaptive modulation and coding lets a user adapt its datarate as a function of channel conditions (e.g. SINR) andhas been popularized in the EDGE cellular standard [5].AM allows several times rate improvement by exploiting

all margins of signal to noise ratio available at anytime/location. The idea behind location adaptivemodulation is depicted at Figure 2. By comparison, a non-adaptive system is deployed with most conservativemodulation/coding for all users in order to preserve goodcoverage and frequency reuse. Note that self-adaptivity(to even slowly changing conditions) by the CPE at onelocation is a key requirement to extract significantcapacity gain. Additional gain is obtained by adaptingfaster. For TCP/IP, AM will result in larger pipe size and

QPSK

16 QAM

64 QAM

� � � � � � / � � 0 � - � � 1 � 2 % 0 � � � � � % � �

additional statistical multiplexing. To detect anappropriate modulation/coding level, the MAC layer usesa set of statistics obtained from the PHY layer. The set ofstatistics involves a combination of Bit Error Rate, PacketError Rate and signal to noise ratio. The details are notdisclosed in this paper. However, it is assumed that eachuser in the cell is assigned a modulation and coding levelthat is a function of individual channel conditions and thatguarantees a given target Packet Error Rate.

3.2. Physical Layer

Spatial Diversity (SD)

SD is obtained through the use of multi-element antennasat either the BTS and/or the CPE. Antenna combining isused to deal with fading, Figure 3, and allows advancedinterference canceling algorithms. The basic idea ofdiversity combining is that several uncorrelated fadingelements are much less likely to fade simultaneously thana single element. Use of SD can reduce the SNRrequirement by 10-15 dB with no loss of TCP/IPperformance. This is exploited to extend the coverage,increase data rates, and most importantly to allow a tightfrequency reuse that will work in arbitrary terrain (flat ornot). Recent measurement campaigns for BWA suggestthat decorrelation is achieved with 1-2 wavelengths of

element separation; less spacing is required with dual-polarized antenna elements [6].

��� ���"�#�435�627�8������-����4���"�����"���'�-��#%�1���0�����#%:9:�"�#�;��%�����)<�%�=0���1��#���#����>?���A@.%��AB�C.DE)+F5�'���"�������

The signal processing technique used to combinethe antennas can vary. Maximum Ratio Combining(MRC), Minimum Mean Square Error (MMSE)combining are typical examples of combining algorithmswhen the channel coefficients are known (e.g. at thereceiver) [7]. On transmission, where the exact channelcoefficients cannot be known unless a bandwidthconsuming feedback link is used, a blind transmit spatialdiversity must be used. Examples are space-time codingalgorithms recently developed [8]. Delay-diversity isused for transmitting and MRC is used for receiving.

MIMO and Spatial Multiplexing (SM)

In the case of multi-element antennas at both ends(multiple-input multiple-output or ’MIMO’), an additionaldata rate increase can be obtained by exploitingsimultaneous transmission over the eigenmodes of thechannel as seen in [7] and [9]. A simple spatialmultiplexing algorithm works as follows (more advancedvariations combined with coding are typically used inpractice): a high rate signal is demultiplexed into set ofindependent bit streams which are independentlytransmitted by the multiple antennas. The signals aremixed with each other in the channel since they occupythe same time and frequency resource. At the receiver, themultiple antennas are used to first learn the spatialsignature (channel) of each stream using trainingsequences then combined so as to separate the differentbit streams from each other. The streams are finallymultiplexed to offer the original high rate signal. Thisoperation consumes only the bandwidth normally neededto transmit one single bit stream. The increase in data rateis achieved at the expense of increased interference. Theprocedure is illustrated in Figure 4.

When there is only one usable eigenmode due torank deficiency of the MIMO channel, the streams are notseparable and the system must fall back to a diversity

scheme [9], using proprietary adaptation algorithm. Thistechnology is particularly well suited for BWA because,

SIGNALPROCESS

SIGNALPROCESS

C1

C3

C2

A1

A3

A2

B3

B2

B1

A1

C1

B3

B2

B1

A3

A2

C3

C2

b2 b5 ...

b1 b4 ...

...b6b3

b1 b4 ...

...b6b3

b1 b6b5b4b2 b3 .... b1 b6b5b4b2 b3 ....b2 b5 ...

� � � � � � G � D � , - � � = � � 0 D - � � � � � 2 � � � � - � � H � � � � ) F � , � $ � � F� F � � � � � � � � � � � � I F � � � � 1 � � � F � � � = % � 0 0 � � � � � � �

� , - � % 1 � , � � � �

unlike handsets, compact multi-element subscriber unitscan be designed with 2 or 3 elements. The gain in datarate is proportional to the minimum of the number oftransmit and receive antennas. The impact of thistechnology on TCP/IP performance is also examined here.

��� �;���#�4J5�?C.��K�2L$&����F?)+%�05�����M��������� =N%:�#,.�4,.��������-5�'��F?������%���#%'���#)+�4%�=*�:������O =N�'05������0���1��#���#����>��

Frequency Diversity

Broadband transmission over multipath channelsintroduces frequency selective fading, as illustrated inFigure 5. Mechanisms that spread information bits overthe entire signal band will mitigate the effect of fadingoccurring at certain frequencies. An example of suchmultipath-friendly mechanisms is frequency-codedmulticarrier (e.g. OFDM) modulation [10]. This techniquerequires lower SNR to achieve the same TCP/IPperformance.

4. End-to-end performance evaluation

4.1. Simulation Methodology

The main objective of end-to-end simulator is to captureall aspects of system design and examine their impact onTCP/IP. The simulations are organized into severalhierarchical layers, where each module sees the abstractedbehavior of a layer below. The simulation is constrainedfrom the bottom by BWA channel models and from thetop by TCP/IP usage models. The three main designlayers with several degrees of abstraction andimplementation detail are captured, namely Transport,MAC and PHY layers. Different tools are used toimplement each layer.

Network Simulator (NS) [11] is used to captureuser behavior and TCP-based transport. Detailed MAClayer functionality is developed in C/C++ and interfacedto NS. PHY layer behavior is abstracted through errorand data rate traces. These traces are obtained throughseveral levels of PHY simulation and reflect performanceof coding, modulation and multiple antennae processingtechniques over realistic fixed broadband wirelesschannels. Realistic channel models are developed fromfixed MIMO channel measurements.

The NS module models one target multi-userTDMA-based cell. However, the interference caused byother cells is also captured through the data and error ratetraces, assuming a 2x3 frequency reuse. Wired backbonenetwork is represented as a set of web servers (trafficsources) connected to the BTS through high bandwidth,high propagation delay and low BER links. The wiredlinks converge to the BTS (Figure 6) which maintainsseparate queues for each network access unit (NAU).Multiple TCP connections, originating from differentsevers can get routed to the same queue modeling abusiness customer or residential users with multipleopened connections. Packets are transmitted over the airand are subject to errors as predicted by PHY simulations.A connection from the BTS to any particular NAU isrepresented with a link whose average quality is dictatedby long-term channel condition specific for eachsubscriber location.

4.2. MAC Model

The MAC layer model captures detailed design featuressuch as segmentation and re-assembly of IP packets,scheduling, ARQ, UL access via contention and linkadaptation.

Received IP packets are fragmented into atomicdata units (ADUs) and resulting smaller packets are storedfor transmission. Receive module ensures in-orderdelivery to higher (TCP) layers. An ADU scheduled fortransmission can be lost with a certain probabilitydetermined by the instantaneous channel state andtransmit mode. Lost ADUs are identified and marked for

retransmission. The scheduler enforces a weighted roundrobin policy. Allocations are made for a number offrequency blocks considering transmit mode and qualityof service requirement of a particular user. On reverselink, from NAU to BTS, initial access is obtained througha contention channel. Once data transfer starts, additionalbandwidth requests are piggybacked on the data ADUs. Ifenabled, both UL and DL are subject to transmit modeadaptation based on SNR measurements made at thereceiver.

NAU NResidential

NAU 1Business

Downlink

Uplink

BTS

Tx/Rx for NAU 1

Tx/Rx for NAU N

(Sink/Server ) 1

(Sink/Server) M

(Sink/Server) ki

(Server/Sink) 1

(Server/Sink) ki

(Server /Sink) M

AirBurst (MAC/ARQ/PHY)

�"� �;���#�4P5�?D���,.�����'����%��AQR��1�����%���,.�S���

4.3. PHY Model

The PHY performance is summarized through estimationof modem setpoints for each transmit mode. A transmitmode is comprised of a coding mode (choice of MQAMand coding rate) and antenna mode (diversity or spatialmultiplexing). The modem setpoint for a particular mode,as defined by the modulation, coding and number oftransmit antennas, is defined as the signal-to-noise ratio(SNR) required to achieve a desired pre-ARQ operatingerror rate. Here, two cases can be distinguished. If codingmode were static over time, modem setpoint wouldrepresent the average SNR level over a long period oftime attained at one receive antenna element. In thesecond case, link adaptation tracks channel variations,allowing maximization of data rate whenever possiblethrough selection of ‘optimal’ transmit mode. In this caseboth packet error rate trace, as well as correspondingmode trace are generated and supplied as input to the end-to-end simulator.

4.4. Wireless Channel ModelPhysical properties of wireless communication

channels exhibit lots of impairments unknown to their‘wired’ counterparts. In a wireless system, a signaltransmitted into the channel interacts with theenvironment in many complex ways. Signal propagationis subject to scattering, reflection, refraction or diffractionof surrounding objects, such as buildings, hills, streets,trees and moving mobiles, as shown on Figure 7. Thereceived signal is much weaker due to wireless

propagation impairments such as path loss, shadowingand fast fading. Path loss is the drop in radiated signalpower due to the distance from the transmitter to thereceiver, environment, water, foliage, etc. Ideal free spacepropagation follows inverse square law spreading with thedistance, whereas in real cellular environment path lossexponent varies from 2.5 to 5.

TVUXW YZW [ UX[ \ ]�^

_`]�a UXbc[ ]�deU W Yf�YXg']�[ Y

_V]�a UXb�[ ]+gN]cd�\ b Y

�"� �;���#�4h5� i � 0 � % $ � 1 � - � % - � � � � � % � $ � � F 0 � = = � � ) � � % � j� � = � � ) � � % � � � 0 , � � � � - � � � � = � � ) � � % �

Shadowing, or slow fading, comes from signalblocking by large obstructions such as buildings and hills,that are positioned between base station and accessterminal. Multipath fading plays a central role indetermining the nature of wireless channel. The result ofsignal interactions with numerous physical objects in itspath to the receiver is a presence of many signalcomponents, or multipath signals, at the receiver.Multipath arrives from different directions, with differentattenuation and propagation delays resulting in receivedsignal amplitude that varies in many dimensions – time,frequency, space.

Wireless channel characteristics are measuredthrough several key parameters: Doppler spread, delayspread and angle spread.

Time selectivity, measured through Dopplerspread, comes from the fact that pure a tone is frequencydispersed due to environment change in the receiver’svicinity. In the mobile channel, received signal envelopehas Rayleigh distribution in the time domain, whichtranslates into Doppler power spectrum of the receivedsignal that has “horn“ shape. On the other hand, fadingcharacteristics of the fixed wireless channels are verydifferent from mobile case and result in “ rounded”Doppler spectrum, [12]. Techniques that exploit thisproperty are adaptive modulation (AM) that trackschannel quality over time and radio link protocols, likeARQ/F, that attempts to fix fades by retransmission oflost packets in good channel states.

mean angle ofarrival

θ

angle spreadatthe BTS array σ

path cluster 1

path cluster 2

path cluster 3

Note: each path cluster has a delay, a powerattenuationfoactor, a mean angle of arrival and an angle

spread

CPE antennas

BTS antennas

d

scatterers

k l m n o p q r s t s u v w v x p y z { | } } p ~ y � � p ~

Frequency selectivity or multipath delay spreadis caused by several distinct and delayed versions oftransmitted signal. Frequency selective implies thatchannel response depends upon the frequency. It ischaracterized by coherence bandwidth representingmaximum frequency separation for which channelresponse remains correlated. Signal components delayedby even a small fraction of symbol period gives rise tointer symbol interference. OFDM is well suited to combatfrequency selectivity either by coding over sub-carriers oradaptively loading sub-carriers with high SINR.

Space selective fading, or angle spread refers todistinct multipath angle of arrival that translates intosignal amplitude that depends on antenna location. Thisproperty allows for space and MIMO diversity techniquesin the systems with antenna arrays.

Depending on the relationship between signalparameters (bandwidth, symbol period) and channelparameters (delay spread, Doppler spread) signaldistortions can be significant or on the contrarynegligible. The level of detail about the environment achannel model must provide is highly dependent upon theradio architecture. Propagation model parameters dependupon terrain, wind speed, season/ time of the year, trafficdensity and proximity, height and beam width of transmitand receive antennas, as well as type of system underconsideration. To predict the performance of singleantenna narrow-band system, received signal powerand/or time-varying amplitude distribution may suffice.However, for emerging wideband multiple antenna arraysinformation regarding the signal multipath and angle ofarrival structure is needed. In other words, widebandMIMO channel is properly characterized by space-timechannel model, as represented on Figure 8.

The channel models used in simulations reflectspecific broadband wireless channels developed frommeasurements and published by Sprint and StanfordUniversity [12]. These models capture radio link

impairments mentioned above. We emphasize the use ofnon-line-of-site (NLOS) models with low Ricean Kfactors in order to design system with high reliabilityrequirements. The use of multiple transmit and receiveantenna calls for antenna correlation modeling. Thesystem under study assumes cross-polarized antennas.Cross-polarization gives rise to attenuation due topolarization mismatch (XPD) between transmit anddifferently polarized receive antenna. For LOScomponent we use 10dB attenuation, but for Raleighcomponent XPD loss goes as low as 4dB. This effect isincluded in PHY modeling. Key parameters for thedifferent channel models used in the analysis, SU1-SU6,are listed in Table 2.

SU1 SU2 SU3 SU4 SU5 SU6

rms delay spread 0.1 0.2 0.3 1.3 3 5.2

overall K (linear) 10 5 0 0 0 0

antenna correlation 0.7 0.5 0.25 0.25 0.25 0.25

I���9:���/��5���#>?-5�'�#�',.�#��������=�%'�4)+F5�'�����#��,.%'05�#���

4.5. Traffic Model

The exponential growth of World Wide Web users causesthe traffic originating from Web applications to dominatethe Internet both in terms of volume and active users. Thistrend, likely to continue in next years causes downstreamand upstream link asymmetry. Majority of the Internetapplications today, such as web-browsing and filedownloads, require larger bandwidth pipes on thedownstream. However, there is an increasing demand forstreaming data (audio and video) applications, videoconferencing and high quality of service (QoS) servicesthat will require networks to support more symmetrictraffic.

Accurate traffic modeling has high relevance insystem evaluation and capacity planning. In particular, weare interested to model traffic requests that mimic acertain population of real users. Literature shows thatsuch model has been difficult to develop due to a numberof unusual Web workload characteristics, [13]. Empiricalstudies of operating servers show that their workload ishighly variable, both in volume and number of openconnections. This feature indicates that proper attentionshould be paid to properties of Web streams such asrequest inter-arrivals and file sizes. An important Webtraffic characteristic is self-similarity, that is, significantvariability over a wide range of time scales. Self-similarity negatively impacts network performance andbecomes important to address.

In our system evaluation we focus on a trafficmodel that attempts to closely imitate HTTP requestsoriginating from a population of Web users. We work

with a traffic model characterized at two levels:macroscopic and microscopic.

Inter Session Arrival Time

Session

RequestsACTIVE OFF Times

SessionFiles

Figure 9. User activity model

The macroscopic description determines thebusiness to residential user penetration ratio, the numberof active connections per business NAU, and NAUactivity factors across the subscriber population. In thesimulations we assume that a set is composed of 20%business NAUs with 5 active connections each.

The microscopic description describes individualuser behavior over time. A user can request data or act asa server to a remote client. We assume data transfers onthe downlink have an average workload of 20 Kbps perconnection and 4 Kbps on uplink. Introduction of uplinktraffic diminishes link asymmetry and allows for moresymmetric loads that account for future trafficdevelopment. In addition to these data workloads, controltraffic in terms of MAC/ARQ or TCP acknowledgementsis inherently generated. The workload is averaged overseveral www sessions, including ‘ think’ time.

According to the standardized SURGE wwwusage model [13], each user starts several sessions duringan online period. User behavior during a session isdescribed as a superposition of active downloads andthink times. Tails of the download (or request) sizes aredescribed by a Pareto distribution, with the body ofdownload sizes following a lognormal distribution. Theaverage request size is 27 KBytes. The think time isPareto distributed with an average of 3.5 seconds. Themodel parameters are shown in Table 3 and illustrated inFigure 9.

RequestBody

Log-normal

RequestTail

Pareto

ActiveOFF

Pareto

Session LengthInverse

Gaussian

µ = 7.881σ = 1.339

K = 34α=1.177

K=1α=1.4

µ = 3.86σ = 6.08

Table 3: Parameters of www traffic usage pattern and filesizes. The average request size is 27kB.

5. Simulation Results

Next, simulation results to estimate the performancebenefits of key aspects of system design that enableefficient broadband wireless operation are presented. Herewe specifically address adaptive modulation (AM), spacediversity (SD) and MIMO technique, operating error rate,impact of Doppler spread and give an example of TCPresponse to wireless channel.

Adaptive modulation and coding allow thesystem to adapt transmit modes across users in a cell, as afunction of channel conditions at each location. Thehighest AM allowable rate is 12 times that of non-adaptive system deployed with the most conservativemode to preserve required coverage and frequency reuse.Diversity of transmit modes is obtained through MIMOand SD features of physical layer. On Figure 10,performance of both systems as a function of offered loadis presented. We observe that non-adaptive systemsustains several times lower workloads with even worserequest service delays. In other words, good coverage ofnon-adaptive system is bounded by quality of servicerequirement, which in return greatly reduces number ofusers the system can support.

From Figure 10 we also observe how unequaldata rate for users of disparite SNR levels reducessystem’s offered troughput. Suppose we separate usersinto N classes according to their SNR levels andcorresponding supportable rate levels. User in class nreceives data at rate Rn bits per second, where n=1,2…N.Also suppose that all users request equal traffic amounts.This is true since all users are assumed to producestatisticaly the same traffic traces.Then, relative frequencyof packets of class n, denoted by Pn, is proportional to thenumber of users in the class. Assume infinite traffic loadof each class and suppose a scheduling scheme whereslots are assigned one at a time succesively to each user.

0 500 1000 1500 2000 2500 3000 35000.4

0.5

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3

1.4

1.5Adaptive Vs. Non-adaptive System Comparison

System Offered Load [Kbps]

Ave

rage

Req

uest

Ser

vice

Del

ay

Non-adaptive System, Conservative Mode(1.1Mbps)Adaptive system - 12 Available Modes

Figure 10. Adaptive modulation offers several foldimprovement in capacity, efficiency and QoS. Note thatthe offered load does not include ~16% of upper layerprotocol overhead.

In other words, bandwidth share of a given classis proportional to the class size. Then the system’savarage data rate, which we define as throughput is

b/s ∑=

=N

1innav RPR (3)

This means that fixed size packets of lower data rateusers will have proportionally higher latencies. That is, if

latency of a B-bit packet is considered, the number ofslots consumed by each user in class n will be B/Rn, andhance the latency Ln is inversely proportional to rate Rn.

Suppose on the other hand that we require thatall users have essentially the same latency irrespective ofthe Rn they can support and that finite amount of traffic isrequested by each user within certain time window.Then, each class will consume bandwidth directlyproportional to its size and inversly proportional to itsrate. Round Robin scheduler assumption and QoSconstraint on bounded delays require system to operatebelow its full capacity. Achieved system throughputdiffers from the average data rate, presented in equation(3), due to multiplexing of users with disparate SNRlevels. Assume that each user requests B bits and thatusers of n-th class require Sn = k/Rn, k being a constant,slots to trasmit those bits. In this case we define effectivethroughput as :

s/b

R/P

1

SP

SRP

RN

1nnn

N

1nnn

N

1nnnn

n

∑∑

∑

==

= == (4)

For example, SNR distribution across the cellderived in AirBurst system with 5% average ADU errorrate, yields an average throughput of 6 Mbps whichtranslates into 4 Mbps of effective throughput. FromFigure 10, we see that after including 16% of upper layerprotocol overhead, system offered load reaches themaximum of 4 Mbps of effective user traffic at close to100% link utilization.

Above analysis indicates cost paid in throughputfor latency equalization. Allocation which is lessbandwith “unfair” can be made if delay garantees arerelaxed for the user with worse link conditions.

MIMO and SD diversity are key to enhancingsystem performance. They take advantage of rich fadingstructure of wireless channel and allow multiple transmitmodes to be deployed. To achieve the same operatingtarget error rate, physical layer simulations estimate that auser in a single antenna system (SISO) requires 12 dBadditional SINR margin compared to MIMO system with2 transmit and 3 receive elements. System analysistranslates this margin into poor coverage of SISOdeployment so that attainable cell size in a macrocellenvironment shrinks down 2.5 times. To preserve thesame large cell size, a cell edge user in SISO systemwould need to operate at 50% packet error rate,unacceptably degrading post-ARQ delays. Furthermore,adaptive modulation can not exploit higher modes due tostringent SINR requirement. While preserving the targeterror rate for a given cell, average downlink peak datarate drops almost 3 times in SISO system.

Given above-mentioned underlying physicallayer differences, TCP level quality of service is

compared on Figure 11. The cumulative distributionfunction of throughput during file download is presentedfor files above 64KB. We exclude “small” files from theplot, since measured throughput for "small" files is gatedby the TCP rather than PHY and MAC layers. Inparticular, connection round-trip time and slow start phaseof congestion avoidance algorithm dictate the rate atwhich packets arrive at the wireless link. "Small" files arenever able to fully utilize available bandwidth because thetransmit window doesn't advance into congestionavoidance phase. Also, large round trip times cause largedelays between consecutive congestion windowincrements, due to window advancement clocked by ACKarrivals. This effect is also present during a “big” filedownload, but only for a fraction of actual downloadtime. Thus, the achievable throughput during filedownloads asymptotically approaches effectivethroughput because of TCP.

0 500 1000 1500 2000 2500 3000 35000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

MIMO, 64KB < FileSize < 10MB, UtilDL 25% MIMO, 64KB < FileSize < 10MB, UtilDL 75% SISO, 64KB < FileSize < 10MB, UtilDL 75%

Figure 11. MIMO system offeres gain in coverage,thoughput and increases load the system can support

The number of users that build link utilization to75% in SISO, only cause 25% link utilization in MIMOdeployments. This corresponds to a three-fold differencein average data rate in these systems. In other words, tobring utilization back to 75%, the MIMO system cansupport up to three times more active users per cell site,while at the same time offering better quality of service.

Operating ADU error rate plays a pivotal rolein user perceived quality of service and area coverage.Table 3 lists the peak and effective data rates averagedacross the cell and coverage gain for several target errorrates under study. The results presented are for a macro-cell environment defined by cell radius of 6 miles, rooftopNAUs and a BTS antenna height of 30m. Two interestingtrends can be derived from the data in Table 4. Firstly, theaverage gross rates improve as the ADU error rateincreases, due to the fact that minimum SINR requirementfor a given mode (modem set point) is relaxed and allows

users with poorer link quality to take advantage of highermodes. The rationale is to let ARQ recover from deepfades by retransmission, but allow faster communicationlink during “good” channel states. On the other hand,non-linear slope of SINR curves does not allow forsignificant rate gain as target error rate moves up to 10%.

Target PER 1% 5% 10%Avg. Peak Data Rate [Mbps] 5.7 6.3 6.3Peak Data Rate (1-Error Rate) Ref +6% +0.5%DLEffective Data Rate [Mbps] 3.7 4.0 4.0Avg. Peak Data Rate [Mbps] 2.8 3.1 3.1Peak Data Rate (1-Error Rate) Ref +6% +0.7%UL

Effective Data Rate [Mbps] 1.9 2.1 2.1

Coverage 86% 90% 92%

Table 4: Operating error rate vs throughput andcoverage.

Furthermore, net throughput, defined asachievable post- channel data rate does not monotonicallyincrease. Compared to the reference error rate of 1%,there is a 6% gain in throughput for 5% PER and only0.5% gain with an error rate of 10%. Higher packet errorrate translates into improved coverage since more userswith poorer channel conditions are accepted into system.

20 30 40 50 60 70 80 90 1000.4

0.6

0.8

1

1.2

1.4

1.6

1.8

Link Utilization [%]

Ave

rage

Req

uest

Tra

nsfe

r D

elay

[se

c]

Impact Of Target Error Rate On Request Service Delays

Average Error Rate 1% Average Error Rate 5% Average Error Rate 10%

Figure 12. Link layer design permits operating error rateas high as 5%

Request service delays closely relate to wirelesslink quality and delay bounds determine system operatingerror rate. The average request service delay as a functionof link utilization is plotted in Figure 12 for the threeaverage error rates of interest (listed in Table 3). Asexpected, 10% ADU error rate and extended fades do notallow ARQ to recover quickly. Other two cases have veryclose performance, with PER 1% being marginally betteracross the whole utilization range. From the figure wealso read that the system can be safely loaded up to 60%of available bandwidth without any cost paid in servicequality. Also, request service delays do not exhibit a sharpknee at any utilization point showing that the system isstable for the whole loading range.

Cumulative Distribution of TCP Throughput

Average TCP Throughput During Request Download[Kbps]

From the above we conclude that TCPthroughput and latency requirements for mainstream webapplications can easily be obtained with up to 5% ADUloss over the air, due to ARQ/F mechanism. We put theemphasis on ADU error rate since without ARQ/Fmechanism, unfragmented higher layer packets wouldexperience unacceptable loss rates, forcing TCP todecrease its window often and retransmit excessively.Furthermore, ARQ/F and high allowable error ratespermit high spectral efficiency with acceptable delays.

Next, we investigate the effect of Doppler spreadon request service delays. Doppler spread governs howrapidly channel changes over time and for how long thewireless link stays in a fade. The system with averageoperating 5% error rate and fairly high link utilization of65% is examined over two channels with Doppler spreadsmost typical for fixed wireless access system- low 0.1Hzand high 1Hz. Cumulative distribution function of requestservice delays is plotted in Figure 13 for the two cases.Request service delays increase in the upper 30% ofcases. This is explained by prolonged fade duration forlower Doppler.

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Pro

b(X

<A

bsci

ssa)

Request Service Delay [seconds]

Impact Of Doppler Spread On Request Service Delay

Doppler 0.1Hz Doppler 0.025Hz

Figure 13. Low Doppler spread and increased fadeduration effect around 30% of request downloads

TCP interacts with MAC scheduling andARQ in several interesting ways. Here we examine theTCP reaction on very slow Doppler with fades up toseveral seconds. Figure 14 presents sequence numberevolution as a function of time and wireless channel. Weobserve that TCP flow control mechanism restricts thesender during both downlink and uplink fades, sinceACKs do not get through either because packet or ACKitself was lost. This effect further shapes applied trafficmodel. TCP avoids data transmission during fades andmore data is attempted when channel improves, as shownon Figure 6. TCP window evolution is slower duringburst errors. In a channel that otherwise results in 5%ADU error rate, because of TCP, the observed error rateis below 5%. Random errors confuse TCP, extended

errors are correctly identified as network congested orpoor channel condition. Furthermore, TCP windowevolution slow down during long channel fade preventsMAC buffer overflows. It is seen that link layer ARQrecovers from channel errors efficiently and preventsexcessive TCP timeouts, that is, link layer ARQ permits aTCP-friendly design!

Packet E

rror Rate

Downlink channelUplink channel

Figure 14. TCP sequence number evolution in fading –Doppler 0.025Hz channel

6. Advanced FeaturesOther concepts that enhance spectral efficiency of 2G-BWA systems that were not part of the simulation effortpresented here, include interference mitigation either atthe PHY or MAC, aggressive frequency reuse to easedeployment strategies and channel dependent schedulingto enhance system capacity.

Interference cancellation can be employed onboth links to reduce modem setpoints. Array processingtechniques at PHY or intelligent scheduling thatcoordinates transmissions across interfering BTSs lead toa system with aggressive frequency reuse. Additionalcapacity gain can be achieved with multi-user channeldependent scheduling that allocates bandwidth to the userwith the best channel quality while supporting desiredQoS.

7. ConclusionSeveral advanced technologies that allow broadbandwireless service deployment have been described andtheir benefits have been evaluated. Results show thatwireless link shaping has a significant impact on QoSprovisioning, and therefore careful design of PHY andMAC layers is critical for acceptable networkperformance. Adaptive modulation (AM) allowsaggressive system design that takes the most advantage ofeach link condition. While coverage remains guaranteed,AM greatly improves user perceived service quality,measured through average request delay, as well asnumber of users the system can support. Also, underlyingtechnology that enables finer granularity of AM is multi-

antenna signal processing at both ends. Our resultsindicate that MIMO and SD offer additional gain incoverage, almost 2.5 times improvement over SISOsystem in attainable cell size and data rates, which furtherbenefits QoS. Additionally, the design of link layer,ARQ/F mechanism and MAC scheduling in particular,allow network layer to perform well over wireless channelwith packet loss rates as high as 5%. ARQ/F mechanismhides wireless link losses from TCP windowingmechanism, preventing TCP control window restriction.Slower Doppler channels, due to prolonged channel fades,make harder for ARQ/F to quickly recover IP packets,giving rise to TCP timeouts and degrading tail caseperformance. But, due to acknowledgement based controlmechanism, TCP shapes offered traffic load according towireless channel state, allowing link layer to preventexcessive TCP timeouts and more successfully recoverfrom channel impairments.

ACKNOWLEDGEMENTS

The authors would like to thank Madame SeverineCatreux and Monsieur Luc Haumonte, of Iospan WirelessInc, for providing us with specific physical layersimulation results.

REFERENCES

[1] Erceg V., et al “A model for the multipath delayprofile of fixed wireless channels “ , Selected Areas inCommunications, IEEE Journal of, Vol. 17 3 , March1999 , Page(s): 399 -410

[2] Kumar A., “Comparative analysis of versions of TCPin a local network with lossy link” , IEEE/ACMTransactions on Networking, August 1998

[3] Stevens W.R., TCP/IP illustrated, Volume 1, AddisonWesley, 1994

[4] Balakrishnan H., Padmansbhan V., Seshan S., KatzR.H., “ A comparison of mechanisms for improvingTCP performance over wireless links” , IEEE/ACM

Transactions on Networking, June 1997, Page(s):336-350

[5] Furuskar A., Frodigh M., Olofsson H., Skold J.,“System performance of EDGE, a proposal forenhanced data rates in existing digital cellularsystems” , Vehicular Technology. Conf., 1998,vol: 2,Page(s): 1284 -1289

[6] Baum D.S., et al., “Measurement andcharacterization of broadband MIMO fixed wirelesschannels at 2.5 GHz” , Personal WirelessCommunications, 2000 IEEE InternationalConference on, 2000, Page(s): 203 -206

[7] Paulraj A.J., Kailath T., “ Increasing system capacityin wireless broadcast systems using distributedtransmission /directional reception (DTDR)” , U.S.Patent No. 5,345,599, http://www.uspto.gov/

[8] Tarokh V., Seshadri N., Ralderbank A.R., “Space-time codes for high data rate wirelesscommunication: performance criterion and codeconstruction” , Information Theory, Transactions on,Vol. 44, No 2, March 1998, Page(s) 744 – 764

[9] Foschini G.J., Gans M.J., “On limits of wirelesscommunication in a fading environment when usingmultiple antennas” , Wireless PersonalCommunications, 1998, vol 6, Page(s) 311-335

[10] Keller T., Hanzo L., ” Adaptive modulationtechniques for duplex OFDM transmission” ,Vehicular Technology, IEEE Transactions on, Vol.49 5, Sept. 2000, Page(s): 1893 -1906

[11] NS - http://www-nrg.ee.lbl.gov/ns/

[12] Hary K.V.S., Sheikh K., “ Interim channel models forG2 MMDS fixed wireless applications” , IEEE 802.16Broadband Wireless Access Working Group<http://ieee802.org/16>

[13] Barford P., Crovella M.E., "GeneratingRepresentative Web Workloads for Network andServer Performance Evaluation," in Proceedings ofPerformance ’98/ACM SIGMETRICS '98, pp. 151-160, Madison WI.%.