Performance and Capacity of Cellular OFDMA Systems ... - IEEE Xplore

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 6, NOVEMBER 2008

3641

Performance and Capacity of Cellular OFDMA Systems With Voice-Over-IP Traffic Qi Bi, Senior Member, IEEE, Stan Vitebsky, Member, IEEE, Yang Yang, Yifei Yuan, Member, IEEE, and Qinqing Zhang, Senior Member, IEEE

Abstract—We propose a methodology to evaluate the Voiceover-IP (VoIP) capacity and performance of orthogonal frequency-division multiple-access (OFDMA)-based systems. The approach combines the queuing and interference analyses to find the delay and power outage. Using this analytical framework, we provide insight into tradeoffs among capacity-limiting factors, including channel dimensions, power, interference threshold and latency, and we identify techniques for capacity optimization. The numerical demonstrations are carried out on the newly standardized Ultra Mobile Broadband (UMB) system. The same methodology can also be applied to Universal Mobile Telecommunications System (UMTS) Long Term Evolution (LTE) and wireless mobile Worldwide Interoperability for Microwave Access (WiMAX) systems. Index Terms—Capacity, evolution-data optimized (EV-DO), orthogonal frequencydivision multiple access (OFDMA), ultra mobile broadband (UMB), Voice-over-IP (VoIP).

I. I NTRODUCTION

I

N THIS paper, we study the performance of next-generation wireless mobile networks based on orthogonal frequencydivision multiple access (OFDMA) with Voice-over-IP (VoIP) traffic. To date, there have been several studies on wireless VoIP performance. In [1], the VoIP OFDMA system capacity is estimated from the link spectral efficiency, taking into account time and frequency dimensions provided by the frame structure. Because the VoIP capacity is often determined by the voice packet delay metric [2], classical queuing theory can be applied to derive the delay performance [3], particularly for scheduled systems such as OFDMA [4], [5] and EV-DO Revision A forward link (FL) [2]. Packet latencies in scheduled systems depend not only on the physical layer data rate but also on the signaling delay due to the contention for limited shared resources [6]. To fully characterize the VoIP capacity and performance, we investigate the role of various system-level parameters. In particular, we perform the queuing analysis on Manuscript received April 24, 2007; revised September 18, 2007, January 17, 2008, and January 22, 2008. First published February 15, 2008; current version published November 12, 2008. The review of this paper was coordinated by Prof. R. Jantti. Q. Bi, S. Vitebsky, Y. Yang, and Y. Yuan are with Alcatel-Lucent, Whippany, NJ 07981 USA (e-mail: [email protected]; [email protected]; [email protected]; [email protected]). Q. Zhang was with Alcatel-Lucent, Whippany, NJ 07981 USA. She is now with the Applied Physics Laboratory, Johns Hopkins University, Baltimore, MD 21218 USA, and also with the Department of Electrical and System Engineering, University of Pennsylvania, Philadelphia, PA 19104 USA (e-mail: [email protected]). Digital Object Identifier 10.1109/TVT.2008.918705

both talk spurt and voice packet levels to obtain the latency distribution statistics of system users. The talk spurt latency analysis includes both assignment and signaling latency. The delay-sensitive and low-data-rate nature of VoIP application calls for an allocation of a constant-rate pipe, whose link quality is maintained either by adaptive resource allocation in the time–frequency domain or by power control to compensate the fading and interference variations. The latter is a more attractive option for scheduled systems, in which the amount of assignment signaling overhead must be contained [7], [8]. In [9] and [10], it has been proposed to use such a configuration to support VoIP in the next-generation cellular OFDMA system based on Ultra Mobile Broadband (UMB) [11], [12]. Simplified capacity calculations for this system are performed in [12] and [13]. In this paper, we assume the same type of operation and resource management for VoIP traffic. Our analysis framework, however, accounts for the distributions of interference, path loss and power outage, and interdependency and feedback that exist between these parameters and the VoIP packet latency. The link quality of a VoIP user is governed by the received signal-to-noise ratio (SNR), which is a function of various power domain quantities such as transmit power, channel fading, and interference. The difficulty of analysis lies in the fact that the interference is highly coupled with most other quantities, particularly in cellular systems operating with universal frequency reuse in all cells. The other-cell interference analysis in a power-controlled, code-division multiple-access (CDMA) environment with a large number of low-rate users via a relative other-cell-to-same-cell interference factor has been introduced in [14]. Recently, several analytical and semianalytical formulations for evaluating the distributions of other-cell interference on the reverse link (RL, uplink) of the power-controlled CDMA have been proposed [15], [16]. These approaches use lognormal approximations of interference distribution functions. The complete modeling of interference in OFDMA, however, has to include the frequency-domain aspect–power spectral density (PSD). Considering that the OFDMA tones or groups of tones are allocated to one user per sector at a time, the above approximations for the interference statistics may no longer hold. In addition, frequency hopping, which is often used in OFDMA to randomize and smooth the narrowband interference, needs to be taken into account. In [17], the impact of frequencyhopping patterns on the link-level performance is compared between trellis-coded modulation (TCM) and bit-interleaved coded modulation (BICM). In this paper, we primarily focus on the system-level aspects. Hence, we introduce a model for the statistical characterization of other-cell interference in the

0018-9545/$25.00 © 2008 IEEE

3642

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 6, NOVEMBER 2008

OFDMA system with VoIP traffic and study its impact on power outages at the base station and the mobile. This paper is organized as follows. First, we describe the model we use in the analysis of the OFDMA cellular network with VoIP traffic. Second, we provide an overview of the proposed analytical framework for capacity and performance evaluation. Then, we introduce the individual components of this framework, including analytical models for talk spurt resource assignment latency, voice packet latency, interference, transmit power outage, and user latency outage. Finally, we provide the numerical results based on the proposed model. II. S YSTEM M ODEL D ESCRIPTION A. Cellular OFDMA Network Assumptions Consider a cellular system consisting of sectorized cells covering a part of the geographical area and using OFDMA on FL (downlink) and RL (uplink). Multiple mobiles are located within each cell’s coverage area and are connected to one (serving) sector at any instant of time. We assume that the system implements fast serving sector switching in such a way that the serving sector is always the sector with the strongest received signal (or smallest path loss). Without loss of generality, we also assume that the FL and RL serving sectors are always the same. B. OFDMA Transmission Format The transmitted signal consists of narrowband tones, which are nearly orthogonal to each other in the frequency domain. A group of tones transmitted over the duration of one time slot (or frame) constitutes the smallest scheduling resource unit, which is also known as a “tile,” as shown in Fig. 3. In general, different tones belonging to a tile may be scattered across the entire band, in which case, the tile transmission would experience a diversified channel and interference on each subcarrier. Alternatively, a tile can always be made of a contiguous set of tones so that the channel and interference experienced by the tile are more localized. In this paper, we assume that the system operates in the latter mode in which all tones belonging to the tile on a given time slot can only experience the interference from just one tile in every neighbor sector. In this mode, channel estimation and power control are facilitated by the dedicated pilot embedded within the tile. In addition, slot-to-slot tile hopping ensures that different users in other sectors interfere with each other over time, providing interference and channel diversity. As in many third-generation systems, the Hybrid ARQ (HARQ) is employed to increase system capacity. To this end, encoder packet transmission occurs using multiple HARQ interlaces repeating every certain number of frames and having a fixed-maximum-allowed number of subpacket retransmissions. In the system, such as UMB, which employs synchronous HARQ, the assignment by the scheduler of the tile–interlace resources is, in general, valid for the duration of each encoder packet transmission. However, to reduce the signaling overhead related to the resource assignment for low-rate applications such as voice, the scheduler can assign resources for longer

durations corresponding to the talk spurt activity of the speech. This kind of “persistent” assignment is described in more detail below.

C. Resource Allocation for VoIP Traffic Statistical characterization of telephone conversations by P. T. Brady conducted in the 1960s at Bell Laboratories [18], [19] resulted in the main finding that a telephone conversation can be represented by the ON–OFF pattern, with an ON period corresponding to a talk spurt and an OFF period corresponding to a period of silence. Furthermore, it was found that the duration of ON and OFF periods can approximately be modeled by an exponential distribution with a mean value of 1.2 and 1.8 s, respectively. In this paper, we follow the voice source model described in [18] and [19]. We assume that within a talk spurt, all voice packets are “full rate” and are deterministically generated every 20 ms when using the enhanced variable rate coder (EVRC) [20]. During the silence period, the vocoder frames are suppressed without transmission. Details of the EVRC frame statistics and IP overhead for the VoIP traffic source model can be found in [2]. We assume that the VoIP operation in an OFDMA system entails allocating a certain portion of tiles within each interlace to the signaling channels with the remaining tiles eligible for traffic transmission. We assume that the transmission channel composed of one tile–interlace resource is assigned to a user when a new talk spurt from that user arrives at the system, as shown in Fig. 3. Any of the available tiles on one of the HARQ interlaces can be assigned to a new talk spurt, and the assignment is kept for the entire duration of the talk spurt. When the channel is assigned, an assignment message is sent to the user on a signaling channel. The channel is deassigned when the talk spurt ends, and the user enters the silence period. In the context of the proposed UMB system, a packet format with payload size of 32 B is optimized to carry one 22-B fullrate EVRC frame with 10 B available for compressed real-time transport protocol/user datagram protocol/internet protocol and radio link protocol/media access control overheads using one tile allocation [9]–[13]. In this paper, we assume that during the talk spurt, all VoIP packets are transmitted using this format. Assuming there are eight HARQ interlaces with slot duration of 0.911 ms, the HARQ retransmission interval is 7.3 ms, and an average of approximately 2.74 retransmissions can be made within the 20-ms frame interarrival period without overrunning the packet queue. In the 5-MHz UMB system (fast Fourier transform size of 512) considered here as an example, there are 32 16-tone tiles available for each time slot. With two tiles used for guard bandwidth, 30 tile–interlace resources are available in each slot (a total of 240 resources for each interlace period). Some of these resources are used for signaling overhead. Hence, the maximum number of voice calls that can be served at any given time is limited. We assume that there are 25 and 20 tile–interlace resources available in each time slot (i.e., 200 and 160 total) on FL and RL, respectively. In addition to allocating bandwidth resources such as tile–interlace pairs, the access network also allocates an

BI et al.: PERFORMANCE AND CAPACITY OF CELLULAR OFDMA SYSTEMS WITH VOICE-OVER-IP TRAFFIC

3643

appropriate amount of FL transmit power to achieve a desired number of HARQ transmissions per packet with certain target probability (i.e., target HARQ termination) for the mobiles in different cell locations. The channel quality indicator (CQI) and the HARQ ACK/NAK feedback are used to adjust this power allocation to compensate for mobile movement and changing fading conditions. On the RL, the network also sends power control commands and HARQ ACK/NAK feedback to instruct the mobile to achieve the SNR required for the target HARQ termination at the base station receiver. III. O VERVIEW OF P ROPOSED A NALYTICAL F RAMEWORK To evaluate the VoIP capacity in an OFDMA system, multiple resources need to be considered. In particular, we consider the following: 1) Orthogonal dimension: the number of tile–interlace pairs available to support VoIP traffic and associated signaling; 2) Transmission power: the nominal transmission power of the base station and the mobile; 3) Other-cell interference: excessive other-cell interference can drive the system to operate in an unstable mode in terms of power needed to sustain target performance and may significantly degrade the performance for the entire sector. The above resources are tightly correlated. The use of one resource can significantly impact the availability of other resources. For example, the interference experienced by a receiver heavily depends on the same channel utilization in other sectors/cells, which is governed by the dimension resource usage. The interference, in turn, dictates the transmit power required to achieve the target performance. The transmit power then produces the interference to the receivers in other sectors/ cells. When there is a shortage of transmit power, the packet HARQ and the queuing latency are increased, leading to higher channel utilization. To evaluate the VoIP capacity for different wireless data technologies, a consistent performance criterion is needed. Since the latency is a critical measure of the VoIP quality of service, the performance criterion for VoIP capacity evaluation based on user latency has been introduced in [2] and is now widely used by the industry. This criterion calls for the evaluation of the tail of the per-user packet latency distribution. In particular, the user is considered to experience latency outage when the 98 percentile of its packet latency exceeds an acceptable delay bound (e.g., 100 ms). Accordingly, the system is considered to reach its full capacity under maximum Erlang loading, for which no more than 2% of all users experience latency outage. The same metric is adopted in this paper as well. To investigate the interworking of multiple resources in the OFDMA system, a closed-loop methodology is established, as illustrated in Fig. 1. There are three major interdependent components in the proposed analytical framework. 1) Queuing and latency analysis to evaluate the resource assignment latency for VoIP traffic under loading and packet queuing latency based on the HARQ termination

Fig. 1. Components of the proposed analytical framework and their interdependencies.

statistics. This analysis provides the user-level latency outage and determines the channel utilization based on traffic loading. 2) Packet transmission performance analysis to evaluate the achievable packet transmission performance as a function of received SNR. This analysis provides HARQ termination statistics due to the inability of some packets to meet the target performance. 3) Power and interference analysis to evaluate the othercell interference experienced by the receiver and the transmit power requirement in the transmitter. This analysis also provides the received SNR degradation due to the power overload control scheme implemented in the transmitter. Fig. 1 illustrates the interdependency and feedback mechanism among the resources that are accounted for in the proposed analytical framework. The user latency outage with respect to the acceptable VoIP performance bound and the transmit power outage (i.e., the probability of exceeding maximum allowed transmit power) are computed as a result of the convergence of this iterative procedure and are used to determine the VoIP capacity of the system and transmit power required to support this capacity. The numerical calculations in Section VI are verified to converge with the values of the assumed parameters. The general proof of convergence is beyond the scope of this paper. Since the VoIP traffic is symmetrical with respect to the FL and RL loading it produces, the overall system capacity is determined by the limiting link. The proposed framework is able to capture this effect and provides a useful means to obtain the system configuration (e.g., required minimum base station transmit power) that balances the capacity of two links. IV. VO IP L ATENCY C HARACTERIZATION The latency in the OFDMA system depends on the availability of shared channel resources and the utilization of the channel resource by the VoIP packet transmissions. Thus, we employ queuing analysis to characterize the relationship between latency and capacity.

3644

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 6, NOVEMBER 2008

Fig. 2. Queuing model used for talk spurt resource assignment latency analysis.

A. Analytical Model for Talk Spurt Resource Assignment Latency

n+1

(4)

To transmit a new talk spurt on either FL or RL, the scheduler assignment is signaled to the user over the signaling channel, followed by the voice packets over the assigned traffic channel. If there is no traffic resource available for the new talk spurt, the talk spurt is queued until an existing talk spurt ends, and the traffic channel resource it has occupied is released. The queuing model for talk spurt delay analysis is depicted in Fig. 2. Several notations are used in this figure. 1) λ: The average talk spurt arrival rate. The arrival is modeled as a Poisson process. The arrival rate λ depends on the average talk spurt duration Tb and the voice activity a as λ = aTb−1 nv , where nv is the offered traffic load in Erlangs. 2) μ1 : The average service rate of the signaling channel. The signaling message transmission time is approximated by an exponential distribution with μ1 = Ts−1 to facilitate the analysis, where Ts is the time slot duration. 3) m1 : This is the number of tile–interlace resources dedicated to signaling transmission within one interlace period. 4) μ2 : The average service rate of the traffic channel. In this case, the traffic service time represents the talk spurt duration, which is modeled as an exponential distribution with μ2 = Tb−1 . 5) m2 : This is the number of tile–interlace resources available for traffic transmission within one interlace period. We model the system as a tandem of M/M/m queues, which is equivalent to the case of each queue independently operated with Poisson arrivals. The probabilities that n talk spurts are present in the jth queue are given by [21] min(m ,n)

j ρnj mj , Pj (n) = Pj (0) min(n, mj )!

for j = 1 or 2

(1)

where ρj =

λ mj μj

(2)

⎤−1 ∞ k  (mj ρj )k  (m ρ ) j j ⎦ . + Pj (0) = ⎣1 + k−mj k! m !m j j k=1 k=mj ⎡

The multiserver service time sj seen by a talk spurt waiting for resource assignment follows an exponential distribution as fsj (t) = mj μj e−mj μj t . On the other hand, the service time sj experienced by a talk spurt that has already been as(t) = signed resources follows an exponential distribution as f sj μj e−μj t . If we denote the waiting time in the signaling server and the traffic server as wb1 and wb2 , respectively, their distributions are given by ⎞ ⎛ ∞  ⎟ ⎜ fwbj (t) = Pj (n + mj ) ⎝fsj (t) ⊗ fsj (t) ⊗ · · · ⊗ fsj (t)⎠   n=0

mj −1

(3)

where ⊗ denotes convolution. The total queuing delay wb experienced by a new talk spurt is wb = wb1 (t) + s1 + wb2 .

(5)

Hence, the distribution of the total queuing delay due to resource assignment can be computed from fwb (t) = fwb1 (t) ⊗ f (t) ⊗ fwb2 (t). s1

(6)

B. Analytical Model for Voice Packet Latency Within the Talk Spurt In Fig. 3, we illustrate the timeline of resource allocation and HARQ applied to a talk spurt as well as some key notations used in the packet latency analysis of this section. Following the arrival of the talk spurt, the system assigns to it a tile resource on one of the Ns interlaces (interlace number 4 in the example shown in the figure). This assignment is “persistent” throughput the duration of the talk spurt, i.e., there always exists a tile allocated on the same (fourth) interlace that may be, but does not have to be, used for the transmission of packets that are part of this talk spurt. The packets may take a variable number of HARQ transmissions. If due to early HARQ termination the transmission of a packet ends before the HARQ cycle, in which the next packet has not yet arrived, the tile allocated to the talk spurt in this cycle goes unused. In addition, shown in the figure, the physical location of the tile in frequency varies from one HARQ cycle to another due to the tile frequency-hopping feature. After the tile–interlace resource is assigned to a new talk spurt, the individual full-rate packets within the spurt are generated by the vocoder in regular intervals T0 . Given a known head-of-spurt waiting time distribution, the subsequent waiting time of the packets can be derived and evaluated using numerical methods. The waiting time of the first packet in the talk spurt is given by w1 = wb + τ , where τ is the delay due to alignment of the packet transmission with a specific interlace determined by the talk spurt assignment. Assuming that a random interlace is assigned to every talk spurt, the alignment delay can be modeled as a random variable with uniform distribution fτ (t) within the

BI et al.: PERFORMANCE AND CAPACITY OF CELLULAR OFDMA SYSTEMS WITH VOICE-OVER-IP TRAFFIC

Fig. 3.

3645

Talk spurt resource allocation and HARQ timeline and notation used in packet latency analyses.

interval [0, Ns Ts ), where Ns is the number of interlaces or the period of HARQ retransmission in slots. The distribution of the number of packets in a spurt K can be approximated using the exponential talk spurt duration assumption as

From (8), the distribution of the waiting time for the ith packet within the talk spurt can recursively be obtained as ⎧ t ≥ Ns T s ⎪ ⎨ fdi−1 (t + T0 ), C(t) fwi (t) =  ⎪ fdi−1 (t + T0 − jNs Ts ), 0 ≤ t < Ns Ts ⎩ j=0

pK (k) = e−(k−1)μ2 T0 − e−kμ2 T0 ,

for k ≥ 1.

for i > 1

If a new packet arrives before the previous packet completes the transmission, the new packet must wait in queue. Otherwise, the new packet must align the beginning of its transmission with the earliest timing of the interlace assigned to the talk spurt. The assigned interlace periodically occurs, i.e., every Ns Ts starting from the initial alignment. Hence, the waiting time for the ith packet of the spurt wi depends on the waiting and the service time of the previous packet wi−1 and vi−1 , respectively, as in (8), shown at the bottom of the page. The packet transmission time vi−1 seen by another packet waiting in queue follows the HARQ early termination probability {hn } for packet termination at the nth transmission as

fv (t) =

N max 

hn δ(t − vn ),

vn = nNs Ts

(9)

n=1

where fdi−1 (t) is the distribution of the total packet delay of the i − 1th packet seen by the waiting ith packet fdi−1 (t) = fwi−1 (t) ⊗ fv (t); C(t) is the number of HARQ cycles between the arrival of the i − 1th packet and the start of transmission of the ith packet C(t) = (t + T0 )/Ns Ts , with x denoting the floor of x; and the initial waiting time distribution fw1 (t) = fwb (t) ⊗ fτ (t). The expression for the distribution fwi (t) in the range of waiting times smaller than one HARQ cycle Ns Ts in (10) is based on the observation that the alignment time governs the waiting time only if the previous packet finishes transmission in one of the HARQ cycles preceding the cycle in which the next packet arrives. The probability of this event is the summation of probabilities of the previous packet finishing in one of the j ∈ [0, . . . , C(t)] HARQ cycles between its arrival and the start of transmission of the next packet. The overall distribution of the packet waiting time is obtained by combining the distributions of the waiting time of the ith packet of (10) weighted by the probability that the talk spurt is large enough to include at least i packets or more, i.e.,

where Nmax is the maximum number of HARQ transmissions, and δ(t) is the continuous Dirac delta function.

 wi =

(10)

(7)

fw (t) =

∞ ∞  1  fwi (t) pK (k) K i=1 k=i

wi−1 + vi−1 − T0 ,

wi−1 + vi−1 ≥ T0

Ns Ts − mod(T0 − wi−1 − vi−1 , Ns Ts ),

otherwise

,

for i > 1

(11)

(8)

3646

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 6, NOVEMBER 2008

 where K = ∞ j=1 jpK (j) is the average number of packets in the talk spurt. Finally, the distribution of the total packet delay d = w + v is given by fd(t) = fw (t) ⊗ f (t) v

(12)

where v represents the transmission time experienced by the transmitted packet itself and only includes the time slots between the start of the first and the end of the last HARQ transmission. Hence, its distribution is slightly different from (9), i.e., f (t) = v

N max 

hn δ(t − vn ),

vn = (1 + (n − 1)Ns ) Ts .

n=1

(13) The tile–interlace resource utilization during talk spurt transmission is given by ρp =

Nmax Ns T s  nhn . T0 n=1

(14)

This leads to the aggregate average channel utilization of ρ2 ρp with talk spurt level traffic queue utilization ρ2 defined in (2).

A. Analytical Model for Interference and System Stability Criterion Under the assumption of the tile-hopping mode of operation described in Section II, we perform interference analysis on the tile level. In other words, all random variables introduced in the analysis take a single average value per tile and stochastically change from tile to tile (in frequency and time). In principle, similar equations could be written for the tone-hopping mode but with per subcarrier resolution. In the latter case, tile-level parameters could be obtained by aggregating over the noncontiguous subcarriers constituting the tile. The traffic tile activity factor αk is a binary random variable found from the talk spurt level and packet level queuing analysis as (15)

The transmit power in a tile is determined by the SNR requirements for traffic as γ=  ∞

sm ξm,m αn sn ξn,m + Nt W

=

sm ξm,m m Nt W IoT

m IoT =

∞ 1  αn sn ξn,m + 1. Nt W n=1

(17)

n=m

Note that the traffic channel SNR requirement γ is, in general, also a random variable. However, to simplify the analysis, we assume an ideal power control operation and use a constant γ, taking a single mean value of “effective” SNR for the composite channel. If necessary, the formulation can be extended to include the distribution of γ based on the assumed channel mix and nonideal power control. In addition, in (16), we assume that the variation of the background noise power from tile to tile is negligibly small compared with that of the signal power. From (16), we write the expression for the per-tile transmit power at the mobile (or base station) in the mth sector on the uplink (or downlink) as m −1 γξm,m . sm = Nt W IoT

(18)

This transmit power, in turn, contributes interference to mobiles (base stations) in other sectors receiving the same tile. n experienced by a mobile in Combining (17) and (18), the IoT (or the base station of) sector n is

V. I NTERFERENCE AND P OWER O UTAGE C HARACTERIZATION

Pr(αk = 1) = 1 − Pr(αk = 0) = ρ2 ρp = ρ.

serving base station (or from the base station of the nth sector to its own served mobile), ξn,m is the path loss (including antenna gain, shadow fading and other losses) between the receiver in sector m and the transmitter in sector n, Nt is the noise PSD including the noise figure of the receiver, W is the tile m is the interference over thermal noise rise bandwidth, and IoT experienced by the mobile in sector m defined as

(16)

n=1 n=m

where γ is the per receive antenna average SNR requirement for traffic, sm is the traffic transmit power from (or to) the mobile in the mth sector to (or from) its serving base station on the uplink (or downlink), sn (n = m) is the traffic transmit power in the same tile from a mobile served by the nth sector to its own

Nt W

n (IoT

− 1) =

∞ 

αm sm ξm,n

m=1 m=n

= Nt W γ

∞ 

m αm IoT

m=1 m=n

ξm,n . ξm,m

(19)

The ratio of path losses between nonserving and serving sectors is a random variable contributing to the distribution of IoT and defined as Dm,n ≡ ξm,n /ξm,m . Hence, (19) can be rewritten as n IoT =γ

M 

m αm IoT Dm,n + 1

(20)

m=1 m=n

where the summation is over a large enough number of sectors M that make the most significant contribution to interference. Assuming the mobiles are uniformly distributed among sectors of the network, the interference level experienced by different sectors follows the same distribution. The mean interference level can be obtained from (20) as I¯oT = (1 − ργΔ))−1

(21)

 where Δ = M m=1 Dm,n and Dm,n are the mean path loss m=n ratios between nonserving and serving sectors. From (21), we immediately see that to maintain operational stability, the

BI et al.: PERFORMANCE AND CAPACITY OF CELLULAR OFDMA SYSTEMS WITH VOICE-OVER-IP TRAFFIC

average tile utilization ρ must be less than the value of (γΔ)−1 . This requirement is similar to the loading stability requirement for the RL of the CDMA system based on the pole capacity [22], [23] and can be used as a guideline for engineering system capacity calculations of power-controlled OFDMA with VoIP or other homogeneous traffic. For numerical calculations, the distribution of Dm,n can be precomputed assuming appropriate link budget, shadow fading, antenna patterns, and other system parameters. Note that an analytical expression in [16] can also be used to calculate Dm,n if a particular subset of assumptions on antennas, sectorization, and fading is of interest. Iterative calculations can be applied to (20) to find the distribution of IoT with the initial value of fIoT (I) = δ(I), similar to the approach reported in [15]. The convergence is established based on the smallness of the changes in the calculated IoT distribution from one step to another. Based on (18), the distribution of the tile transmit power sm is found via the convolution of the distribution of IoT and the distribution of path loss between the mobile and the serving base station ξm,m in the decibel domain with appropriate scaling. B. FL Power Outage Analysis The probability that the total BS amplifier power required to support VoIP sector capacity exceeds the maximum power SBS max must satisfy the power outage criterion Poutage = Pr(S > SBS max )

(22)

where S is the total transmit power on any of the interlaces with at least one active traffic tile, and Poutage is an acceptable outage probability (e.g., 1% or 2%). The base station transmit power S on an active interlace is a random variable that can be written as S = S0 +

J 

αj sj

(23)

j=1

where S0 is the interlace transmit power used for sector-wide overhead and signaling transmission, αj and sj are the activity coefficient and the transmit power on the traffic tile j of this interlace, respectively, and J is the maximum number of traffic tiles available for transmission on each interlace (excluding guard tones, overhead and signaling tiles, etc.). Assuming that the assignment algorithm balances the load between interlaces, we consider all the interlaces to be equivalent. Therefore, the maximum number of traffic tiles per interlace J is related to the number of traffic servers m2 as J = m2 /Ns . Substituting the interlace transmit power (23) into (22) and applying the theorem of total probability, we establish the following power outage condition: Poutage

  j J   1 = Pt (j) Pr sl > SBS max − S0 1 − Pt (0) j=1 l=1

(24)

3647

where Pt (j) is the probability of j traffic resource tiles being used on the interlace, which we obtain from the downlink talk spurt and packet level queuing models as J    m  P (m)ρjp (1 − ρp )m−j , Pt (j) = j 2 m=j

j ∈ [0, 1, . . . , J] (25)

where P2 (m) is the probability of m traffic resource tiles on the interlace being assigned to active talk spurts. Based on the assumption of balanced talk spurt assignment among all interlaces, we compute this probability as P2 (m) =

Ns 

P2 (Ns (m − 1) + l) ,

m ∈ [1, 2, . . . , J]

l=1

P2 (0) = P2 (0)

(26)

where P2 (n) is the probability of n traffic tile–interlace server resources being used by active talk spurts. P2 (n) depends on the talk spurt utilization ρ2 according to (1)–(3). Using (24), we verify that the power outage criterion is satisfied for a given delay outage obtained from the queuing model. If the power outage is not satisfied, the utilization of traffic tiles ρ2 ρp needs to be reduced by reducing the call arrival rate and the Erlang loading in the queuing model. Consequently, we have an iterative procedure for capacity analysis based on both delay and power outage. When the total demanded power exceeds the maximum available traffic transmit power, the base station amplifier overload control algorithm scales the power of each tile by a scale factor  gBS = min

 SBS max − S0 ,1 . S − S0

(27)

The scale factor has an immediate impact on the achieved SNR at the receiving mobile. The distribution of the achieved SNR is, therefore, directly computed from the probability of base station power outage as  1 Pt (j) 1−Pt (0) j=1 J

Pr (γ = gBS γt ) =  × Pr

j 

 −1 sl = gBS (SBS max −S0 )

,

for gBS ∈ (0, 1]

(28)

l=1

where γt is the target SNR value that achieves a desired HARQ delay profile. Probability (28) is used as the weighting for 98 percentile VoIP packet latency as a function of SNR in the evaluation of the user latency outage. C. RL Power Outage Analysis Similarly, the probability that the mobile station transmit power exceeds the maximum power limit SMS max defines the

3648

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 6, NOVEMBER 2008

TABLE I SYSTEM PARAMETERS USED FOR THE EVALUATION OF INTERFERENCE RISE TRANSMIT POWER DISTRIBUTIONS

RL outage probability Poutage on any interlace on which the mobile has an active traffic tile, i.e., Poutage = Pr(s0 + s > SMS max )

(29)

where s is the transmit power of the mobile on the traffic tile, and s0 is the overhead transmit power. The uplink power outage analysis is performed by substituting the tile transmit power distribution function found from (18) into (29). When the demanded power exceeds the maximum available traffic transmit power, the mobile clips the transmit power at maximum. In other words, the following scaling is applied:  gMS = min

   SMS max SMS max − s0 , 1 = min − g0 , 1 s s (30)

where g0 = s0 /s is the overhead-to-traffic transmit power ratio. The scale factor in (30) has an immediate impact on the distribution of the achieved SNR at the receiving base station, e.g.,   SMS max Pr(γ = gMS γt ) = Pr s = , gMS + g0

The analysis may also be performed for different cell radii. Additionally, different SNR targets could be tried to investigate if the capacity could be optimized by trading tile utilization and transmit power. VI. N UMERICAL R ESULTS In this section, we present the application of the proposed analytical framework to evaluating the VoIP capacity in the OFDMA system based on the UMB. As discussed above, the metric used for this purpose is based on the user latency outage, which heavily depends on the underlying distributions of packet latency, interference, and achieved SNR. However, before embarking on the numerical evaluation of these distributions, we illustrate how our formulation could be used to provide an engineering estimate of capacity based on the average values of various system parameters. By combining (2), (14), and (21), making substitutions, and rearranging the terms, we obtain the expression for the estimated RL Erlang capacity as

nv = for gMS ∈ (0, 1]. (31)

D. User Latency Outage Characterization Given the distribution of achieved SNR and the dependence of the 98 percentile latency from the packet latency analysis on the SNR, we find the distribution of 98 percentile latencies using (28) and (31) as weighting probabilities and compare the 98 percentile of the resulting distribution with the delay bound. The tile utilization and the distribution of active tiles are iteratively adjusted to maximize the capacity while satisfying the delay bound for the 98 percentile user.

λ = aμ2

  −1 m2 T0 1 − I oT aγNs Ts NH Δ

(32)

 max where NH = N n=1 nhn is the average number of HARQ transmissions corresponding to the received target average SNR γ. Substituting the values of parameters from the UMB link level simulations, the average other-cell interference analysis based on the assumptions listed in Table I and the UMB transmission format and VoIP traffic parameters, i.e., T0 = 20 ms, I oT = 5.5 dB, a = 0.4, γ = 3.5 dB per antenna (6.5 dB, combined), Ns = 8, Ts = 0.911 ms, Nmax = 6, NH = 2.35 (based on the target HARQ termination distribution {hn } listed in Table I), and Δ = 0.75, the Erlang capacity in (32) is evaluated to approximately 1.25 of the maximum number of tile–interlace resources m2 available for traffic. Assuming that the number of

BI et al.: PERFORMANCE AND CAPACITY OF CELLULAR OFDMA SYSTEMS WITH VOICE-OVER-IP TRAFFIC

3649

Fig. 4. Talk spurt resource assignment latency distribution Fwb (t) as a function of talk spurt level traffic resource utilization ρ2 .

these resources on the RL of the UMB system is 160 (20 tiles in each of the eight interlaces), with the remaining resources used by ACK channel and CDMA control segment, the estimated capacity is about 200 Erlangs. Above, we assumed a typical mean value of 5.5 dB for the nominal I oT level [2], [24]. The result of the above calculation represents an estimate of the capacity unrestricted by coverage and transmission power limitations. To find whether the capacity based on the mean interference approximation estimated via (32) is further reduced due to the user latency outage resulting from these limitations, we carry out numerical analysis according to the framework proposed in this paper. In Fig. 4, we plot the distribution of the RL talk spurt assignment latency wb of (6) as a function of the talk spurt resource utilization ρ2 obtained according to the model described in Section IV-A with the following parameter values: m1 = 16, m2 = 160, Tb = 1.2 s, and Ts = 0.911 ms. Note that the utilization shown in the figure does not include the packetlevel (ρp ) effects. We observe that the talk spurt assignment latency is negligible for 70% utilization. However, the tail of the assignment latency rapidly starts degrading when the utilization increases to around 85%. Nevertheless, from the interference stability criterion based on (21) and the parameter assumptions listed above, we find that the expected system operating regime with one tile–interlace allocation occurs when the utilization is less than (γΔ)−1 = 60%. Given the target HARQ termination distribution {hn } from Fig. 5, the packet-level resource utilization ρp is around 86% from (14). Therefore, the talk spurt level utilization required for stability is less than 70%, and the impact of resource assignment on the overall latency is expected to be small. The FL assignment latency is even less restricting because the number of available traffic resources m2 = 200 is larger than on the RL. In Fig. 5, we show the dependence of RL HARQ termination statistics on the average combined received SNR. Note that the FL results are approximately 0.5 dB worse due to the difference in packet format definitions. In calculating the HARQ performance, we used a mix of channel conditions recommended by the 3rd generation partnership project 2 (3GPP2) simulation methodology document [24], which is shown in Table I. Some of the discrete values from Fig. 5 are also tabulated. The target average number of transmissions NH and the termination

Fig. 5. HARQ termination probabilities {hn } and average number of transmissions NH as a function of average received SNR.

probabilities {hn } correspond to the target combined SNR of 6.5 dB. The termination probabilities are input for the voice packet-level latency analysis described in Section IV-B. They characterize the probabilities {hn } of each of the deterministic service times used in (9) and (10) of the queuing model formulation. The dependence of these probabilities on SNR allows us to perform latency analysis for users that are not capable of achieving the SNR target. The average number of HARQ transmissions shown in Fig. 5 characterizes the stable region of operation for the packet queue. If a mobile or a base station runs out of available power due to high interference rise and/or large path loss and cannot maintain the received SNRs above the level necessary to achieve less than the average of 2.74 transmissions, the packets they transmit would experience excessive queuing and are very likely to exceed an acceptable latency threshold. This observation suggests that to reduce the outage and maximize capacity, it is important to increase the probability that the users on the edge of coverage meet their SNR targets. In Fig. 6, we show the median and the 98 percentile of the voice packet latency distribution fd(t) of (12) obtained by applying the analysis in Section IV-B with the service times based on HARQ termination statistics corresponding to different levels of SNR degradation relative to the target (from Fig. 5). This result includes packets received in error as packets having an infinite delay. Note that we make a simplifying assumption that the same SNR degradation equally applies to all subpacket transmissions within a given talk spurt. From this figure, we observe that the packet latency exceeds the delay bound of 100 ms when the received SNR falls below the target by approximately 1 dB due to transmit power outage. If more than 2% of users experience a similar or worse combination of interference and path loss, the capacity limit based on user latency outage is reached. Hence, next, we investigate the interference and power outage distributions.

3650

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 6, NOVEMBER 2008

Fig. 6. Voice packet latency as a function of SNR degradation with respect to the target. Fig. 8. Mobile (MS) and base station (BS) traffic transmit power required to support VoIP capacity as a function of cell site-to-site distance and Erlang loading.

Fig. 7. Distribution of RL interference over thermal noise ratio as a function of Erlang loading.

In Fig. 7, we plot the distribution of interference over thermal noise ratio obtained via the method outlined in Section V-A and based on (20) without maximum power limitation at the transmitter. The distributions of ratios of the path losses between nonserving and serving sectors are calculated from the simulations by measuring path loss ratios in every location of the cellular system layout. This allows us to incorporate realistic antennas and fading characteristics (i.e., site-to-site correlation of shadow fading). In calculating the distributions of Dm,n , the assumptions that are consistent with 3GPP2 performance evaluation framework are used [24]. The main parameters used are listed in Table I. From Fig. 7, we observe that although the median IoT for 200 Erlangs is around 4 dB, the tail of the distribution at this loading is significant, which makes it challenging to support this capacity in power- or coverage-limited scenarios. The distributions of IoT for the FL are similarly obtained. The distributions of required transmit power for both links are obtained from (18) via convolutions of distributions of the interference rise IoT and the serving sector path loss ξm,m and using the transmit SNR requirement γ, which is approximately 1 dB higher than the received SNR shown in Table I due to imperfect power control. The tile utilization used in computing IoT is obtained from the iterative latency and queuing analyses

that include the effects of underachieving the received SNR required for HARQ target termination due to power limiting. The user latency outage is computed by applying the distributions of degraded SNR [see (28) and (31)] to the 98 percentile packet latency found from (12) with the corresponding HARQ termination statistics. The maximum base station and mobile transmit powers assumed in (28) and (31) are varied to achieve the same 2% user latency outage on FL and RL for different site-to-site distances and Erlang loading. In Fig. 8, the mobile and base station transmit powers required to support certain VoIP capacity are shown as a function of cell site-to-site distance. Under the balanced FL–RL condition, these transmit powers provide the same 2% user latency outage for the 100-ms latency bound on both links. Note that Fig. 8 does not include the power required to signal the overhead transmission, which could be a significant fraction of the total. From this figure, we observe that the capacity is strongly coverage limited for a typical mobile with 23-dBm maximum power. For example, if 30% of the maximum power is allocated for signaling, the capacity of 200 Erlangs can only be supported in the network with site-to-site distances of less than 1 km. If these distances exceed about 1.2 km, the capacity falls below 160 Erlangs. Clearly, such sensitivity to coverage needs to be addressed either by using a lower modulation order packet format to reduce the required SNR or by managing the interference experienced by edge users. Both of these approaches, however, lead to either higher utilization or lower availability of tile–interlace resource dimensions. Consequently, for each site-to-site distance and transmit power limit, there is an optimal solution that accounts for the tradeoff between power, interference, and dimension resources. The framework of this paper can be extended to analyze this optimization tradeoff in the future. In Fig. 9, we plot the probability that the user’s 98 percentile packet latency exceeds a certain delay outage bound. In this figure, we assume that sufficient maximum transmit power is provided to achieve the target 2% user outage for the delay bound of approximately 100 ms on FL and RL. Under this

BI et al.: PERFORMANCE AND CAPACITY OF CELLULAR OFDMA SYSTEMS WITH VOICE-OVER-IP TRAFFIC

3651

R EFERENCES

Fig. 9. Probability of the user’s 98 percentile packet latency exceeding the delay outage bound on FL and RL.

condition, the shape of the distribution very weakly depends on the site-to-site distance and Erlang capacity. Hence, we only plot the result for a particular combination of 1-km site-tosite distance and 200-Erlang capacity. Fig. 9 also verifies the achievability of a certain VoIP system capacity using the user latency outage criterion. VII. C ONCLUSION In this paper, we have proposed an analytical method to evaluate the VoIP capacity and performance of OFDMA-based cellular systems. Numerical calculations are carried out on the newly standardized UMB system. Based on the analysis, we make the following interesting observations. 1) The UMB system has the potential to provide significantly higher VoIP capacity than the 1xEV-DO Revision A system. 2) Even for VoIP application, the modulation order and packet format need to be adaptive to RF conditions. Otherwise, the cell radius supportable by an OFDMAbased cellular system with VoIP may be small for a given nominal mobile power. 3) The VoIP capacity of an OFDMA-based cellular system is a result of the tradeoff among orthogonal dimensions, available transmit power, interference threshold, and tolerable delay. The optimal capacity can only be reached when these four factors are balanced. In this paper, the analysis uses a particular packet format. Future work would include the use of more power-efficient lowermodulation-order packet formats over multiple tiles [9], [10], [13] for users on the edge of the cell and fractional frequency reuse [12]. Another important topic of future research is the impact of mobility and nonideal fast cell switching on system capacity and voice quality. ACKNOWLEDGMENT The authors would like to thank D. Cui for providing linklevel simulation results and useful technical discussions.

[1] J. Song et al., “Performance comparison of 802.16d OFDM, TD-CDMA, cdma2000 1xEV-DO and 802.11a WLAN on voice over IP service,” in Proc. IEEE Veh. Tech. Conf., 2005, vol. 3, pp. 1965–1969. [2] Q. Bi, P.-C. Chen, Y. Yang, and Q. Zhang, “An analysis of VoIP service using 1×EV-DO revision a system,” IEEE J. Sel. Areas Commun., vol. 24, no. 1, pp. 36–45, Jan. 2006. [3] G. Y. Fletcher, H. G. Perros, and W. J. Stewart, “A queueing network model of a circuit switching access scheme in an integrated services environment,” IEEE Trans. Commun., vol. COM-34, no. 1, pp. 25–31, Jan. 1986. [4] Q. Wang and M. Du, “Queueing analysis and delay mitigation in the access point of VoWLAN,” in Proc. ISCIT, 2005, pp. 1122–1125. [5] D. Niyato and E. Hossain, “Queueing analysis of OFDM/TDMA systems,” in Proc. GlobeComm, 2005, pp. 3712–3716. [6] K. M. Rege and D. Sun, “A simple analytical model to estimate VoIP signaling delays in an HFC access network,” in Proc. GlobeComm, 2005, pp. 343–347. [7] N. Maeda, S. Sampei, and N. Morinaga, “A delay profile information based subcarrier power control combined with a partial non-power allocation technique for OFDM/FDD systems,” in Proc. PIMRC, 2000, vol. 2, pp. 1380–1384. [8] O. Awoniyi, O. Oteri, and F. A. Tobagi, “Adaptive power loading in OFDM-based WLANs and the resulting performance improvement in voice and data applications,” in Proc. IEEE Veh. Tech. Conf., 2005, vol. 2, pp. 789–794. [9] M. Dong and A. Gorokhov, Forward Link VoIP and BE Performance in LBC FDD, Jan. 2007. Third Generation Partnership Project 2 (3GPP2) TSG-C WG3 contribution by Qualcomm, Inc. No. C30-20070108-022. [10] J. Borran and A. Gorokhov, Reverse link VoIP performance in LBC FDD, Jan. 2007. Third Generation Partnership Project 2 (3GPP2) TSG-C WG3 contribution by Qualcomm, Inc. No. C30-20070108-023. [11] Third Generation Partnership Project 2 (3GPP2) TCG-C WG3, Joint Proposal for 3GPP2 Physical Layer for FDD Spectra, Jul. 2006. contribution by China Unicom, Huawei Technol., KDDI, LG Electronics, Lucent Technol., Motorola, Nortel Networks, QUALCOMM Inc., RITT, Samsung Electronics, ZTE Corp. [12] S. Das, S. Li, P. Monogioudis, S. Nagaraj, S. Ramakrishna, A. N. Rudrapatna, S. Venkatesan, S. Vasudevan, H. Viswanathan, and J. Zou, “EV-DO revision C: Evolution of the CDMA2000 data optimized system to higher spectral efficiencies and enhanced services,” Bell Labs Tech. J., vol. 4, no. 11, pp. 5–24, Jan. 2007. [13] J. Borran, Reverse Link VoIP and Outage Reduction Using Extended Frames, Sep. 2006. Third Generation Partnership Project 2 (3GPP2) TSGC WG3 contribution by Qualcomm, Inc. No. C30-20060911-086. [14] A. J. Viterbi, A. M. Viterbi, and E. Zehavi, “Other-cell interference in cellular power-controlled CDMA,” IEEE Trans. Commun., vol. 42, no. 2–4, pp. 1501–1504, Feb.–Apr. 1994. [15] D. Staehle, K. Leibnitz, K. Heck, B. Schroder, A. Weller, and P. Tran-Gia, “Approximating the other cell interference distribution in inhomogeneous UMTS networks,” in Proc. IEEE Veh. Technol. Conf., 2002, vol. 4, pp. 1640–1644. [16] T. Liu and D. Everitt, “Analytical approximation of other-cell interference in the uplink of CDMA cellular systems,” in Proc. IEEE Veh. Technol. Conf., 2006, vol. 2, pp. 693–697. [17] K. Stamatiou and J. G. Proakis, “A performance analysis of coded frequency-hopped OFDMA,” in Proc. IEEE WCNC, 2005, pp. 1132– 1137. [18] P. T. Brady, “A technique for investigating on–off patterns for speech,” Bell Labs Syst. Tech. J., vol. 44, no. 1, pp. 1–22, 1965. [19] P. T. Brady, “A model for generating on–off speech patterns in two-way conversations,” Bell Labs Syst. Tech. J., vol. 48, no. 9, pp. 2445–2472, 1969. [20] Third Generation Partnership Project 2 (3GPP2), Enhanced Variable Rate Codec, Speech Service Option 3 for Wideband Spread Spectrum Digital Systems, 2004. [21] L. Kleinrock, Queueing Systems. Volume 1: Theory. New York: WileyInterscience, 1975. [22] The Members of Technical Staff, Bell Labs, Handbook of CDMA System Design, Engineering, and Optimization. Englewood Cliffs, NJ: Prentice-Hall, 2000. [23] A. J. Viterbi, CDMA: Principles of Spread Spectrum Communication. New York: Addison Wesley, 1995. [24] “Third generation partnership project two (3GPP2),” WG 5 Evaluation AHG, ‘1xEV-DV Evaluation Methodology–Addendum [V6],’ Jul. 25, 2001.

3652

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 6, NOVEMBER 2008

Qi Bi (SM’92) received the M.S. degree from Shanghai Jiao Tong University, Shanghai, China, and the Ph.D. degree from Pennsylvania State University, University Park. He was an Assistant Professor for one year with Utah State University, Logan. In 1988, he was a Member of Technical Staff with Bell Laboratories, where he became a Distinguished Member of Technical Staff in 1995 and was promoted to Technical Manager two years later. He is currently with Alcatel-Lucent, Whippany, NJ. He is a recognized expert in wireless communications. He has extensively published in many technical journals and conference proceedings and has served as Editor for many technical publications. He has often been invited as a keynote speaker to many international conferences and has filed more than 60 U.S. patents. Dr. Bi is dedicated to international and social activities. From 1998 to 1999, he served as the technical chair for the Wireless Mobile ATM Conference. From 1999 to 2000, he organized the first and the second CDMA conferences at Lucent Technologies. From 2000 to 2002, he served as the Technical Chair of IEEE Globecom’s wireless program. In 2003, he served as the Technical Chair of the IEEE Wireless Communications and Network Conference. Since 2006, he has served as the Chair or organizer of the Wireless and Optical Communications Conference. From 2002 to 2006, he was also the New York Chapter President of the Alumni Association of Chiao Tung University. He has been listed in Who’s Who since 2003. He received Awards of Excellence from the Advanced Technology Lab of AT&T in 1996 and 1997 and received Bell Labs President’s Gold Awards in 2000 and 2002. Under his leadership, his team was recognized for its outstanding contributions in innovations and was awarded the Bell Labs Innovation Team Award in 2003. In 2004, he received the Speaker of the Year Award from the IEEE New Jersey Coast Section. Based on his pioneering contributions in wireless communications, he broke ground in 2003 by becoming the first native of mainland China since 1949 to receive the prestigious Bell Laboratories Fellow Award. In 2004, he received the Asian American Engineer of the Year Award during Engineers Week in the United States.

Stan Vitebsky (M’94) received the M.S. and Ph.D. degrees in electrical engineering from the Polytechnic University, Brooklyn, NY, in 1994 and 1996, respectively. From 1995 to 1996, he was a Research Associate with the Department of Electrical and Computer Engineering, Duke University, Durham, NC. From 1996 to 1997, he was a Senior Engineer with the RF Communications Division, Harris Corporation, Rochester, NY. Since 1997, he has been with the Wireless Business Group, Lucent Technologies (now Alcatel-Lucent), Whippany, NJ, where he was a Member of Technical Staff and became a Distinguished Member of Technical Staff in 2004 for his contributions to the successful deployment of third-generation, high-speed packet data wireless networks. He is the holder of numerous patents in the area of power control and resource management for wireless networks. He is currently involved in the performance analysis of next-generation wireless communication systems. Dr. Vitebsky is a member of the Eta Kappa Nu Electrical Engineering Honor Society. He received the Bell Labs President’s Gold Award in 2003.

Yang Yang received the B.S. degree in electrical engineering from the University of Science and Technology of China, Hefei, China, in 1994 and the M.E. and Ph.D. degrees in electrical engineering from Stevens Institute of Technology, Hoboken, NJ, in 1997 and 1999, respectively. Since 1999, she has been a System Engineer with Lucent Technologies (later Alcatel-Lucent), Whippany, NJ. Her current research interests and activities are focused on the performance analysis of third- and forth-generation wireless systems, algorithm designs for air interface control and optimization, traffic modeling, and engineering of wireless networks.

Yifei Yuan (S’96–M’00) received the B.Eng. and M.Eng. degrees from Tsinghua University, Beijing, China, in 1993 and 1996, respectively, and the Ph.D. degree from Carnegie Mellon University, Pittsburgh, PA, in 2000. Since 2000, he has been with Alcatel-Lucent, Whippany, NJ, first in the Open Innovations Lab and then in the RF Performance Analysis Group, where he worked on signal processing, intelligent antennas, UMTS channel element design, EV-DO Broadcast Multicast service, and OFDM-based 4G wireless standards. His research interests include multiple-antenna systems, equalization, error control coding, and radio resource management.

Qinqing Zhang (S’95–M’98–SM’03) received the B.S. and M.S.E. degrees in electronics engineering from Tsinghua University, Beijing, China, and the M.S. and Ph.D. degrees in electrical engineering from the University of Pennsylvania, Philadelphia. She was previously with Bell Laboratories, Alcatel-Lucent Technologies, Holmdel, NJ. Since May 2007, she has been a Senior Research Scientist with the Milton Eisenhower Research Center, Applied Physics Laboratory, Johns Hopkins University, Baltimore, MD. She is also a Research Assistant Professor with the Department of Computer Science, Johns Hopkins University, and an Adjunct Assistant Professor with the Department of Electrical and System Engineering, University of Pennsylvania, where she teaches a graduatelevel course and supervises students. She was an Invited Committee Member for a Ph.D. defense at University of Pennsylvania and Rutgers University, New Brunswick, NJ. She is the coauthor of Design and Performance of 3G Wireless Networks and Wireless LANs (Springer, 2005) and the author of a book chapter in The Handbook of Computer Networks (Hoboken, NJ: Wiley, 2007). She has published numerous papers in IEEE journals and conference proceedings. She has more than 30 awarded and pending patent applications in the area of radio resource management, media access and radio link control, and IP networking. She has been working on the design and performance analysis of wireline and wireless communication systems and networks, radio resource management, algorithms and protocol designs, and traffic engineering. Her current research interests are in mobile ad hoc networks, distributed sensor networks, cooperative communications and networking, and assurable communication system and networks. Dr. Zhang has been the recipient of numerous awards and scholarships, including the Bell Labs Advanced Technology Recognition Awards in 1999 and 2001, the Bell Labs President’s Gold Award in 2002, and a fellowship and scholarships from the University of Pennsylvania and Tsinghua University. She serves on the editorial board of the IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS. She is the Cochair of IEEE International Conference on Communications (ICC2008) workshop on Cooperative Communications and Networking. She serves as the Chair of Wireless and Sensor Networks Track of the 2007 and 2008 Military Communications Conference (Milcom2007 and Milcom2008). She has been serving on the technical program committees of various IEEE conferences, including IEEE Globecom, ICC, WCNC, VTC, MWC, and Milcom.