Application of Closed Loop Resource Allocation for High Data Rate ...

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 6, NO. 11, NOVEMBER 2007

4049

Application of Closed Loop Resource Allocation for High Data Rate Packet Transmission Giovanni E. Corazza, Senior Member, IEEE, Stefano Cioni, Member, IEEE, and Roberto Padovani, Fellow, IEEE

Abstract— High data rate (HDR) CDMA systems are designed for cellular data distribution networks with maximized overall throughput and quality of service (QoS). Both figures of merit rely critically on the adoption of highly performant medium access control (MAC) algorithms, which are the focus of this paper. In particular, the Closed Loop Resource Allocation (CLRA) MAC algorithm for the reverse link of an HDR system is described. CLRA entails different properties and advantages that make it attractive for application in HDR or similar mobile Internet networks. First of all, CLRA is a macro control loop with distributed intelligence stimulated by busy tones issued by the base stations to inform terminals about their current load. Rate allocation is performed at terminals, whereby fairness is ensured by a stochastic rate assignment algorithm based on data queue states, rate history, and minimum rate guarantee. Network stability is guaranteed by controlled ramping driven by base stations busy tones. Further, a specific feature is introduced to account also for possible effects on base stations which are not in the rake active set; this feature is identified as candidate set protection. Finally, power headroom margins are used to minimize the occurrence of outage events. The analytical characterization of this MAC algorithm is derived using Markov chain modeling, and the results confirm the CLRA robustness, fairness and efficiency. The proposed algorithm is applicable to the 1xEV-DO system standardized by 3GPP2 within the cdma2000 family. Index Terms— CDMA, resource allocation, medium access control, throughput optimization.

I. I NTRODUCTION

M

ODERN cellular networks must support a variety of services, whereby users with different applications, time varying rates and quality of service (QoS) requirements have to be efficiently accommodated [1]. While voice was the primary service in second generation systems, third generation system designers have focused their attention upon packet data transmission, with emphasis on Internet/Intranet access, characterized by asymmetry and less stringent latency requirements. This trend is being emphasized further in beyond 3G and 4G evolutions. In view of the allowed flexibility, a mobile station (MS) may set up and modify sessions for voice, data, image and video through wireless channels to the base station (BS). In order to provide such services, the network must be able to statistically multiplex users with Manuscript received April 20, 2006; revised November 3, 2006; accepted January 16, 2007. The associate editor coordinating the review of the paper and approving it for publication was D. I. Kim. The Closed Loop Resource Allocation (CLRA) is covered by a U.S. patent, No. 6.563.810, issued 13 May 2003, owned by QUALCOMM Inc. G. E. Corazza and S. Cioni are with the University of Bologna, Department of Electronics Computer Science and Systems, Viale Risorgimento, 2 - 40136 Bologna Italy (email: {gecorazza, scioni}@deis.unibo.it). R. Padovani is with QUALCOMM Incorporated, 5775 Morehouse Drive, San Diego, CA 92121-1714, USA. Digital Object Identifier 10.1109/TWC.2007.060173.

different rates and QoS requirements while maximizing the overall spectral efficiency. In recognizing these requirements, the CDMA technical specification group (TSG-C) within the Third Generation Partnership Project 2 (3GPP2) has defined an evolution of the cdma2000 standard [2], [3], commonly known as 1xEV-DO, stemming from the work in [4]. In this paper, we focus on a protocol which is applicable to several high data rate (HDR) networks, including specifically 1xEVDO [5]. The characteristics of the code division multiple access (CDMA) for cellular systems are well known [6]. As compared to orthogonal systems, as time division multiple access (TDMA) or frequency division multiple access (FDMA), resource allocation between cells and within the same cell is more flexible and simplified to a certain extent. However, to guarantee large network throughput and controlled latency, efficient medium access control (MAC) algorithms must be designed and enforced. As a matter of fact, it can be stated that the overall network performance depends fundamentally on the characteristics of the adopted MAC algorithm, the purpose of which is to let the MSs select an optimized data rate for packet transmission, according to selected criteria and requirements. Indeed, in CDMA systems the received power at the BS is the resource to be shared. From this point of view, power control in itself is a form of resource assignment, ensuring that each MS is actually using its respective share. However, power control is not sufficient when variable data rates are available for the MSs, so that the targets for power control are also variable. However, the design of a rate allocation algorithm for the return link of a CDMA network is very complex due to the vast number of degrees of freedom and the difficulty in modeling accurately the real field conditions. Indeed, the search for optimal resource allocation algorithms in CDMA networks has received specific attention in the scientific literature, where different approaches to maximize system capacity are presented. In particular, Hanly in [7] has discussed a combined power control and cell-site selection algorithm that converges to a number of users per cell which is optimal in the sense that inter-cell interference is minimized. The algorithm is distributed in the sense that BSs measure total interference and broadcast this information to the terminals which compute their power allocation based on path-loss estimates. Although the analysis in [7] allows different users to have different data rates, these are supposed to be fixed, i.e., there is no mechanism for dynamically increasing or decreasing rate allocations (only power control is dynamic). In [8], the concept of variable spreading factor is introduced to statistically multiplex the uplink traffic in

c 2007 IEEE 1536-1276/07$25.00 

4050

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 6, NO. 11, NOVEMBER 2007

accordance to the network traffic load. Each user is considered to be assigned to a service class with a fixed rate, and the possibility to transmit is modeled through a two states Markov chain with exponential sojourn time. Again, this model does not allow the single user to span dynamically across a range of data rates, and the exponential sojourn may not apply to every traffic source. The algorithm can be implemented either in a centralized or distributed fashion. Further, the same assumptions of users belonging to fixed-rate service classes is exploited in [9]- [10] to dynamically control the transmission power. The optimization criteria is the minimization of the total transmitted power maximizing at the same time the sum of the user rates taking into account the delay constraints. The optimization is performed centrally at the base stations. The base stations make the overall assignments also in [11], but in this case the users are allowed to have several data rates available for transmission. In this paper, we introduce a MAC algorithm, identified as CLRA (Closed Loop Resource Allocation) [12], [13], which goes beyond the previous approaches by allowing the users to increase/decrease their instantaneous data rates depending on both their own traffic requirements and on the network load, with minimal signaling traffic and completely de-centralized decisions. In other words, in CLRA the intelligence is distributed to MSs, which select their transmission rate according to a sequential procedure. CLRA is robust in that it does not make any assumption about the traffic source model: it is applicable to any kind of service or application. CLRA has become the standardized resource allocation technique for the reverse link of the 1xEV-DO system [2], [5], [14]. The CLRA algorithm has been designed to satisfy several requirements: efficiency, in terms of responsiveness and traffic demand, network stability, fair capacity distribution among users, minimized outage, and lightweight signalling. Essentially, CLRA is a macro control loop with all MSs on one side and the network of BSs on the other side, in which the data rate selection is driven by minimal feedback provided by each BS in the form of a busy tone. In this way, radio signalling is kept to a minimum both in the forward and reverse links, with no backhaul signaling. Therefore, CLRA is independent of the network architecture. The busy tone is issued on the basis of the traffic load the BS is experiencing, identified as the rise over the noise-floor, i.e., the ratio of total received energy per chip over thermal noise power spectral density, I0 /N0 (dB). In particular, two approaches are possible: hard busy tone (HBT) and soft busy tone (SBT). In practice, the HBT consists of a single bit, while the SBT is quantized to n bits, with n ≥ 2. The problem is then to estimate whether the advantages brought about by the use of SBT are worth the increased signalling weight and rate assignment complexity. From the results in this paper, the conclusion is that the HBT choice is more sensible from a practical point of view. The CLRA algorithm is conceived to be simple, efficient, fair,and stable. Simplicity and efficiency are achieved by relegating the data rate allocation decision to the MSs, which occurs on every packet transmission to give more flexibility and rapidity. Fairness in allocating user data rates is ensured by a stochastic rate assignment algorithm based on data queue states, rate history, and minimum rate guarantee. Network load

stability is guaranteed by controlled user data rate ramping and by a unique feature, identified as candidate set protection, introduced to account for interference effects on base stations which are not in the rake active set. Finally, transmission power headroom margins are used to minimize the occurrence of outage events. After a detailed description of the CLRA algorithm, the paper develops an analytical method to characterize its performance. Every MS is modeled as a finite state machine (FSM), and its evolution is modeled using a Markov process. MS states are the possible transmission data rates and the transition probabilities are a function of the busy tone information received from all BSs. The Markovian model enables us to estimate the load of any BS and the transient performance, to ensure that the overall system behaves according to expectations. The agreement between simulation results and analytical estimates serves to validate the overall approach. The results in terms of system performance have been obtained using a purposely developed simulation tool, identified as M AXIM (MAC-SIMulator). This study shows in detail that the CLRA algorithm satisfies the most important MAC requirements. In particular, the system stability is demonstrated by analyzing the case of maximum traffic step starting from zero network load, in order to stress the network. We also show that the assignment procedure is indeed fair; in fact all available data rates are experienced frequently, albeit not uniformly by each MS. The paper is organized as follows. Section II describes the problem formulation for resource allocation in a CDMA reverse link. Section III contains a description of the CLRA algorithm. Section IV discusses the system model and the architecture of the MAXIM simulator. Section V contains the analytical characterization, while the numerical results are reported in Section VI. Finally, Section VII summarizes the conclusions. II. P ROBLEM F ORMULATION Considering a CDMA mobile wireless network formed by a collection of BSs and MSs, the aim of an effective resource allocation algorithm in the reverse link is to jointly maximize the throughput and the QoS as perceived by the users (including fairness), subject to a number of constraints. The throughput optimization problem can be expressed as: ⎧ N BS  ⎪ ⎪ achieve max Rj ⎪ ⎪ ⎪ i=1 j∈NM S (i) ⎪ ⎨ given that (1) ⎪ (power-constraint) Pj (R) ≤ Pmax ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ Rj ∈ {Ridle , Rmin , ..., Rmax } (rate-constraint) SNIRj ≥ ζ(R) (channel QoS-constraint) where NBS is the number of BSs in the network, NMS (i) is the set of MSs controlled by the i-th BS, and Rj is the transmission rate of the j-th MS, which is constrained to belong to a set of discrete data rates, where Rmin and Rmax are the minimum and the maximum data rates (given that the MS is transmitting), and Ridle means that no information is transmitted except for control signaling. Pj (R) is the needed transmission power by the j-th MS for sustaining its own rate

CORAZZA et al.: APPLICATION OF CLOSED LOOP RESOURCE ALLOCATION FOR HIGH DATA RATE PACKET TRANSMISSION

R (depending also on power control commands), and Pmax is the maximum transmission power for a MS. Finally, SNIRj is the instantaneous signal-to-noise-plus-interference ratio of the j-th MS, and ζ(R) is the requested Eb /N0 to correctly detect the transmitted information at rate R. The maximization problem reported in Eq. (1) is extremely difficult to solve without strong assumptions, bearing in mind that we are looking for a practical solution which must be implemented in the real field. In [9], an optimal solution is presented under the assumption that decisions are taken at a control station, i.e., using a centralized control loop, and with a set of simpler constraints. The authors have shown that this approach has a computational complexity which grows exponentially with the number of MSs and the number of BSs controlled by the same control station. To reduce the complexity, [11] proposes an assignment algorithm which is enforced directly by the BSs, thus reducing the need for backhaul signalling. The proof that this approach is stable and converges to a fixed network load is based on the results from [15]. But even in this case of semi-distributed allocation, the complexity and the required information exchange remain larger than what would be desired in the real field. Taking one step further the idea to spread intelligence at the edges of the network to bring complexity to a minimum, the CLRA algorithm adopts a completely distributed approach. In fact, the data rate allocation is selected by each MS independently, according to its own constraints (data buffer, power and link status) and on the network load information. In this sense, we can say that each MS solves a local optimization problem, but to avoid greedy behavior CLRA allows to increase/decrease the transmission rate only according to a stochastic procedure, and only with steps which are a factor of λ. The steps occur iteratively on a per-packet basis. To prove that this iterative local procedure achieves the global optimum expressed in Eq. (1) is very difficult. It could be argued that its philosophy is similar to the belief propagation concept, in which a complex global optimization problem is solved by iterative actions at local level. What we do claim is that CLRA satisfies the constraints in Eq. (1) while achieving at the same time a throughput which can be controlled by the base station (by setting the busy tone thresholds) as well as its goals for simplicity, efficiency, flexibility, and stability, as shown in the remainder of the paper. III. C LOSED L OOP R ESOURCE A LLOCATION We are now in a position to detail the structure of the CLRA algorithm, mapping the CLRA features on the most significant MAC algorithm requirements: efficiency (in terms of responsiveness and MS traffic requests), stability, fairness, minimized outage probability, and lightweight signalling. First of all, we introduce the concept of hard/soft busy tone, which is the fundamental driver for the entire CLRA control loop. Essentially, the busy tone guarantees the network stability by constantly reacting to the traffic load, and ensures proper utilization of overall system resources with minimal signalling. The busy tone information is generated by all BSs, and represents their status of experienced traffic load in terms of I0 /N0 (t). In the i-th BS, the instantaneous traffic load is averaged over a time window, identified as traffic

4051

load averaging window (TLAW), cutting down instantaneous fluctuations, to obtain: ρi =  (I0 /N0 )i (t) TLAW

(2)

The unquantized transmission of ρi would be evidently a waste of system resources. In order to make the forward link signaling as lightweight as possible, the average traffic load can be sampled and quantized to the desired accuracy. In the simplest case, this amounts to a comparison with a single threshold, leading to the HBT case. Specifically, the i-th BS transmits a single busy tone bit as follows:  1, ρi ≥ ξH BSi busy (higly loaded) (3) bi = 0, ρi < ξH BSi not busy (scarcely loaded) Interestingly, the threshold ξH can be also interpreted as the target average traffic load that we would like to achieve in the steady state. In other words, the system architect can design the overall steady state network load by judiciously selecting the threshold. The HBT information could seem to be excessively coarse. In order to refine the information sent to the MSs, it is possible to increase the number of the bits to achieve a soft busy tone representation (SBT). For the case of n = 2 bits per busy tone, the rule can be specified as follows: ⎧ 11, ρi ≥ ξ3 BSi overloaded ⎪ ⎪ ⎨ 10, ξ ≤ ρ ≤ ξ BSi highly loaded 2 i 3 ¯bi = (b1i , b2i ) = 01, ξ ≤ ρ ≤ ξ BSi stable ⎪ 1 i 2 ⎪ ⎩ 00, ρi < ξ1 BSi scarcely loaded (4) As will be shown in the numerical results, the availability of three thresholds enables faster transients in the macro control loop. The target load can be adjusted to approximate the intermediate threshold, ξ2 . The busy tone may be computed at a faster rate than that with which it is broadcast. In particular, the network refresh time (NRT) is defined as the time interval between successive busy tone transmissions. Busy tone transmission from different BSs is asynchronous and is planned according to a cluster approach similar to that used for frequency reuse in second generation cellular systems. This is done with the aim of minimizing overall network traffic correlations. The CLRA stochastic rate assignment defines a procedure to achieve fair resource allocation, system stability, and responsiveness. In fact, the MS rate allocation procedure is governed by a Bernoulli random variable (rv) with a probability density function which depends on the network traffic load expressed by the received busy tone values. Fairness is enforced by letting the Bernoulli rv be a function of the MS rate history with a minimum rate guarantee. Also, the MS offered traffic in the short-term is accounted for by the MS, which checks the length of the data buffer queue before selecting its desired data rate. It is very important to note that this approach does not need any assumption on the traffic source model, and as such it is robust. Another method to pursue network stability in CLRA is controlled ramping: the new MS data rate allocated by the algorithm is at most doubled/halved with respect to the previous request. This is a prudent approach to avoid stressing the system with instantaneous large capacity requests. At the same time, this procedure enables the network to share the

4052

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 6, NO. 11, NOVEMBER 2007

unassigned capacity among more users to again achieve fair allocation. Introducing the time index k, such that Rj (k) is the instantaneous transmission rate, it is therefore immediate to see that CLRA ensures NM S N N MS MS   1  Rj (k) ≤ Rj (k + 1) ≤ 2 Rj (k) 2 j=1 j=1 j=1

BS2 BT2

To introduce the candidate set protection feature, it is mandatory to also define the active set. The active set of any MS consists of a group of Nactive BSs directly involved in soft-handoff (SHO) procedure to avoid drop-calls due to the MS mobility or fast channel variations. On the other hand, the candidate set is defined as the group of Ncandidate BSs which are not in SHO with current user terminal, but would be valid candidates to replace those BSs in the active set which experience increasing path-loss. In other words, these BSs are not directly interacting with the MS, but they are strong enough and ready to enter the active set when channel variations occur. The candidate set protection introduces the key concept of the power distance between active and candidate set to overcome the limitation of the rake receiver, so that the new allocated rate must strike a balance between MS requests and the amount of inter-cell interference. It is important to remember that every MS receives the busy tone only from the BSs present in the active set group. Furthermore, to avoid or minimize outage events, a transmission power margin, identified as power headroom, is introduced. An outage event occurs when a MS has a data rate request (information bits present in the data buffer) which would require a transmission power increase exceeding the power headroom, due to the experienced propagation channel conditions. The CLRA application scenario is graphically depicted in Fig. 1. Hexagonal cells are considered with a BS in the center. The symbol BTi indicates the busy tone transmitted from the i-th BS. The notation Pj,i (R) stands for the received power from the j-th MS to the i-th BS that is a function of the transmission data rate R. It shall be observed that the link between a MS and a BS listed in its active set is indicated with a solid arrow, whereas a dashed arrow represents the link with a BS listed in the candidate set. A. CLRA steps First of all, the considered HDR frame structure is briefly reported. The fundamental timing unit for RL transmissions is a 1.667 ms slot. The reverse traffic channel (RTC) is spread over 32 slots and contains both control and data information. The control information is formed by the pilot and the data rate control (DRC) channels. In particular, the power control loop fixes the transmission level of the pilot/DRC channels for the j-th MS, indicated in the following with PjDRC . The available data rates belong to the set R ∈ {Ridle , Rmin , ..., Rmax }. The number of bytes associated to a HDR packet as a function of the current MS transmission rate is indicated with Lsize (R), and the possible values are summarized in Tab. I. The sequence of operations to determine a new MS data rate is now described in more detail. The algorithm mode consists

P2,2 (R)

P1,2(R)

(5)

MS1

BT1

BT0

BS1

Fig. 1.

BT0 BS0

P1,3(R) P1,1 (R)

MS2

BT0 P3,3(R)

MS3

P3,1(R)

An exemplary scenario for the application of CLRA.

of five steps and produces a new MS transmission rate every Tclra seconds. STEP 1: Buffer state The first MS decision is taken by considering its own buffer state and the desired rate Rbuffer is selected in order to empty the data queue as fast as possible (see Tab. I). Now, the first strategy to avoid network instability and to ensure a fair capacity allocation method is introduced. Let Rprevious be the data rate used during the previous packet transmission. The rate provided by this first CLRA step is: Rstep1 = min {Rbuffer , 2 ∗ Rprevious }

(6)

As before, this approach is identified as controlled rate ramping. It shall be noted that, if the MS comes from the idle state (no previous transmission occurred), the requested rate is Rstep1 = Rmin , to enable a MS to start the procedure of capacity request and allocation. STEP 2: Power headroom Let Pmax be the maximum transmission output power for a MS. Again, Pj (R) is the power required by the j-th MS to transmit the rate R. This power level is determined by two components. The first depends on the MS channel conditions (path-loss, shadowing, and terminal speed) and is fixed by the power control loop; instead the second component represents the increased power level to successfully transmit at the rate R. This second value is listed in Tab. I and is normalized to the reference power level required to transmit only the pilot/DRC channels, Pj (Ridle ) = PjDRC . According to Tab. I, the maximum sustainable rate due to the power transmission constraint can be computed as follows: Rpower = max{R | Pj (R) < (Pmax − Pmargin )}

(7)

where a margin is necessary due to both power estimate inaccuracies and to prevent future channel level fluctuations. After this second step, the new selected rate is defined as: Rstep2 = min {Rstep1 , Rpower }

(8)

CORAZZA et al.: APPLICATION OF CLOSED LOOP RESOURCE ALLOCATION FOR HIGH DATA RATE PACKET TRANSMISSION

4053

TABLE I PACKET SIZES AND INDICATIVE POWER INCREASE STEPS NORMALIZED TO PjDRC AS A FUNCTION OF THE TRANSMISSION DATA RATE FOR HDR APPLICATIONS . Data rate [kbit/s]

0

4.8

9.6

19.2

38.4

76.8

153.6

307.2

Packet size [bytes]

0

32

64

128

256

512

1024

2048

Increased power [dB]

0

3

4.8

7

9.5

12.3

15.2

18.1

It shall be observed that the Rstep1 rate is the maximum possible transmission rate for an MS at a specific instant. This step and the successive ones can only decrease it or at most confirm it. Finally, note that the rate could drop to Rstep2 = Ridle , in which case the MS keeps sending only the pilot/DRC channels. Of course, this is an outage event produced by the CLRA algorithm and is identified as power headroom outage for monitoring purposes. STEP3: Candidate set protection Let Eccs (j)/I0 be the strength in the j-th pilot/DRC channel in the candidate set, including all multi-path components pertaining to the same BS, for j = 1, ..., Ncandidate . On the other hand, let Ecas (i)/I0 be the strength of the i-th pilot/DRC channel in the active set, including all multi-path components pertaining to the same BS, for i = 1, ..., Nactive . Now, the key concept of power distance between active and candidate set is introduced for monitoring the inter-cell interference. The power distance metric can be defined as:  as  cs Ec (i) Ec (j) [dB] − max [dB] (9) Δac = min i j I0 I0 where the distance between the lowest active set link and the maximum candidate set interfering link is computed. With the aim to avoid excessive increments of inter-cell interference, we define: Rcandidate = max {R | Δac > (Pj (R) − ΔP rot ) [dB]} (10) as the maximum suitable rate according to a desired power distance, ΔP rot . The resulting rate allocation is therefore: Rstep3 = min {Rstep2 , Rcandidate }

(11)

STEP 4: Busy tone detection Every MS receives the busy tone information only from the BSs in the active set. Hereafter, both HBT and SBT case are considered. Hard Busy Tone The MS selects the maximum received busy tone within the active set, b = max {bi }. As a result, the i two possible HBT states are identified as: • b = 0 : active set load below threshold → stochastic rate increase possible • b = 1 : active set load above threshold → stochastic rate decrease necessary In practice, the state b = 0 means that all BSs in the active set are scarcely loaded (below the ξH threshold), and indicates that there is unassigned system capacity available. On the other hand, the detected value b = 1 corresponds to the case that at least one BS is overloaded, and consequently a rate decrease is required.

Soft Busy Tone In this case, the busy tone selection provides the maximum couple of bits inside the active set, (b1 , b2 ) = max{2∗b1i +b2i }, where the four possible i SBT states are identified as: • (0, 0) : active set scarcely loaded → aggressive stochastic rate increase possible • (0, 1) : active set stable → moderate stochastic rate increase possible • (1, 0) : active set highly loaded → moderate stochastic rate decrease necessary • (1, 1) : active set overload → aggressive stochastic rate decrease necessary In practice, the SBT enable to consider four BS traffic conditions. With respect to the HBT case, this allows to add two decision regions (ρi < ξ1 and ρi > ξ3 ) in which aggressive or mandatory transmission rate increase/decrease can be enforced. It is important to underline that this fourth step is the only one in which the CLRA algorithm receives a direct input from the network. In fact, during the previous steps, every MS is able to determine the possible data rate with its own information. This lightweight feedback from all BSs is a simple and effective idea to implement a distributed allocation algorithm, keeping network stability under control, and, at the same time, providing fast response to the instantaneous network traffic load variations. STEP 5: Stochastic rate assignment This is the final step that finally provides the new MS transmission rate. To avoid that all MSs change their rate at the same time, without however having centralized control, we impose that a change in rate is determined by a random Bernoulli variable γ: • •

γ = 0 with probability p, corresponds to Action (the MS increases or decreases its data rate) γ = 1 with probability (1−p), corresponds to No action (transmit again with Rprevious )

The probability p is the rate change probability (RCP) and is in essence the fundamental parameter to determine the system performance and stability. To ensure fairness and avoid greedy MS behavior, p should be a function of the MS rate history. After a lengthy and tedious optimization, we selected p to be simply a linear function of rate history, represented here by the average rate R over the last Naverage rate assignments. Obviously, p must be limited to the range [0,1]. The selected probability function can therefore be written as follows: 

 R p(R, A, B) = max 0, min 1, A + (B − A) (12) Rmax

4054

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 6, NO. 11, NOVEMBER 2007

Rstep3 < Rmin

yes

Rstep5 = Ridle

no Rstep3 < Rprev.

yes

Rstep5 = Rstep3

no no

yes

b=0 (busy tone detection)

yes

γ=0

Rstep5 =max[Rprev./2 ,Rmin] Fig. 2.

no

no

Rstep5 = max[Rprev. , Rmin ]

γ=0

yes

Rstep5 = Rstep3

two. Each antenna receives a signal affected by Rayleigh fading with a specified Doppler spectrum and shadowing; notably, Rayleigh fading is assumed to be spatially uncorrelated, while shadowing can in practice be regarded as perfectly correlated. Further, balanced antenna diversity is considered, i.e., the average received power level on both antennas is assumed to be the same. Therefore, the total received energy per chip over thermal noise power spectral density at the -th antenna of the i-th BS at time instant k can be written as: 1 I0 × (13) (i, , k) = N0 N0 W N MS  × PjDRC (k)Pjrate (k)L(i, j, k)α(i, j, , k)G() j=1

Flow-chart for rate assignment procedure.

= where A is probability value for null average rate while B is probability value for maximum rate (e.g., 307.2 kbit/s). The values of the parameters A, B, depend on the detected busy tone state and determine stochastic increase, SI = p(R; AI ; BI ) or stochastic decrease, SD = p(R; AD ; BD ). By observing Eq. (12), it can be noted that if a stochastic increase is possible, low data rate users have larger probability to increase their transmission rate in the next slot (because AI > BI ); instead, in case a stochastic decrease is forced, high data rate users have larger probability to decrease their transmission rate (because AD < BD ). This ensures fairness. Note that it is possible to induce deterministic increase/decrease by simply setting A = B = 1. Based on the BT state and on the outcome of drawing the rv γ, the rate assignment goes as in the flow-chart contained in Fig. 2. It shall be noticed that the stochastic rate assignment is performed if and only if the MS desired rate, Rstep3 , is not smaller than the previous transmission rate. Then, according to the busy tone state and the following value of γ, MS can double, halve or hold its own data rate. Finally, it shall be noted that, in any case, the minimum transmission rate, Rmin , is guaranteed.

NMS 1  Ec (i, j, , k)Pjrate (k) N0 j=1

 = 1, 2.

where W is the channel bandwidth, NMS is the total number of MSs, PjDRC (k) is the pilot/DRC channel transmission power of the j-th MS, Pjrate (k) is the normalized power increase due to the instantaneous data transmission rate equal to R (see Tab. I), L(i, j, k) is the path loss including a distance factor and shadowing, α(i, j, , k) is the Rayleigh fading power, and finally G() is the antenna gain. Thus, the notation Ec (i, j, , k) represents the received pilot/DRC channel strength from the j-th MS. It is evident that the instantaneous traffic load is different on the two antennas. However, the differences will certainly be much smaller when the traffic load is averaged over a sufficiently large time window. In the BS receiver, the two antenna outputs are linearly combined according to the maximal ratio combining (MRC) principle. Thus, the signal-to-noise-plus-interference ratio (SNIR) for the j-th MS measured by the i-th BS at time instant k can be expressed as: SNIR(i, j, k) =

Ec (i, j, 2, k) Ec (i, j, 1, k) + N0 + I0 (i, 1, k) N0 + I0 (i, 2, k)

(14)

The performance of the CLRA algorithm in realistic scenarios has been evaluated through the use of an event driven simulator. The approach allows to see a snap-shot of the network with several MSs randomly scattered inside a multicell coverage of a specified service area. In particular, the simulator has been used to evaluate the CLRA algorithm from the stability and fairness point of view. In the following we describe the adopted channel model, power control loop, busy tone generation approach, and traffic source modeling.

This SNIR value is used to drive the power control algorithm for the j-th MS. In particular, every HDR slot (i.e., 600 Hz update rate) the BS computes the instantaneous SNIR value and, after comparing it to the target SNIR level, ξSNIR , transmits a single power control bit (up or down command). Correspondingly, in every slot, each MS reads the power control bit and increases/decreases the transmission reference power of the pilot/DRC channels. In this procedure, a power control bit detection error is introduced: a binomial rv with probability of error event πP is considered to simulate a more realistic power control loop.

A. Channel model, signal combining, and power control

B. Busy tone generation and detection

The channel model foresees shadowing and Rayleigh fading, as appropriate for cellular environments. However, fullfledge simulation of time correlated Rayleigh fading for a MAC algorithm which accounts for both active and candidate sets and antenna diversity is necessary, with a complexity which increases dramatically with the number of MSs, BSs, antenna diversity order, and multi-path diversity order. A reasonable assumption is to limit the antenna diversity order to

As pointed out before, the essential information to be provided by the network to all user terminals is the traffic load measured at each BS. At regular time intervals, the averaged traffic load, ρi , derived from Eq. (13) and Eq. (2), is compared to one or more thresholds to determine the state of the current BS. Consequently the busy tone information serves as an up or down signal for the aggregate traffic pertaining to this particular BS. In order to simulate possible inaccuracies in

IV. P ERFORMANCE C HARACTERIZATION

CORAZZA et al.: APPLICATION OF CLOSED LOOP RESOURCE ALLOCATION FOR HIGH DATA RATE PACKET TRANSMISSION

4055

traffic load measurements, a lognormal error rv with standard deviation σL is added. To introduce time correlation, the error has been filtered inside a traffic error window (TEW). Similarly to the power control loop, a busy-tone bit detection error, as a binomial rv, with error probability πB is simulated.

cumulative distribution function which can be calculated using the incomplete gamma function, Γ(a, x): x 1 ta−1 e−t dt (21) Γ(a, x) = Γ(a) 0

V. A NALYTICAL M ODELLING

y α e− β pY (y) = α+1 β Γ(α + 1)

y

The analytical characterization of the traffic load evolution for a network regulated by the CLRA algorithm is an extremely complex task. In order to achieve a tractable model, the operation of a MS can be represented by a Markov process. In particular, a MS is modelled as a FSM, whereby the states are characterized by the current transmission data rate among the set of possible values. The transition probabilities are a function of the information delivered by the network in the form of a busy tone received from all BSs in the active set. The analysis is focused on single cell, BS 0 , with NMS mobile stations and the HBT approach. The load from external cells is equivalently taken into account in the overload and outage thresholds introduced in the following. Let s be the index numbering the states in the Markov chain, s = 0, ..., NR , where NR is the number of possible transmission rates and s = 0 corresponds to the idle mode. Letting SI s (SDs ) be the stochastic increase (decrease) probability functions according to (12), the state transition probabilities can be written as: s Pinc (k) =

[1 − Pbt (k) − Pout (k)]SI s

(15)

s Pdec (k) Pcs (k)

Pbt (k)SD s s 1 − Pout (k) − Pinc (k) − Pdec (k)

(16) (17)

= =

s

where k is the time index, Pbt is the overload traffic probability (i.e., the probability to receive a busy tone set to 1), Pout is the outage probability, and finally Pinc , Pdec , and Pc are the rate increase, decrease and confirm probabilities, respectively. The state probabilities associated to the s-th state of the Markov chain, Πs (k), can be derived at time instant k exploiting the Markovian nature of our model: (18) Π0 (k) = Pout (k − 1) s−1 s s s s−1 Π (k) = Pc (k−1)Π (k−1)+Pinc (k−1)Π (k−1)+ s+1 +Pdec (k − 1)Πs+1 (k − 1) ΠNR (k) = PcNR (k−1)ΠNR (k−1)+

NR −1 +Pinc (k−1)ΠNR −1 (k−1)

where Γ(a) is the complete gamma function. According to the one-sided CLT, the χ2 -distribution is given by

(19)

(22) 2

2

where the characteristic parameters are α = μσ2 − 1, β = σμ , N N MS MS   with μ = μj and σ 2 = σj2 , where μj and σj2 are j=1

j=1

the mean and variance for the j-th MS power. To complete this characterization and to evaluate dynamically the Markov process evolution, the last step is the computation of Pbs (k) and Pout (k), as follows:

 ξ  Pout (k) = Prob {Ptot (k) > ξ } = 1−Γ α+1, (23) β Pbt (k) = Prob {ξ < Ptot (k) < ξ  }



 ξ ξ = Γ α+1, − Γ α+1, (24) β β where ξ represents the traffic overload threshold, and ξ  is the outage threshold, which account for the overall load of BS 0 . These thresholds are a design parameter which is under control of the network operator, and that can be optimized numerically or empirically. According to the definition given in Eq. (21), it shall be noted that Eq. (23) and (24) are derived from the following basic relationship: y ξ y α e− β dy (25) CDFY (ξ) = α+1 Γ(α+1) 0 β

 βξ 1 ξ α −x = x e dx = Γ α+1, (26) Γ(α+1) 0 β Through a numerical approach it is possible to analyze the Markov process for evaluating convergence and stability of the CLRA algorithm. A vast testing campaign has been carried out. A few results are reported here and compared in the next Section to simulation results obtained in accordance with the scenarios described in Section IV, showing astounding match. VI. N UMERICAL R ESULTS

(20)

where Π0 (k) is the idle state probability and ΠNR (k) is the maximum rate state probability. Now, the interesting fact is that, according to Tab. I there is a link between transmission rate and MS transmitted power. Therefore from the transmission rate probabilities, the transmission power step probability can be computed for each MS. The total power received at the BS, Ptot (k), is a rv resulting from the sum of NMS power distributions. According to the central limit theorem (CLT) [16], Ptot (k) could be modelled as a gaussian variable. However, observing that the power received is strictly positive, a better approximation can be found by using the causal form of the central limit theorem (one-sided CLT). Hence, the total received power has a χ2 -distribution with a

A. MAXIM parameters The baseline parameters for the simulator are reported in the following. The service area is covered by 7 BSs: an internal cell surrounded by an external tier of 6 cells, as shown in Fig. 1. It is assumed that 8 MSs per BS (NMS = 56) receiving data from the network with a dedicated streams of traffic, which are assumed to be always “ON” in order to stress the system response. The results are referred to the internal cell, BS 0 , which is loaded from all sides, and to the only MSs which have in their active set this internal cell. The NRT interval, which sets the rate at which busy tones are updated, is taken to be either 128 HDR slots or 256 HDR slots. The TLAW size, i.e., the window over which the instantaneous traffic load is averaged, is taken

to be equal to 32 HDR slots. The TEW length is taken to be 4 and the lognormal standard deviation σL is 1 dB. The ξSNIR value, as measured on the reverse link pilot/DRC channels, is fixed at -19 dB for all MSs. The maximum normalized Doppler frequency shift is set to 0.15, corresponding to approximately 50 km/h. This choice corresponds to worst case where power control is imperfect (as opposed to a slower channel case), and interleaving is also imperfect (as opposed to a faster channel case) due to channel correlation. Concerning CLRA parameters, it is assumed that Pmargin = 3 dB, Nactive = 3, Ncandidate = 7, ΔP rot = 5 dB, and Naverage = 2. In order to guarantee a fair comparison between the HBT and the SBT approach, it is important that both cases achieve the same average traffic load, ξT . It is evident that for the HBT case this means ξH = ξT . On the other hand, in the SBT approach the investigated threshold set is ξ2 = ξT , ξ1 = 5 dB, and ξ3 = 15 dB. The results have been obtained averaging over 30 “snapshot” simulations which last respectively 4000 and 6000 HDR slots for the two values of NRT=128 and 256 HDR slots, respectively. Finally, all simulations assume πP = πB = 0.05.

Average traffic load [dB]

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 6, NO. 11, NOVEMBER 2007

16 14 12 10 8 6 4 2 0 -2 -4 -6 -8 -10 -12 -14

HBT SBT

1

251 501 751 1001 1251 1501 1751 2001 2251 2501 2751 3001 3251 3501 3751

Time slot number

Fig. 3. cases.

Comparison on BS 0 average traffic load for both HBT and SBT

6.00% HBT SBT

5.00% MS percentage

4056

4.00% 3.00% 2.00% 1.00% 0.00% 1

B. Rate Change Probability Optimization As can be noted from the description in the previous sections, there is indeed a large number of parameters in the CLRA algorithm that can be used to tune system performance in spite of its simplicity. However, a few of them have similar impact on performance. Therefore, a first step in parameter optimization is to identify the baseline configurations which can be used as a reference to assess the impact of variations of single parameters. Following an exhaustive simulation campaign, it was verified that the RCP functions which determine the parameter p for stochastic increase/decrease are instrumental in determining system performance. Therefore, to identify a baseline configuration it is mandatory to select a particular probability function for stochastic increase/decrease. As a result of this initial optimization campaign, the following RCP parameters according to Eq. (12) were selected as the baseline configuration in fading channel conditions. For the HBT case, the values A = 0.4 and B = −0.1 are chosen for stochastic increase (SI), while for stochastic decrease (SD) A = −0.1 and B = 0.4 are considered. On the other hand, for the SBT case, there are four RCPs in correspondence of the four decision regions associated with three thresholds. The selected RCPs are: SI1 ) A = 0.6 and B = 0.1; SI2 ) A = 0.4 and B = −0.1; SD3 ) A = −0.1 and B = 0.4; SD4 ) A = 0.1 and B = 0.6. The comparison between the HBT and SBT cases in terms of average traffic load, outage statistics, rate distribution statistics, and MS mean throughput is reported in Figs. 3, 4, 5, 6, respectively. The following considerations are in order. First of all, observing Fig. 3, it shall be noted that both approaches are stable: after the transient, during which the network load is growing towards the target value, ξT , the steady state is reached, albeit with possible traffic load oscillations. Fig. 4 shows all outage contributions for HBT and for SBT: it can be observed that the overall instantaneous outage is slightly in favor of SBT, but in both cases the overall outage is below

3

5

7

9

11 13 15 17 19 21 23 25 27 29 31

Network refresh time interval

Fig. 4. cases.

Comparison of the overall outage statistics for both HBT and SBT

4%. The rate distribution is reported in Fig. 5; it can be seen that it is a fair distribution where all available data rates are experienced by the MSs. The average value achieved for MS throughput, reported in Fig. 6, is in favor of the HBT case, while the rise time is shorter for SBT. Most of the HBT vs. SBT comparisons are either tied or slightly in favor of SBT; but nowhere the advantages of SBT appear to be dramatic. Indeed, it may be surprising to see that with an appropriate choice of probability function, the HBT solution can achieve practically the same performance of the SBT solution. In view of its simplicity, HBT appears to be therefore the preferable solution. In the following discussions, the HBT approach is assumed as the reference. To analyze the robustness of the CLRA algorithm, in the next Section the sensitivity of the baseline configuration is tested as a function of variations of all relevant parameters. C. CLRA Sensitivity vs. main parameters To evaluate sensitivity and robustness, several parameters are perturbed: the traffic load threshold, the traffic load estimation error, σL , the NRT, and the Naverage value. The results are collected in Tab. II, from which several observations can be drawn. An increase in NRT from 128 to 256, given the same threshold values, produces longer rise times and lower outage. Changing the ξT value from 10 to 15 dB produces an increase in throughput, which however is paid in terms of outage. However, the increase in outage can be significantly alleviated by increasing NRT from 128 to 256. On the other hand, reducing ξT from 10 to 8 dB produces opposite results, as expected. In general, performance variations are evident but

CORAZZA et al.: APPLICATION OF CLOSED LOOP RESOURCE ALLOCATION FOR HIGH DATA RATE PACKET TRANSMISSION

120

100% 80% 307.2 kbs

70%

153.6 kbs

60%

76.8 kbs

50%

38.4 kbs 19.2 kbs

40%

9.6 kbs

30%

4.8 kbs 0 kbs

20%

Average traffic load

90%

100 80 60 40 20

10%

4096

3840

3584

3328

3072

2816

2560

2304

2048

11 13 15 17 19 21 23 25 27 29 31

1792

9

1536

7

1280

5

768

3

1024

1

512

0

0

0%

256

Mean rate distribution

4057

Time slot number

Network refresh time interval

(a) HBT approach

Fig. 7.

BS 0 analytical traffic load simulation.

100%

100%

80% 307.2 kbs

70%

153.6 kbs

60%

76.8 kbs

50%

38.4 kbs 19.2 kbs

40%

9.6 kbs

30%

4.8 kbs 0 kbs

20% 10% 0% 1

3

5

7

9

11 13 15 17 19 21 23 25 27 29 31

Analytical mean rate distribution

Mean rate distribution

90%

Network refresh time interval

90% 80% 307.2 kbs

70%

153.6 kbs

60%

76.8 kbs 38.4 kbs

50%

19.2 kbs

40%

9.6 kbs

30%

4.8 kbs 0 kbs

20% 10% 0% 1

(b) SBT approach

Mean rate per MS [kbit/s]

Fig. 5.

Comparison on MS mean rate distribution.

55 50 45 40 35 30 25 20 15 10 5 0

HBT SBT

1

Fig. 6.

3

5

7

9

11 13 15 17 19 21 23 25 27 29 31 Network refresh time interval

Average MS throughput vs. HBT and SBT cases.

not dramatic confirming the robustness of CLRA. A central threshold of 10 dB appears to be a very good compromise between throughput and outage. Increasing the traffic load estimation error standard deviation to a very pessimistic value of σL = 5 dB, it is apparent that the performance degradations are very small. This confirms the robustness of CLRA also in respect of the accuracy of this measure. Indeed, even larger values have been tested, and the effect is simply that the target load is not achieved, i.e., the network does not reach its wanted traffic. An increase in Naverage from 2 to 8 does not affect performance significantly. A decrease from 2 to 1 also does not appear to impair performance. This is a significant result in view of hardware implementation of the algorithm. D. Analytical results The results from the analytical approach discussed in Section V for CLRA performance analysis using the HBT approach are here reported. The two thresholds, ξ and ξ  ,

3

5

7

9

11 13 15 17 19 21 23 25 27 29 31 33

Network refresh time interval

Fig. 8.

Average MS analytic rate distribution.

have been numerically optimized in order to achieve really the same network throughput and total system outage as in the previous Section. A comparative validation between simulated and analytical results can therefore be performed by contrasting Fig. 3 and Fig. 7, where the analytical dynamic traffic load evolution is reported. The agreement is more than satisfactory. The Y-scale changes only for a normalization factor. A few considerations are in order. First the rise time is approximately the same, about 1400 HDR slots. Secondly the same slope variation after about 300 HDR slots can be observed. This variation is due to the increasing MS rate transmission and to the consequent reaction from the network to these fast requests, which decreases the SI probability. Similar considerations and very interesting similarities can be observed by comparing the analytic rate distribution in Fig. 8 to the simulated one, reported in Fig. 5a. Again, the agreement between the two investigated methods is very satisfactory. Also note that the distribution in steady state is such that all of the allowed data rates (from 4.8 kbit/s to 307.2 kbit/s) are used frequently but not uniformly, whereby the intermediate rates are the most used. VII. C ONCLUSIONS In this paper, an efficient MAC algorithm identified as CLRA has been presented. The CLRA approach has been designed and tested to be an efficient, simple and stable distributed resource allocation algorithm for HDR systems. A version of this MAC algorithm is presently implemented in the 1xEV-DO system. Several funding concepts, such as busy tone, stochastic rate assignment, controlled ramping, candidate set

4058

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 6, NO. 11, NOVEMBER 2007

TABLE II S ENSITIVITY OF THE BASELINE CONFIGURATION TO VARIATIONS IN THE MAIN DESIGN PARAMETERS . ξT [dB]

σL [dB]

NRT

Naverage

Rise Time [s]

Average Load [dB]

St. Dev. [dB]

Thp. [kbit/s]

Average Outage

10

1

128

2

2.209

11.16

2.96

44.19

3.346%

10

1

256

2

4.358

11.11

2.94

45.48

2.841%

15

1

128

2

2.639

14.54

3.57

47.24

6.997%

15

1

256

2

4.993

14.62

3.28

48.85

4.812%

8

1

128

2

2.141

9.61

2.66

42.75

2.383%

10

5

128

2

2.314

11.91

3.72

44.54

3.722%

10

1

128

8

2.156

11.35

3.02

43.81

3.368%

10

1

128

1

2.314

11.26

2.91

44.84

3.360%

protection, and power headroom have been introduced and analysed. Further, a simulator that enacts this MAC algorithm for the reverse link of a CDMA system for high speed packet data services has been implemented. From the simulation results, it has been shown that the HBT performance is generally comparable to that of the more complex SBT. So, in view of its simplicity, HBT appears to be the preferable solution. CLRA has also been proved to be stable and robust with respect to errors in traffic load estimation, busy tone and power control bit detection. Concerning the transient in loading the network with the useful traffic, the rise time from zero load to the 90% of target load network is in the order of 1-2 seconds and the average throughput per MS in Rayleigh fading is about 40 kbit/s. The CLRA outage probability is less than 4% even in Rayleigh fading channel conditions and, in addition, the outage duration is generally short with respect to the NRT parameter. Finally, a Markovian analytical characterization of the CLRA resource allocation algorithm has been also provided. The comparison between simulations and numerical results serves as mutual validation and stimulates future improvements and generalizations of this approach, for example by extending the Markov chain to the multi-cell scenario. R EFERENCES [1] S. Kumar and S. Nanda, “High data rate packet communications for cellular networks using CDMA: algorithms and performace,” IEEE J. Sel. Areas Commun., vol. 17, pp. 472–492, Mar. 1999. [2] cdma2000 High Rate Packet Data Air Interface Specification, 3rd Generation Partnership Project 2 (3GGP2) Std., Oct. 2000. [3] E. Esteves, “The high data rate evolution of the cdma2000 cellular system,” in Proc. Multiaccess Mobility and Teletraffic for Wireless Communications, Dec. 2000. [4] P. Bender, P. Black, M. Grob, R. Padovani, et al., “CDMA/HDR: a bandwidth efficient high speed wireless data service for nomadic users,” IEEE Commun. Mag., vol. 38, pp. 70–77, July 2000. [5] G. E. Corazza, “Closed loop resource allocation,” U.S. Patent no. 6.563.810, May 13, 2003. [6] K. Gilhousen, I. Jacobs, R. Padovani, A. Viterbi, et al., “On the capacity of a cellular CDMA system,” IEEE Trans. Veh. Technol., vol. 40, pp. 303–312, May 1991. [7] S. Hanly, “An algorithm for combined cell-site selection and power control to maximize the cellular spread spectrum capacity,” IEEE J. Sel. Areas Commun., vol. 13, pp. 1332–1340, Sept. 1995. [8] C. I and K. Sabnani, “Variable spreading gain CDMA with adaptive control for true packet switching wireless network,” in Proc. IEEE International Communications Conference (ICC) 1995, pp. 1060–1064. [9] A. Sampath, P. Kumar, and J. Holtzman, “Power control and resource management for a multimedia CDMA wireless system,” in Proc. IEEE International Symposium on Personal Indoor Mobile Radio Communications (PIMRC) 1995, pp. 21–25.

[10] S. Ramakrishna and J. Holtzman, “A scheme for throughput maximization in a dual-class CDMA system,” IEEE J. Sel. Areas Commun., vol. 16, pp. 830–844, Aug. 1998. [11] R. Rezaiifar and J. Holtzman, “Proof of convergence for the distributed optimal rate assignment algorithm,” in Proc. IEEE Vehicular Techonolgy Conference (VTC) 1999, pp. 1841–1845. [12] G. Corazza and S. Cioni, “MAXIM: a MAC simulator for the application of CLRA to HDR,” QUALCOMM, Tech. Rep., June 2000. [13] ——, “Applications of closed loop resource allocation for high data rate packet transmission,” in Proc. International Symposium on Telecommunication (IST), Sept. 2001, pp. 739–742. [14] S. Chakravarty, R. Pankaj, and E. Esteves, “An algorithm for reverse traffic channel rate control for cdma2000 high rate packet data systems,” in Proc. IEEE Global Telecommunications Conference, vol. 6, Nov. 2001, pp. 3733–3737. [15] R. Yates, “A framework for uplink power control in cellular radio systems,” IEEE J. Sel. Areas Commun., vol. 13, pp. 1341–1347, Sept. 1995. [16] A. Papoulis, The Fourier Integral and Its Applications. McGraw Hill, 1962. Giovanni E. Corazza was born in Trieste (Italy) in 1964. He received the Dr. Ing. degree (summa cum laude) in Electronic Engineering in 1988 from the University of Bologna, and a Ph.D. in 1995 from the University of Rome Tor Vergata. He is currently a Full Professor at the DEIS of the University of Bologna. He is responsible for the area of Wireless Communications inside the ARCES center of the University of Bologna. In the years 2000-2003, he held the Chair for Telecommunications inside the Faculty of Engineering. He is the Chairman of the Integral Satcom Initiative (ISI), a European industrial forum on satellite communications with more than 160 members, under the auspices of the European Commission and of the European Space Agency. In 1989-1990 he was with the Canadian aerospace company COM DEV (Ontario). In 19911998 he was with the University of Rome Tor Vergata as a Research Associate. In November 1998 he joined University of Bologna. During 1995 he visited ESA/ESTEC (Noordwjik, NL) as a Research Fellow. During 1996 he was a Visiting Scientist at the Communications Sciences Institute, University of Southern California (Los Angeles, CA), where he also held a graduate course on Spread Spectrum Systems in 2000. During 1999 he was a Principal Engineer at QUALCOMM (San Diego, CA). Prof. Corazza is author or co-author of more than 170 papers published in International Journals and Conference Proceedings. Since 1997, he joined the Editorial Board of the IEEE Transactions on Communications as Editor for Spread Spectrum. He received the Marconi International Fellowship Young Scientist Award in 1995. He was co-recipient of the Best Paper Award at IEEE ISSSTA’98, and of the Best Paper Award at IEEE ICT2001. He was co-recipient of the 2002 IEEE VTS Best System Paper Award in 2002. He was Chairman of the ASMS2004 and of the ASMS2006 Conferences and of the upcoming IEEE ISSSTA 2008 Conference.

CORAZZA et al.: APPLICATION OF CLOSED LOOP RESOURCE ALLOCATION FOR HIGH DATA RATE PACKET TRANSMISSION

Stefano Cioni was born in Imola (Italy) in 1973. He received the Dr. Ing. degree in telecommunication engineering and the Ph.D from University of Bologna, Italy, in 1998 and in 2002, respectively. From March 2002 to October 2002, he was visiting researcher at the European Space Agency (ESA) on adaptive coding and modulation techniques for future broadband satellite networks. In 2002, he joined the Advanced Research Center for Electronic Systems (ARCES) of the University of Bologna, where he is currently a Post-Doc researcher. During the summer of 2006 he was a Visiting Researcher at the Agilent Labs SMRD, Belgium. His research activities are mainly focused on the next generation wireless telecommunication systems, both the terrestrial and the satellite networks. In particular, his interests include synchronization techniques, medium access control resource allocation algorithms, OFDM systems, and iterative decoding techniques joint to channel parameter estimation. Dr. Cioni coauthored more than 30 papers and scientific conference contributions and he is co-recipient of the Best Paper Award at the IEEE International Conference on Telecommunications 2001, ICT2001, 4-7 June 2001, (Bucharest, Romania).

4059

Roberto Padovani received a Laureate degree from the University of Padova, Italy and a M.S. and Ph.D. degrees from the University of Massachusetts, Amherst in 1978, 1983, and 1985, respectively, all in Electrical and Computer Engineering. Dr. Padovani joined QUALCOMM Incorporated in 1986, after two years at M/A-COM Linkabit where he was involved in the design and development of satellite communications systems, secure video systems and error-correcting coding equipment. Over the past 20 years at QUALCOMM, Dr. Padovani has been involved in the research and development of digital communication systems with particular emphasis on Code Division Multiple Access (CDMA) wireless technology systems. He was involved in the initial design, development and standardization of IS-95 CDMA systems. His research and inventions in this field have led to the worldwide standardization and commercialization of CDMA technology for second- and third-generation cellular systems. More recently, he has led the design and development of CDMA2000 1xEV-DO, an IP-based, high-speed wide-area wireless data technology, which has led to the deployment of multiple broadband wireless networks and services across the globe. Dr. Padovani holds more than 60 patents on wireless systems. He has published numerous technical papers in the digital communications field and was the co-recipient of the 1991 IEEE Vehicular Technology Society Best Paper Award for a fundamental paper on the capacity of CDMA cellular systems. In addition, Dr. Padovani has received the Innovators in Telecommunications, 2004 award from the San Diego Telecom Council; election to the National Academy of Engineering, 2006; and the Executive of the Year, 2006 from the School of Electrical and Computer Engineering at the University of California, San Diego. He is an IEEE Fellow and an Adjunct Professor in the Electrical and Computer Engineering Department at the University of California, San Diego.