This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts

for publication in the WCNC 2009 proceedings.

A Game Theoretical Formulation for Proportional Fairness in LTE Uplink Scheduling

Elias Yaacoub and Zaher Dawy Department of Electrical and Computer Engineering, American University of Beirut, Beirut, Lebanon Email: {eey00, zd03}@aub.edu.lb
Abstract Uplink scheduling in LTE systems is considered. A game theoretical formulation is derived where the scheduling problem is represented as a Nash bargaining solution. An algorithm to implement the proposed scheduling scheme is presented. Throughput and fairness analysis are performed via simulations. Results show that channel aware scheduling schemes outperform the round-robin scheme, but a tradeoff must be made between the increase of total throughput and fairness towards the different users. Index Terms Uplink scheduling, LTE, SCFDMA, game theory, Nash bargaining solution

I. I NTRODUCTION The Long Term Evolution (LTE) of UMTS will stretch the performance of 3G technology, in order to meet user expectations in a 10-year perspective and beyond [1]. The objectives of LTE include reduced latency, higher user data rates, improved system capacity and coverage, and reduced overall cost for the operator. Data rates of 100 Mbps in the downlink and 50 Mbps in the uplink are expected. To meet these requirements, orthogonal frequency division multiple access (OFDMA) based on OFDM is the key technique of LTE because it has the immunity to inter-symbol interference and frequency selective fading [1]. However, in spite the numerous advantages of OFDM and OFDMA, their major disadvantage is that their waveforms have high peak to average power ratio (PAPR). High PAPR is problematic for uplink transmission where the mobile transmission power is usually limited. For PAPR reduction, 3GPP-LTE agreed on using single carrier frequency division multiple access (SCFDMA) transmission with cyclic prex in the uplink where frequency domain generation of the signal by a DFT precoding followed by an IFFT structure was assumed [2]. SCFDMA is a modied form of OFDMA. As in OFDMA, the transmitters in SCFDMA use different orthogonal frequencies (subcarriers) to transmit information symbols. However, they transmit the subcarriers sequentially, rather than in parallel [3]. Relative to OFDMA, SCFDMA reduces considerably the envelope uctuations in the transmitted waveform. Therefore, SCFDMA signals have inherently lower PAPR than OFDMA signals [3]. SCFDMA has two types of sub-carrier mapping [3]: Localized FDMA (LFDMA) and Interleaved FDMA (IFDMA). In LFDMA, the scheduler assigns consecutive subcarriers to convey information from a particular user. In IFDMA, users are assigned sub-carriers that are distributed over the entire frequency band in order to avoid allocating adjacent subcarriers that are simultaneously in a deep fade. IFDMA was not included into the LTE standard due to slight performance disadvantages caused by the requirements of channel

estimation accuracy [4]. Hence, in the remainder of this paper, we will only consider LFDMA. Consequently, the available spectrum is divided into resource blocks (RB) consisting of 12 adjacent subcarriers. The duration of a single RB is 1 ms, agreed to be the duration of one transmission time interval (TTI) [4]. With the high bit rates expected for LTE, it would be interesting to investigate efcient scheduling of users, especially in the uplink. Work in this direction was performed in [1], where link and system throughput are estimated by simulations, with a simple round-robin scheduling policy. More efcient scheduling algorithms were developed in [3], where the utility functions to maximize are the sum of user throughput or the sum of the logarithm of user throughput. Fairness issues are also discussed. In this work, we present a game theoretical formulation that links proportional fairness to the logarithm of the throughput as a utility function to be maximized. The proposed game theoretical formulation applies in both centralized and distributed scenarios. A heuristic scheduling algorithm achieving this proportional fairness with low complexity is then designed and tested in a simulation environment consistent with the current LTE standards. It should be noted that the proposed algorithm has the exibility to accommodate various utility functions. Thus, it is used to compare the performance of sum-throughput maximization to that of the proposed proportional fairness scheme. The paper is organized as follows. The problem formulation and the game theoretical modeling of the uplink scheduling solution are presented in Section II. The sub-optimal scheduling algorithm to implement the proposed solution is described in Section III. The simulation results are presented and discussed in Section IV. Finally, in Section V, conclusions are drawn and extensions for future research are described. II. P ROBLEM F ORMULATION We dene the RB allocation problem as a constrained utility maximization problem rst, then we present a game theoretical approach that leads to the same initial model while at the same time taking proportional fairness into account. A. Constrained Optimization Problem Letting IRB,k be the set of RBs allocated to user k, Isub,k the set of subcarriers allocated to user k, Nsub the number of subcarriers, Pk the instantaneous transmission power of user k, Pk,max its maximum transmission power, and Rk its achievable throughput, the maximization of the sum of user utilities U can be formulated as

978-1-4244-2948-6/09/$25.00 2009 IEEE

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2009 proceedings.

max
k=1

U (Rk |IRB,k )

(1) subject to

max
k=1

R(Pk , IRB,k )

(5)

subject to Pk Pk,max ; k = 1, ..., K

K

(2)

Pk Pk,max ; k = 1, ..., K
K

(6)

i,k = 1; i = 1, ..., Nsub

k=1

(3)
k=1

i,k = 1; i = 1, ..., Nsub

(7)

with i,k = 1 if subcarrier i is part of an RB allocated to user k; i.e. i Isub,k . Hence, (3) represents the constraint that each subcarrier can be allocated to a single user only during one TTI. The utility function depends on the throughput, and could vary depending on the different services and QoS requirements. Letting the utility equal to the throughput, (1) becomes a maximization of the sum-throughput of the cell. However, in this case, users close to the base station (BS) will be allocated most of the resources, whereas edge users will generally suffer from starvation. To solve this problem, utility functions providing proportional fairness are desired. In [5], it was proven that the logarithmic utility function is associated with the proportional fairness for the utility-based optimization. Hence, letting U = ln(R) provides proportional fairness, where ln represents the natural logarithm. In this section, we will illustrate this result using a game theoretical formulation that applies to both centralized and distributed scenarios, then we will propose a scheduling algorithm to maximize the sum utility in Section III. B. Game Theoretical Approach In this section, we model the resource allocation problem in the LTE uplink as a bargaining game. We consider that each user is a player (in this section, both terms user and player are used interchangeably) who wants to maximize its payoff, considered to be its throughput. Cooperation is assumed between players. Consequently, players should share the resources in an optimal way, i.e. a way they cannot jointly improve on. The resources to be shared are the RBs in each TTI. Allocating the shared resources in a way to maximize the users payoffs is equivalent to allocating RBs to users in a way to maximize each users throughput, given the shares allocated to the other users. It is a well known result in game theory that the solution to the cooperative bargaining problem maximizes the Nash product [6]:
K

Since the logarithm is a continuous strictly increasing function, (5) is equivalent to

K K

max
k=1

R(Pk , IRB,k ) = max ln

k=1 K

R(Pk , IRB,k ) (8) ln R(Pk , IRB,k )

= max
k=1

subject to the constraints in (6) and (7). Hence, this proves that the logarithm of the throughput, when used as a utility, ensures proportional fairness, since maximizing it leads to maximizing the Nash product thus leading to the Nash bargaining solution (NBS). In other words, we obtained the result of [5] from another perspective, by using a game theoretical approach. C. Centralized vs. Distributed Scenarios The NBS in cooperative game theory assumes the presence of some authority to enforce the agreement between the different players [6]. In a centralized scheduling scenario, no negotiations take place between the players. The BS is responsible for making the scheduling decisions. However, since the NBS in a bargaining scenario is equivalent to proportional fairness, it could be assumed that virtual negotiations take place between the players that reach the optimal NBS, and that the BS is the authority assumed to enforce the cooperative solution. However, in a game theoretical formulation it is implicity implied that users have more independence, which could be applicable in our case to a partially distributed scenario, where users take part in the scheduling decisions. In such a scenario, the BS would allocate a subset of the RBs to a group of users that are close enough to communicate with each other. The users would then perform a bargaining game to share the allocated RBs, with each user trying to maximize its share. Obviously, the solution to this problem is the NBS, which is the optimal (Paretto efcient) solution. The challenge in such a scenario is to determine an efcient low-overhead communication protocol between the users that allows them to exchange the information necessary to perform the scheduling operation (e.g. broadcast of channel state information of each user on each RB). III. S CHEDULING A LGORITHM In this section, we present a low complexity scheduling algorithm that applies the proposed game theoretical formulation in a centralized scenario within the LTE uplink framework. The algorithm is used in the simulations of Section IV. The

NashProduct =
k=1

Wk (xk ) Tk

(4)

where xk represents the fraction of resources allocated to player k, Wk (xk ) corresponds to the payoff of player k when xk is allocated to it, and Tk is the payoff of player k in the case where no agreement is reached in the bargaining problem. In the LTE scheduling problem, the player payoff is the throughput achieved, i.e. Wk (xk ) = R(Pk , IRB,k ), and Tk = 0 since no transmission occurs if no agreement on RB allocation is reached. Hence, the optimization problem becomes

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2009 proceedings.

proposed algorithm is a channel aware scheduling (CAS) algorithm that assumes the knowledge of the channel state information (CSI) at the BS. It should be noted that the algorithm can be used with any utility function while having the property of ensuring proportional fairness when the utility used is the logarithm of the throughput. The proposed algorithm consists of allocating RB n to user k in a way to maximize the difference n,k = U (Rk |IRB,k {n}) U (Rk |IRB,k ) (9)

is greater than U (Rk1 |IRB,k1 {n }) = X . However, in 1 this case, the increase in the sum-utility by allocating RB n to 2 user k2 is only U (Rk2 |IRB,k2 {n }) U (Rk2 |IRB,k2 ) = . 2 Since X/2 > , the maximization of marginal utility leads to better results.

A. Complexity Analysis The proposed algorithm allocates each RB after performing a linear search on the users and RBs in order to nd the user-RB pair that maximizes the marginal utility. Hence, the complexity to allocate the rst RB is O(N K), the complexity to allocate the second RB is O((N 1)K), and so on. Consequently, the total complexity of the algorithm is O N K + (N 1)K + ... + 2K + K N (N + 1) K 2 O(N 2 K) (12) =O

where the marginal utility, n,k , represents the gain in the utility function U when RB n is allocated to user k, compared to the utility of user k before the allocation of n. The CAS algorithm is described as follows: Consider the set of available RBs Iavail RB {1, 2, ..., N }. At the start of the algorithm, Iavail RB = {1, 2, ..., N }. Step1: Find an RB-user pair which has the highest marginal utility dened in (9) among all available RBs and users. For each available user k and RB n, nd the pair: max n,k (10) [n , k ] = arg
nIavail RB ,k

i.e. the algorithm has linear complexity in the number of users and quadratic complexity in the number of RBs, and thus could be easily implemented in real-time. B. Throughput Calculations For the throughput calculations, we consider the following expression:
Nsub

Step2: Allocate RB n to user k : IRB,k = IRB,k {n } Step 3: Delete the RB from the set of available RBs: Iavail RB = Iavail RB {n } (11)

Repeat Steps 1, 2, and 3, until all RBs are allocated. The following lemmas highlight some important properties of the proposed algorithm. Lemma 1: The proposed algorithm achieves proportional fairness when the utility considered is the logarithm of throughput. Proof: Let Cn be the throughput of a user when a given RB n is allocated to it and C its throughput without RB n. When the utility is the throughput, the scheduler allocates n to the user that maximizes the difference (Cn C), which is the user having the best channel conditions, i.e. in most cases the user closest to the BS. On the other hand, when the utility is the logarithm of the throughput, the scheduler allocates n to the user that maximizes the difference (ln(Cn ) ln(C)), or, equivalently, the ratio Cn /C, thus allocating n to the user that achieves the largest proportional increase in throughput.

R(Pk , Isub,k ) =
i=1

i,k

B log2 [1 + i,k ] Nsub

(13)

where B is the total bandwidth and is called the SNR gap. It indicates the difference between the SNR needed to achieve a certain data transmission rate for a practical M-QAM system and the theoretical limit (Shannon capacity) [7]. It is given by: = 1.5 ln(5Pb ) (14)

where Pb denotes the bit error rate (BER). Each user is assumed to transmit at the maximum power (Pk = Pk,max ), and the power is assumed to be subdivided equally among all the subcarriers allocated to that user. Hence, the SNR over a single subcarrier, i,k , is given by: Pk Hk,i |Isub,k | = 2 i

i,k Lemma 2: Marginal utility is capable of achieving better performance than full utility; i.e. nding the pair (n , k1 ) that 1 maximizes the marginal utility n,k leads to a sum-utility greater than nding the pair (n , k2 ) that maximizes the full 2 utility U (Rk |IRB,k {n}). Proof: As a proof outline, we consider the following case: U (Rk1 |IRB,k1 {n }) = X , U (Rk1 |IRB,k1 ) = X/2, 1 U (Rk2 |IRB,k2 {n }) = X + , and U (Rk2 |IRB,k2 ) = X, 2 with X . In this case, the proposed algorithm allocates n 1 to k1 , leading to a sum-utility increase of n ,k1 = X/2 . 1 On the other hand, an algorithm maximizing the full utility would allocate n to k2 since U (Rk2 |IRB,k2 {n }) = X + 2 2

(15)

where Hk,i is the channel gain over subcarrier i allocated to 2 user k, and i is the noise power. Subdividing the power equally over the subcarriers is justied in [8] by the fact that the achieved gains are negligible compared to the increase in complexity when optimal power allocation is performed. In fact, in the uplink scenario, the maximum transmission power of mobile users is limited, contrarily to the downlink case where the BS has considerably more power and where the variation in user distances from the BS allows the latter to achieve gains by optimizing the power allocation.

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2009 proceedings.

120

100 SumThroughput (Mbps) RR U=R U = ln(R)

80

60

40

20

10

20

30 40 Number of Users

50

60

70

be denoted by (U = R) and (U = ln(R)), respectively. It can be seen from Fig. 1 that the CAS algorithm outperforms RR for both utilities (R and ln(R)) in terms of cell throughput, and the case (U = R) has a clear advantage over (U = ln(R)). The RR algorithm allocates one RB for each user. This explains the linear increase in the sum-throughput plot of RR until the number of users reaches the number of RBs. In this case, the number of served users during one TTI will be equal to the number of RBs, with one RB being allocated to each served user. However, Fig. 1 does not show the effect of the distance from the BS on the throughput of the different users. With (U = R), users close to the BS are expected to receive most of the RBs, preventing edge users from fair access to resources most of the time. These issues will be discussed in the following section. C. Fairness Analysis To obtain an indication about the fairness of the three investigated approaches, we consider the case of eight users located at xed distances from the BS, with the distances shown in Table I.
TABLE I U SER D ISTANCES FROM THE BS. User Distance (km) 1 0.2 2 0.3 3 0.4 4 0.5 5 0.6 6 0.7 7 0.85 8 1

Fig. 1. Sum-throughput comparison.

IV. R ESULTS AND D ISCUSSION This section presents the simulation results obtained by applying the algorithm presented in Section III. The results include plots of the cell throughput, in addition to a discussion of fairness results. A. Simulation Model The simulation model consists of a single cell with a BS equipped with an omnidirectional antenna. The throughput is averaged over 200 TTIs, with the duration of a TTI being 1 msec. Then the simulation is repeated over 30 iterations. The total bandwidth considered is B = 5 MHz, subdivided into 25 RBs of 12 subcarriers each. We consider a target BER of 106 . The maximum mobile transmit power is considered to be 125 mW. All mobiles are assumed to transmit at the maximum power, and the power is subdivided equally among all subcarriers allocated to the mobile. The channel gain over subcarrier i corresponding to user k is given by: Hk,i,dB = ( log10 dk ) k,i + 10 log10 Fk,i (16)

In (16), the rst factor captures propagation loss, with a constant chosen to be 128.1 dB, dk the distance in km from mobile k to the BS, and the path loss exponent, which is set to a value of 3.76. The second factor, k,i , captures log-normal shadowing with an 8 dB standard deviation, whereas the last factor, Fk,i , corresponds to Rayleigh fading with a Rayleigh parameter a such that E[a2 ] = 1. Perfect CSI estimation is assumed at the BS. Results from the round robin algorithm are presented for reference. The implementation of the round robin (RR) is straightforward. The number of RBs is xed, and the RBs are allocated to users on a rst-come rst-served basis. B. Throughput Results Fig. 1 shows the cell sum-throughput results. In all the gures, the results corresponding to round robin will be denoted by (RR), and those corresponding to the proposed algorithm with throughput utility and logarithmic throughput utility will

Fig. 2 shows the throughput results as a function of the distance. In the case (U = R), users close to the BS achieve very high throughput, whereas those close to the cell boundary suffer from starvation. The throughput of the different users in the RR case is clearly very low regardless of the distance, but the throughput gap between users in the center and the cell edge users is reduced. From Fig. 2 it can be seen that the case (U = ln(R)) allows cell edge users to achieve higher throughput than RR and the case (U = R), while allowing the users close to the center to keep the advantage of achieving a relatively higher throughput than edge users. Hence, the simulation results validate the theoretical formulation about the proportional fairness of the logarithmic throughput utility. In fact, studying fairness from the perspective of the average number of subcarriers allocated per TTI to each of the users of Fig. 2, the results of Fig. 3 are obtained. Naturally, in the RR case, all users are allocated the same number of subcarriers by denition. This achieves perfect fairness in terms of access to the resources for the different users. However, since RR is not a CAS algorithm, this absolute fairness is not reected into throughput gains, as can be seen from Fig. 2. On the other hand, with the throughput maximizing scheduler (U = R), users close to the cell center are allocated most of the resources, leaving edge users with practically nothing, as can be seen for the user located at d = 1 km, where far more subcarriers are allocated to it by RR. However, with the logarithmic throughput maximizing scheduler (U = ln(R)), access to subcarriers is clearly more fair. All users, even at d = 1 km, are allocated more than one RB on average, as can

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2009 proceedings.

16 14 12 Throughput (Mbps) 10 8 6 4 2 0 RR U=R U = ln(R)

n to the user that achieves the largest proportional increase in throughput. This allows the allocation to edge users of RBs over which they have relatively good channel conditions (with respect to other RBs). V. C ONCLUSIONS AND F UTURE W ORK Uplink scheduling in LTE was considered. The problem was formulated as a cooperative bargaining problem, where each user aims to maximize its own utility. A Nash bargaining solution was derived for the resource allocation problem, and a heuristic algorithm to implement the solution was proposed. It was shown through simulations that the maximization of the sum throughput leads to a higher cell throughput, although considering the logarithm of throughput as a utility function ensures proportional fairness, and thus constitutes a tradeoff between throughput and fairness. In the proposed game theoretical model, centralized scheduling was considered where the BS was assumed to enforce the cooperative solution. However, a game theoretical formulation insinuates a more distributed scenario, where users take part in the scheduling decisions. Distributed scheduling in LTE uplink is thus an interesting topic for future research, where the proposed game theoretical approach can be generalized. As an enhancement to the scheduling algorithm, the optimization of the power allocation scheme could be considered. In this case, the gains of the optimal power allocation would be compared to the equal power scenario. Another natural extension to this work is the study of a multiple cell scenario where intercell interference is taken into account. ACKNOWLEDGMENT

0.2

0.4 0.6 0.8 User Distances from BS (km)

Fig. 2. Throughput achieved by each user as a function of the distance from the base station.

150 RR U=R U = ln(R) Average Number of Subcarriers

100

50

0.2

0.4 0.6 0.8 User Distances from BS (km)

Fig. 3. Average subcarrier allocation vs. distance: 8 users and 25 RBs of 12 subcarriers each.

This work was supported by the American University of Beirut (AUB), the AUB Research Board, Dar Al-Handassah (Shair and Partners) Research Fund, and the Rathman (Kadifa) Fund. R EFERENCES

be seen from Fig. 3, where the scheduler with (U = ln(R)) outperforms RR even for the cell edge user (at d = 1 km). Comparing the (U = ln(R)) scheduler to the (U = R) scheduler, it can be clearly seen in Fig. 3 that the former noticeably outperforms the latter, in terms of subcarriers allocated, starting at d = 0.5 km, while it allocates slightly less resources for users at d = 0.3 and 0.4 km. User 1 still gets the highest number of subcarriers with the (U = ln(R)) scheduler, but not as much as in the case of the (U = R) scheduler, where it is depriving other users from fair access to resources. The behavior of these two schedulers can be better understood by examining (9) and Lemma 1. In fact, if A is the throughput of a user when a given RB n is allocated to it and B its throughput without RB n, the (U = R) scheduler allocates n to the user that maximizes the difference (A B), which in most cases is the user having the best channel conditions, i.e. the closest to the BS. On the other hand, the (U = ln(R)) scheduler allocates n to the user that maximizes the difference (ln(A)ln(B)), or, equivalently, the ratio A/B, thus allocating

[1] H. Ekstr m, A. Furusk r, J. Karlsson, M. Meyer, S. Parkvall, J. Torsner, o a and M. Wahlqvist, Technical Solutions for the 3G Long-Term Evolution, IEEE Communications Magazine, vol. 44, no. 3, pp. 3845, March 2006. [2] 3rd Generation Partnership Project (3GPP), 3GPP TR 25.814. 3GPP TSG RAN physical layer aspect for UTRA, v7.1.0, tech. rep. [3] H.G. Myung, J. Lim, and D.J. Goodman, Single Carrier FDMA for Uplink Wireless Transmission, IEEE Vehicular Technology Magazine, vol. 48, no. 1, pp. 3038, September 2006. [4] T. Lunttila, J. Lindholm, K. Pajukoski, E. Tiirola, and A. Toskala, EUTRAN Uplink Performance, International Symposium on Wireless Pervasive Computing (ISWPC) 2007, February 2007. [5] G. Song and Y. Li, Cross-Layer Optimization for OFDM Wireless Networks-Part I: Theoretical Framework, IEEE Transactions on Wireless Communications, vol. 4, no. 2, pp. 614624, March 2005. [6] F. Carmichael, A Guide to Game Theory. Prentice Hall, 2005. [7] X. Qiu and K. Chawla, On the Performance of Adaptive Modulation in Cellular Systems, IEEE Transactions on Communications, vol. 47, no. 6, pp. 884895, June 1999. [8] J. Lim, H.G. Myung, K. Oh, and D.J. Goodman, Channel-Dependent Scheduling of Uplink Single Carrier FDMA Systems, IEEE VTC Fall 2006, September 2006.