A Dynamic PF Scheduler to Improve the Cell Edge Performance

Xu Ning, Vivier Guillaume, Zhou Wen, Qiang Yongquan
China Broadband Wireless Research Center Motorola Labs Beijing, China ningxu@motorola.com approach is to allocate a set of specific frequency resources for cell edge users, mitigating the inter-cell interference and thus, increasing their performance. In this paper, we investigate an approach of using scheduler itself to provide better fairness and performance for cell edge users. Specifically, we propose a Dynamic PF scheduling algorithm. The novelty consists in adapting dynamically and on a per user basis the β parameter of the PF algorithm, in order to better adapt to user specific conditions (e.g. location in the cell or QoS requirement). Our algorithm can provides better fairness and increase the seamless experience by providing more uniform QoS over the cell. The rest of this paper is outlined as follows. In section II we review and discuss traditional PF algorithm and point out its limitations in allocating resource to cell edge users. We also introduce some known modifications to the PF. Then, our Dynamic PF algorithm is presented in details in section III. Simulation results are showed in section IV and the conclusion is given in section V. II. PF AND CELL EDGE ISSUES

Abstract—In order to improve the QoS in the cell edge of broadband cellular systems, we propose to modify the Proportional Fair (PF) scheduling algorithm in a new way. In the traditional PF algorithm, a beta parameter can be used in the denominator to control the PF ratio. In our method, the beta is dynamically adjusted in a time and user depended basis, taking into account for example, the user’s receiving signal level. In each updating period, the value of beta may be different from different users. Therefore, the scheduling priority is adapted to individual user’s condition. The underlying idea is to accelerate the increase of scheduling priority when the user moves to the cell edge. Simulation results show that our dynamic PF scheduling algorithm can improve the performance of cell edge users with a limited degradation of whole system throughput. Keywords - Packet wireless systems, scheduling, cell edge performance, proportional fair

I. INTRODUCTION Packet data services with various QoS requirement play an important role in wireless communication systems, such as 3GPP LTE, 3GPP2 UWB, and WiMAX systems. Due to the variations of the wireless channel (in time, frequency and space), it is a challenge to ensure fairness and seamless experience among data users with different channel conditions and quality of service requirements. Users in the cell edge usually experience bad channel conditions, thus tend to get bad throughput and performance. Providing a consistent user experience all over the cell and especially acceptable performance at the cell borders is becoming a key requirement in future wireless communication as for instance mentioned in IMT Advanced requirement. User and packet scheduler is the main component in Media Access Control layer to control the system performance in terms of fairness and throughput. For non-real time service, Proportional Fair (PF) scheduling [1] is a well-known algorithm, which can provide long-term fairness when channel conditions of users are almost the same. In a realistic cellular scenario, when some users are in the cell edge, the averaged SINR of them will remain lower than the users in cell center. In this situation, PF can not provide good fairness and yields to performance degradation of cell edge users. Existing solution to provide better performance to cell edge users is to consider frequency reuse, or soft frequency reuse method [2], which works on top of packet scheduler. The main idea of such

A. PF Basic When providing non real time service for users, one can use standard Proportional Fair (PF) scheduling algorithm [1] to provide a good trade-off between capacity and fairness. PF scheduling is simple but effective for non real-time services. Its policy is to schedule the user j in a time slot t with:
j = arg max

rk [t ] Rk [t ]


(1 − Rk [t ] = (1 −

1 1 ) Rk [t − 1] + rk [t ], if k is scheduled Tc Tc 1 ) Rk [t − 1], if k is not scheduled Tc


where rk, in the numerator represents the instantaneous rate of the users while Rk represents the average rate of the same user, based on the past resource allocation. The PF basically allocates the resource at a given time to the user which maximizes the ratio of instantaneous rate over average rate.

978-1-4244-1722-3/08/$25.00 ©2008 IEEE.


This method can both exploit multi-user diversity gains and provide flexible fairness adjustment from proportional fairness to maxmin fairness as β increases from 1 to ∞ . it can provide more uniform QoS for user. if α = 0. In our approach. The parameter β is decreased if averaged data rate is smaller than the target and fairness index is larger than the target. In other words. The method in [8] only considers the long-term fairness and aggregate throughput in the system. The parameter β k [t ] is updated not only in time but also depended on user’s individual requirement. several modifications have been introduced. it is proposed in [7] that Tc in (2) should update with β . it is equivalent to the proportional fairness. if α = β = 1 . For OFDMA based system. When channel condition goes to bad due to user in cell edge. In [4]. In [7]. So. the equation (1) can be modified as: j = arg max k ( Rk [t ]) ( rk [t ])α β (3) can explain the influence of the parameter a as follows. in which α in equation (3) is set to a fixed. We 2 . 2. 1. the priority becomes small for a long term because of the rk dropping. to compensate the PF penalty caused by averaged SINR reduction. B. So.g. β k [t ] is updated as a function of user’s channel condition. Different user may get different value of β k [t ] in a given scheduling interval. we can simply apply this policy in each frequency sub-channel independently. 3. where. such as 3GPP LTE or WiMax. It is increased when averaged data rate is larger than the target and fairness index is smaller than the target. At last. In addition. user’s averaged SINR. This scheduling policy can be used in time-division multiplexing based packet wireless systems. If the user has got a moderate or relative large Rk when he has been in the cell center (which is a usual case). It works well when users experience homogeneous channel conditions and when the channel conditions vary rapidly. the access terminal with higher rk [t ] is chosen. as in [3]. in this situation. The update of β is performed by monitoring the averaged data rate and averaged fairness index. the transmit opportunities of a user can be improved immediately once it moves into the cell edge. a scheduling method. the user k can get scheduled only after the users in cell center get much more scheduling opportunity. then it is a round-robin (RR) scheduling. The value of β is the same for all users in one updating period. then we get PF scheduling. higherthan-one value. the authors propose to set β from 1 to ∞ . Moreover. The updating process is controlled by a monitor that checks whether the difference between the proportional data rate allocated to user and the average value over all users is within an acceptable interval. its channel condition is usually bad compared to the channel conditions in the cell center. thus preventing the user to get scheduling opportunity. the fairness of this cell edge user k cannot be compensated for a relatively long term due to the slow reaction capability of PF scheduler. the access terminal with smaller rk [t ] is chosen. the R ’s of other users are updated slower than the dropping rate of the cell edge user. named Alpha-Rule is proposed. With dynamic control of β . Throughput and fairness of cell edge users can be hence guaranteed in a short-term basis. However. β is the same for all users in a scheduling interval. there are high probabilities that the scheduling priority of this cell-edge-user is small. β = 1 . Cell Edge Problem PF scheduler can balance the multi-user diversity and fairness among users. such as HSPA or EV-DO. the fairness index is only slightly changed. To elaborate this problem. β k [t ] is updated in a range from 0 to 1 (or 0 to 2) (not as in [7][8] from 1 to infinite). When a mobile terminal moves into the cell edge. It controls and dynamically changes β . β = 0 . then when he entered the cell edge. In [6]. When a < 0 . Our approach is different from above mentioned methods in the following ways. This method has been extended in [5]. parameter β is not adaptive with time or with users. 4. a generalized formula is proposed to computing the average rate as: Rk [t ] = (1 − 1 Tc ) Rk [t − 1] + 1 Tc ( rk [t ]) 1− a . in which α is adaptively updated on a per-user basis. this method requires an extra component to monitor the fairness and the aggregated throughput. When a > 0 . especially cell edge users in a real-time way.With a simple generalization. then it is a maximum C/I scheduling. if α = 1. e. which permits dynamic and real-time tradeoff between aggregate throughput and per-user throughput. this method reacts slowly when a user enters into the cell edge and can not provide good throughput and performance for cell edge users in a short-term way. The control parameter α is the same for all users. To solve this issue of PF. In this Alpha-Rule. In [8]. the numerator rk in the priority calculation equation (1) will become small for a long term because that the bad channel condition is caused not by the small scale fading but by serious cell edge interference and high path loss. thus balancing the aggregate throughput and long-term fairness. a Data Rate Control (DRC) Exponent Rule is proposed. since the fairness index calculation is related to all the users in the system. the fairness is changed between proportional fairness and Max-Min fairness. which requires more complexity and memory in the scheduler. let’s check the priority calculation equation (1) and the rate update equation (2). With pure PF scheduling policy. It does not treat individual user.. Meanwhile. When individual cell-edge-user’s throughput is reduced. When a=0.

we can adapt the β of all users in the system: we set α = 1 . while in the cell edge when γ k ≥ δ . we have when γ k [t ] < δ . So. ( Rk [t ]) β k [t ] < Rk [t ] ( Rk [t ]) β k [t ] will be bigger than Rk [t ] . Dynamic β for Cell Edge Users To dynamically adjust the scheduling priority in an efficient way. This time counter records the duration in which a user is in the cell edge. B. e. users in cell edge will get more opportunities to transmit their data packet. With the priority calculation equation below: j = arg max k Alternatively.. and β k [t ] as a function of γ k : 1. which is not what we want. K } . In addition to adapt the parameter of the cell edge user. γ k [t ] < δ δ (4) β k [t ] = max γ k [t ] γ [t ] . Assuming the downlink pilot averaged SINR γ k of user k can be reported to Base Station (BS) by user k. we set β k [t ] = 1 β k [t ] . or the QoS requirements. b . In our proposed method.III. this ratio (thus β ) will be lower than 1. γ k [t ] < δ δ (5) β k [t ] = max 1 2 − Because usually Rk [t ] ≥ 1 . for example. Dynamic β for All Users ( Rk [t ]) ( rk [t ])α β k [t ] . or user’s traffic parameters and Quality of Service (QoS) requirements. DYNAMIC PF SCHEDULER In order to improve the performance of users in the cell edge. We give details in the following. A user is in the cell center when γ k ≥ δ . γ k [t ] < δ .g. (6) With this counter. although when δ > 0 . γ k may be less than 0. the user’s averaged SINR (which is closely related to its location). Note that when Rk [t ] < 1 . we need a mechanism to determine whether a user is in cell edge or not. A problem arising in above design is that. In our proposal.b . b .. when a user enters cell edge. the scheduling priority is artificially adapted to individual user’s needs. depended on user’s channel condition (the averaged signal to interference and noise ratio (SINR)). So cell edge user can get higher priority than that in a pure PF scheduler. β is no longer adjusted as a function of the averaged SINR. The underlying idea in this parameter adaptation is to accelerate the increase of scheduling priority when user is in the cell edge. Usually. we propose to modify the PF scheduling algorithm in a new way. we can set β k [t ] to update less aggressively: 1. It is not the purpose of this paper to optimize the δ parameter value. when Rk [t ] < 1 . The performance and fairness is then guaranteed by standard PF itself. the β parameter can dynamically adjusted as a function of user’s averaged SINR. A. δ max 3 . Tk [t ] ≤ 0. To ensure β k [t ] > 0 . if a user remains staying the cell edge for a very long time. where b is a small positive number. or Tk [t ] > θ β k [t ] = max . The adjustment of the value of β is only performed within a predefined duration.. the value of β may be different for different users. γ k [t ] ≥ δ . in (4) and (5). γ k [t ] . γ k [t ] ≥ δ Tk [t − 1] + 1.. Before going to details of our method. With this adjustment. b . γ k [t ] . In each updating period. γ k [t ] ≥ δ . 0 ≤ Tk [t ] ≤ θ δ (7) where the threshold θ is to indicate that the user has stayed too long in the cell edge. A time counter is also introduced to maintain long-term fairness. then its scheduling priority will be increased compared to the standard PF algorithm. One way to do this is to set β as the averaged SINR to a threshold ratio when user is in cell edge and as 1 when user is in cell center. A cell edge threshold δ can be a priori determined by network planning or dynamically set by the BS and then broadcasted to all users. In the standard PF algorithm. When a user stays in cell edge for a relative long time (longer than θ ). With this consideration. it will violate the PF criterion of the whole system and reduce the system throughput. Now we explain the function of b in (4) and (5). β in the dominator is dynamically adjusted in a user depended basis. or measured from BS through uplink pilot. which can compensate the priority decreasing caused by bad averaged SINR in cell edge. k ∈ {1. Both pilot averaged SINR γ k and the cell edge threshold δ are represented in dB. (8) γ k [t ] >2 2. k ≤ 2 δ δ β k [t ] = . we set a parameter b as the lower bound of β k [t ] in (4) and (5). we can set: 1.. we can further improve the parameter design by introduce a time counter Tk [t ] to record the duration in which the user remain in the cell edge: Tk [t ] = 0. In some scenario. γ k [t ] . the numerator in (3) has an exponent denoted as α and the dominator has an exponent denoted as β .

where 2 θ3dB =70 degrees.53s. users in the cell edge can be compensated and users in the cell center can be punished.68s. In Figure 2.5.60 ( δ = 7 dB). TABLE I. The link packet error probability is predicted based on channel SINR by effective SNR mapping methodology [9]. As for Dynamic PF. γ k [t ] < δ Bk [t − 1] + 1. at the cost of a reduction in total throughput of 14.2% ( δ = 7 dB). (10) Bandwidth of sub-carrier sub-frame length Number of OFDM symbols in a sub-frame Distance-dependent path loss Shadowing fading Shadowing standard deviation Multi-path channel model User speed Total downlink TX power in traffic channels Cell radius Modulation type Channel coding HARQ type With (9) and (10). The antenna patter A (θ ) = − min 12 (θ θ3dB ) . Ak [t ] ≥ θ . According to the discussion in last subsection.2% ( δ = 5 dB) and 16. 80% users have packet delay less than 0. while the sacrifice of the system capacity for the fairness can remain comparatively small. 3/8. the users near cell center will be punished since β k tends to be larger than 1. 3/4 IR SIMULATION PARAMETERS Value (9) Number of sub-channels in traffic channels Number of sub-carriers in each sub-channel . The simulation layout has 19 base stations and each has 3 sectors. the fairness of the system is improved from 0. The time counter threshold θ is set to 500 subframes.7% ( δ = 7 dB).59 ( δ = 5 dB) or 0. We also simulate the packet delay performance. I= K rk [t ] ( r [t ] ) 2 k =1 k K . To demonstrate the effect of the Dynamic PF algorithm more clearly. The fairness index and system 4 . or.. The performance is showed by fairness index.0% ( δ = 5 dB) or 31. γ k [t ] ≤ δ throughput of two threshold settings of Dynamic PF together with those of PF are illustrated in Figure 1 and Figure 2. For Dynamic PF algorithm.4% ( δ = 7 dB). or. b . 1/2. the throughput of cell edge users is improved from Dynamic PF to standard PF by 16. SIMULATION RESULTS This section presents simulation results to validate the performance of the proposed method.49 of standard PF to 0. and Am = 20 dB. To provide more credible results. Am . Bk [t ] > 0 δ for all k ∈ {1.By equation (8). 9. IV. 1/4..7% ( δ = 7 dB).9% ( δ = 5 dB) and 13. γ k [t ] > δ 0. In Figure 3. The algorithm of dynamic β for cell edge users as in (7) is investigated. K} . 10) is improved by Dynamic PF by 30. With standard PF algorithm. we run the simulation with 25 drops and each drop runs 2 second (2000 steps).7% ( δ = 5 dB) and 9. we can see that when Dynamic PF is used. we can have: 1. 0 ≤ Ak [t ] ≤ θ . user and system throughput. Parameter System carrier frequency System bandwidth 2GHz 10MHz 50 12 15KHz 1 ms 12 COST-231 lognormal distribution 8 dB GSM TU 6 ray 3 km/h 46 dBm 1km QPSK/16QAM/64QAM Turbo coding 1/5. we can take a drop as an example and examine the impact on the cell edge users. respectively: Ak [t ] = Bk [t ] = 0. We compare the standard PF and the proposed Dynamic PF in an OFDMA based system simulation platform. The delay of all packets belonging to a user is averaged and the CDF curve of the averaged packet delay of users is showed in Figure 4. 0 ≤ Bk [t ] ≤ θ β k [t ] = max (11) γ k [t ] . From the above simulation results. at the cost of a reduction in total throughput of 10. The central sector is interested and surrounding sectors provide interference. Full buffer traffic is considered because the comparison is irrelevant to traffic feature. while the users in cell edge will be compensated since the β k tends to be smaller than 1. The main simulation parameters are listed in Table I. we introduce two time counters Ak [t ] and Bk [t ] to record the duration in which the user remain in the cell edge and the cell center. Parameter b is set to 0. There are 10 users randomly distributed in the central sector.. By (11).. 80% users have packet delay less than 0. Packet size in the transmit buffer is set constantly as 1500 bytes so that packet delay performance can be considered in simulation. we can draw the conclusion that the cell edge user’s performance can be improved by the proposed Dynamic PF algorithm. γ k [t ] ≥ δ Ak [t − 1] + 1. So the packet delay performance can also be improved by our proposed algorithm. the throughput of cell edge users (user 8. we set the cell edge threshold δ as the averaged SINR of 5 dB and 7 dB respectively. Fairness index is computed as [10]: K 2 k =1 From Figure 1.

In our algorithm. Padovani. pp. pp. Aniba and S. we propose a dynamic proportional fair scheduling algorithm to provide better performance for cell edge users compared to standard PF scheduler. May 2000.D. Wang. Sang. and B. Fairness index comparison 9000 8000 7000 System Throughput Throughput (Kbps) 6000 5000 4000 3000 2000 1000 0 1 standard PF Dynamic PF (δ = 5dB) Dynamic PF (δ = 7dB) Cell Edge Users Throughput 2 Figure 2.8 standard PF Dynamic PF (δ = 5 dB ) Dynamic PF (δ = 7 dB ) CDF 1 1. Shen. 6.2 0. CONCLUSION In this paper. R. V. and J. 1854–1858. [5] G. the β parameter is adjusted in a user-dependent. Oct. K.3 0. T. 2005. Zou Ting. R1-050629.4 0. pp. vol.2 0. M. Gitlin. 11. IEEE VTC2006 Fall. pp. Shen.K.” IEEE Transactions on Vehicular Technology.6 0. and Zhang Ping.9 0. June 2005 [3] Xu Ning. F-21 . Madihian. Aissa. Han. 2244–2248. [6] A. Indoor and Mobile Radio Communications (PIMRC’ 2002). 1 0. “An efficient scheduling algorithm for QoS in wireless packet data transmission”.5 0. Sartori. Classon.2 0. vol. 5217 – 5221. Kim. “Adaptive proportional fairness for packet scheduling in HSDPA”. Yang. Huawei. 4033–24037. “Data Throughput of CDMAHDR: a High Efficiency-High Data Rate Personal Communication Wireless System”. 12th IEEE International Symposium on Personal. Naik. “Performance characteristics of cellular systems with different link adaptation strategies. March 2006. 2004. Pankaj. IEEE Trans. vol.6 0. “Forward link packet scheduler for high-speed packet data system”. vol. [9] K. 1–5. Indoor and Mobile Radio Communications (PIMRC’2001). The 13th IEEE International Symposium on Personal.V. “Inter-cell Interference Mitigation”. User’s throughput comparison 5 . Kostas. Y. Kim. 5. IEEE Global Telecommunications Conference. Dec. [4] H. Louisiana. no. X. X. June 2006. 3. Packet delay comparison Fairness Index 0. Li. [2] 3GPP. pp. Yamaguchi and Y. 712 – 717. Takeuchi. Nov. Shao. [7] D. pp.6 0.7 0.F-24. and R.4 0. September 2006.1 0 1 REFERENCES [1] A. and S. 6. 568-577. 2003.6 Average packet delay (second) Dynamic PF (δ = 7 dB) Dynamic PF (δ = 5 dB) standard PF Figure 4. and W.1 0 0 0. 2001. System throughput comparison 600 500 User Throughput (Kbps) standard PF Dynamic PF (δ = 5 dB) Dynamic PF (δ = 7 dB) 400 300 200 100 0 1 2 3 4 5 6 7 8 9 10 User Index Figure 3. and R.3 0. Baum. P. D. “A flexible downlink scheduling scheme in cellular packet data systems”. Simulation results show that the proposed algorithm can improve the performance of cell edge users with a little degradation of the whole system throughput. Proc.4 1.7 0. Jalali. Wireless Commun. [10] M. Proceeding of IEEE Vehicular Technology Conference (VTC Sprint 2000). 5. Wang Ying. New Orleans.8 0. The main idea is to accelerate the competitiveness of cell edge users. 2 3 Figure 1. 2006 IEEE International Conference on Communications (ICC’ 2006).4 0.8 1 0. IEEE WCNC'05. [8] A. vol.2. pp. 52. time-varying way. vol.O. “A MC-GMR Scheduler for Shared Data Channel in 3GPP LTE System”. pp. Dianati.2 1. Lee.5 0. (GLOBECOM’ 04). “A New Fairness Index for Radio Resource Allocation in Wireless Networks”. pp. September 2002.9 0. The adjustment can be controlled by the relative pilot signal strength of users. “Towards Opportunistic Fair Scheduling in Wireless Networks”. vol. 1497-1507.