Radio Resource Allocation Algorithm for Relay aided Cellular OFDMA System

Transcription

1 Radio Resource Allocation Algorithm for Relay aided Cellular OFDMA System Megumi Kaneo # and Petar Popovsi Center for TeleInFrastructure (CTIF), Aalborg University Niels Jernes Vej 1, DK-90 Aalborg, Denmar # Graduate School of Informatics, Kyoto University Yoshida Honmachi Sayo u, Kyoto, , Japan {me petarp}@om.aau.d Abstract We address the problem of radio resource allocation in the Downlin (DL) of relay aided cellular system, based on OFDMA transmission technology. There has been little wor on specific resource allocation algorithms for this system in the literature, although these are the ey elements for realizing the potential capacity and coverage increase offered by the relay. Therefore, we propose two resource allocation algorithms which improve the overall throughput and coverage compared to a system without relay. The advantage of our algorithms is that they perform well while minimizing the complexity and the required amount of Channel State Information (CSI), maing them suitable for practical use. I. INTRODUCTION One of the ey expectations for the 4th Generation (4G) wireless system is to provide ubiquitous high data rate coverage. During these recent years, a large part of the research has focused on Orthogonal Frequency Division Multiple Access (OFDMA) transmission technology, a very promising candidate for the physical layer in 4G cellular system, due to its inherent robustness against frequency selective fading and its capacity for achieving high spectral efficiency. In this multiple access scheme, each subcarrier can be allocated to a different user which can best exploit the current channel condition, hence maximizing the achievable capacity [1]. But with the traditional cellular architecture, increasing the capacity along with the coverage would require the deployment of a large number of Base Stations (BS), which is very costly. However, introducing Relay Stations (RS) in each cell can alleviate this problem since the RS can forward high data rates in remote areas of the cell while eeping a low cost of infrastructure. But this potential gain in capacity and coverage is highly dependent on the radio resource allocation strategy, a topic which draws more and more attention of the research community []. Moreover, the combination of RS resource allocation with OFDMA offers an even more promising perspective as for increasing capacity and cell coverage. Some wors on resource allocation for OFDMA relay systems can be found in the literature, but they mainly focus on multihop, infrastructure less networs or else, only a single destination is considered [3] [4]. To the best of the authors nowledge, there has been little wor in the literature on specific resource allocation schemes for multiple access in a RS aided cellular system based on OFDMA technology. Thus, our goal is to provide efficient allocation algorithms for that system in the Downlin (DL), which enable a high capacity and high coverage with a low complexity. Also, one of the ey points is that the algorithms are designed in such a way that reduces the required amount of Channel State Information (CSI). The remainder of the paper is organized as follows: after defining our system model, our path selection method is presented. This is followed by the description of our algorithms which are evaluated by simulations. Finally, we conclude the paper and give directions for future wor. II. SYSTEM MODEL In this wor, we focus on the single cell DL transmissions with OFDMA, with one BS and one RS as depicted in Fig. 1. Users or Mobile Stations (MS) feed bac to the BS or to the RS their CSI containing their per subchannel Signal to Noise Ratio (SNR) every time frame. A subchannel is defined as a group of adjacent subcarriers. Adaptive Modulation and Coding (AMC) is performed on a per subchannel, per frame basis. We use the frame structure illustrated in, where there are: Time division between BS and RS transmissions, with T BS the time allocated to the BS Frame where transmissions from the BS occur and T RS is allocated to the RS Frame where transmissions from the RS occur. Although suboptimal, we assume time division since it is widely considered in based relay system [5]. Frequency division between BS MS and BS RS transmissions, where each subchannel can be allocated to either lin. With this frame, the RS cannot receive and transmit the same pacets during the same frame: frames are needed to relay pacets from the BS to the destination MS. To the best of our nowledge, this is a new but feasible type of structure as usually, the frequency regions for BS MS and for BS RS are fixed and separated. We focus on a two hop scenario. Let us note that, the goal of this wor is to design algorithms for allocating OFDMA resources in

2 the presence of a RS in a cell, when the CSI is reported. Hence, the principle of the presented algorithms remains the same when the neighboring cells are considered, perhaps with a lower performance (which is also the case for the reference algorithms, so the relative difference will be ept.) Fig. 1. Fig.. Cell Model Frame Structure III. PATH SELECTION With path selection, each MS is lined with either the direct BS MS lin or the relayed RS MS lin, in a manner that maximizes the achievable capacity of the cell. The trade off is that, users close to the RS experience a higher data rate on the relayed lin than the direct lin, but additional delay will occur since relayed pacets need two hops to be received. If we wish to mae an optimal allocation, path selection and resource allocation should be made at the same time. In that case, each user would have to feedbac his CSI to the BS, both for the direct and relayed lins and the algorithm would be highly complex since a path selected for a user would depend on the choice made for every other user. This would result in a tremendous amount of control information. As our goal is to reduce the algorithm complexity and the required information overhead, path selection is made prior to resource allocation. In [6], a low complexity, suboptimal path selection method is proposed where each user selects his path by measuring his SNR level on direct and relayed lins and selects the one with the highest SNR level. The advantage of this method is that no CSI feedbac is required from each MS to the BS. However, comparing only the SNR levels of direct and relayed lins can result in an unfair comparison as the additional delay introduced by relaying is not accounted for. That is, if the RS queue of a user is empty, it would be a waste to allocate the RS MS lin to this user and the capacity, proportional to the SNR on the RS MS lin would not be achieved. Thus, the method in [6] is reliable if there are always pacets queued at the RS, which is not always true: in many cases the RS queues are empty and pacets need to be forwarded from the BS to the RS. Here we introduce a measure for the effective data rate as opposed to the achievable data rate given by the SNR. We illustrate this observation in the following. We mae the assumption that the RS queue is empty, which is the worst case. In this case, two frames are needed to transmit pacets to a relayed user, the first frame for forwarding pacets to the RS using the BS RS lin, and the second frame for providing the pacets to destination, via the RS MS lin. This is because the RS can not receive and transmit the same pacets in one frame. In the direct case, pacets are transmitted in the BS Frame part in both frames. We assume that the frame is equally divided in time between BS Frame and RS Frame, i.e., T BS = T RS = T F. For a user, we compare his effective capacity in the cases of direct and relayed transmissions, as illustrated in Fig. 3, where the achievable capacities for each lin are indicated for that user, namely r BS MS, r RS MS and r BS RS (same for all users). If user is relayed, in the first frame, his pacets are transmitted to the RS, but the RS Frame doesn t contain any pacet for user, so his effective data rate for this frame is If user is put on the direct lin, ρ BS MS ρ RS MS (1) = 0. (1) (1) = rbs MS T F / T F = rbs MS. () But in the second frame, RS Frame contains pacets destined to user since his RS queue is no longer empty. Therefore, we have in the relayed case ρ RS MS and in the direct case () = rrs MS T F / T F ρ BS MS () = rbs MS = rrs MS, (3). (4) To obtain the overall effective rates for user, we tae the average over the two frames, which results in the direct case and in the relayed case ρ BS MS ρ RS MS = rbs MS, (5) = rrs MS. (6) 4 Therefore, if the RS queue is empty for a user, we need to compare r BS MS for the direct lin with rrs MS for the relayed lin, for a fair comparison. To summarize, the procedure of our path selection is that, each user determines his channel qualities r BS MS for the direct lin and r RS MS on the relayed lin. If r RS MS /, the user selects the relayed lin, otherwise the direct r BS MS

3 one. Once a path is selected for each user, the selected lins are ept until the long term average SNR of the users change significantly. Then, path reselection is performed periodically by assessing the long-term channel statistics. Compared to [6] which can also be used, but assumes that the RS queue is full, our proposed method is a worst case approach since we assume that the RS queue is empty. However, it ensures that the throughput achieved by relayed users will not be penalized by the fact that two frames are required to forward their pacets. This is because for the users satisfying r BS MS r RS MS r BS MS, a higher throughput can be achieved by being scheduled with the lower, direct lin s rate in every frame, than with a higher, relayed lin s rate but only every two frames. Fig. 3. Path Selection IV. RESOURCE ALLOCATION ALGORITHMS A. Algorithm with Fixed Time Division Once each MS is lined to a path, the BS performs the Resource Allocation (RA) algorithm presented in this section. The aim of this algorithm is to maximize the overall cell throughput, while minimizing the complexity and the amount of CSI. The first algorithm proposed, referred as RS Max with fixed time division, is a RS aided centralized allocation, where there is an equal time division between T BS and T RS. This is RS aided in the sense that the RS performs its own subchannel allocation in RS Frame and maes requests to the BS for some relayed users pacets to be sent to the RS. The main idea behind this algorithm is that, if the RS allocates the best CSI user in each subchannel, the allocated subchannel may be wasted if no pacets are queued at the RS for that user. Instead, the RS allocates a user with a lower CSI but who has queued pacets, so that the allocated resource is effectively utilized. For the users who had a higher CSI but no pacets, there is a high probability that they will be also scheduled in the next frame since it is reasonable to assume that the RS MS lins are stable over at least several frames. Therefore, the RS sends a request to the BS so that with the current frame, the pacets arrive to the RS and these can be forwarded to the relayed users in the following frame. Also, to increase throughput, real channel utilization is considered, unlie in the usual algorithms such as Max C/I or Proportional Fair Scheduling (PFS). For example, the subchannel allocation in RS Frame is made by taing into account both the user s AMC level and queued pacets. Hence, the throughput is optimized since this algorithm allocates the resource to the lins with high rates and which can use the resource most efficiently. The algorithm is described in detail below: 1) Allocation of RS Frame by RS a) In each subchannel, relayed users are sorted in the order of best AMC level. The user with highest AMC level and with pacets queued at RS is allocated each subchannel. b) The set of users having a higher AMC level than the allocated user but without pacets queued at RS is denoted U Req. For these users, the RS requests the BS to forward their pacets in the BS Frame part of the current frame. c) The request information about U Req is sent to the BS. ) Allocation of BS-Frame by BS, based on the requests by RS a) BS allocates temporarily each subchannel to the best direct user. b) Then, if the RS requested pacets to the BS, the number of subchannels n BR required to send those pacets is determined. The crucial assumption here is that, we assume that all the BS RS subchannels have the same average SNR level, since the BS RS lin is in Line of Sight (LOS). Thus, any subchannel among all N subchannels can be chosen to support the BS RS transmission. c) In each subchannel, we compare the achievable rates for the best direct user and the BS RS lin, i.e., AMC level r BS MS and the average BS RS AMC level r BS RS. The subchannel is allocated temporarily to the lin with the best rate. This results in y subchannels allocated to the BS RS lin, which are not all required. Thus, we compare y with n BR, the required number of subchannels to accommodate the requested pacets: i) If y < n BR, the y subchannels are not sufficient to accommodate all the pacets so some will remain at the BS queue. We allocate pacets from randomly chosen relayed users, until all y subchannels are filled with their pacets. ii) If y > n BR, all y subchannels are not needed for the BS RS lin since the number of queued pacets is lower than the offered capacity, so the y n BR subchannels in excess can be reallocated to the direct users. The n BR worst subchannels for direct users, i.e., with the lowest AMC levels r BS MS, are allocated to BS RS lin. The remaining y n BR subchannels are reallocated to the best direct users. In the throughput, we tae into account the real channel

4 utilization as shown below. Throughput is calculated as τ TF / = τ BS MS T F / + τ RS MS T F / (7) T F where τ BS MS is the throughput achieved in BS Frame for direct users and τ RS MS is the one achieved in the RS Frame for relayed users. For each lin l, they are defined as τ l = 1 N K c l,n u l,n r,n l (1 ber,n l r l N,n). (8) n=1 =1 In this expression, c l,n is equal to one if user is allocated subchannel n on lin l, zero otherwise, r,n l is the achievable data rate or AMC level for user on subchannel n, lin l. It is expressed in [b/s/hz] but also corresponds to the number of pacets that can be fit in a bin, where a bin is defined as a time slot and a subchannel, and the pacet size is fixed to S BU = P N T s bits. P is the total number of subcarriers, is the slot length and T s is the OFDM symbol duration, in [s]. For example, if we have BPSK, r,n l = 1 and one pacet of size S BU is contained per bin. If r,n l =, there would be pacets per bin. ber,n l is the bit error rate experienced by user on subchannel n, lin l. u l,n is an utility metric which measures the real channel utilization, equal to T l,n u l,n = min(rl,n T l,n, q l ) (9) r,n l T,n l Here, is the number of time slots allocated to user, subchannel n, lin l where T,n l is expressed in [s]. q l is the number of queued pacets for user lin l. u l,n is simply the number of allocated pacets over the available capacity for user, subchannel n, lin l, where r,n l T l,n is the available capacity counted in number of pacets of size S BU. If the number of queued pacets is larger than the available capacity, the capacity is fully utilized and u l,n = 1, otherwise all the queued pacets are allocated and u l,n < 1 since ql < rl,n T l,n. B. Algorithm with Adaptive Time Division The previous algorithm was designed with a fixed time division between BS Frame and RS Frame. In this algorithm, referred as RS Max with adaptive time division, after the allocation made by the previous algorithm, the time division is adapted in an iterative manner. 1) To search the direction for the optimal time division, first, T BS is increased by one slot and T RS decreased by one, i.e., T BS = T F /+ and T RS = T F /. This gives a new throughput value, τ a where the utility metrics have been updated by (9). ) In the same way, T BS is decreased by one slot and T RS increased by one, which gives throughput τ b. 3) Then, we compare τ TF /, τ a and τ b, and eep the time division corresponding to the maximum throughput. a) If τ TF / is the maximum, then the algorithm stops and τ Opt = τ TF /. b) Otherwise, if for example the maximum throughput is τ a, we continue the time division by setting T BS = T F / + and T RS = T F /. The new throughput τ a+1 is computed and compared with τ a. If τ a > τ a+1, the optimal solution is τ a, otherwise we continue the adaptation. At iteration i, T BS = T F / + i and T RS = T F / i with throughput τ a+i. The search stops when τ a+i 1 > τ a+i and the optimal throughput is τ a+i 1. c) Finally, T BS and T RS are fixed to their optimal value and the queues are updated following this optimal time division. C. Upper Bound Algorithm To evaluate the performance of the proposed algorithms, we have designed an algorithm giving an upper bound to the optimal throughput. Due to the constraints imposed by the frame (Fig. ), the optimal algorithm becomes too computationally complex since the allocation of the paths, subchannels, and time divisions should be jointly performed depending on the real channel utilization in each case, while satisfying the constraints of the frame. In this Upper Bound algorithm, the allocation is also completely BS centralized and the path selection and subchannel/time allocation are merged and performed for every frame. However, we tae the following simplifying assumptions which also ensure that this algorithm will give an upper bound: all the pacets destined to a relayed user coming from the BS to the RS in a frame is received by the user in the same frame the optimal time division between BS Frame and RS Frame can vary for each subchannel in a frame. These assumptions are not feasible with the frame structure in Fig., since in this case pacets from the BS for relayed users would need at least two frames to be received. Also, a different time division per subchannel is not feasible since we assume a TDD system, so a RS cannot transmit and receive at the same time on different subchannels. Hence this algorithm gives a performance upper bound, but infeasible. For this algorithm, the frame will be structured as depicted in Fig. 4. Here is the description for this Upper Bound algorithm. 1) Let K be the total number of users. We create set D containing the CSI of all K users on the direct lin, and set R, their CSI on their relayed lin. ) For all users in R, we determine the values of T,n BS allocated to the BS RS lin and T,n RS for RS MS lin, for user on subchannel n. With the assumption that everything sent from BS to RS arrives at MS during the same frame, T,n BS RS and T,n are proportional to the

5 BS RS and RS MS rates for that subchannel. Simply T RS,n = r BS RS r BS RS + r RS MS,n T F, (10) and T,n BS = T F T,n RS. 3) To tae into account the real channel utilization, we define the effective capacity, η,n l = ul,n rl,n T,n l, which measures the useful part of the allocated capacity. Then, the K users from sets D and R are simultaneously ordered by decreasing η,n l. In each subchannel, the best user who has either the direct or relayed lin, is allocated. 4) Finally, the queues are updated following this optimal allocation. Fig. 4. Frame Structure for the Upper Bound Algorithm V. REDUCED CSI FOR THE PROPOSED ALGORITHMS One common concern for RS aided systems is the increased amount of CSI. An optimal algorithm centralized at the BS or the Upper Bound algorithm would require: 1) CSI for the direct lin of all K users per subchannel per frame ) CSI for relayed lin of all K users per subchannel per frame, which can result in a tremendous amount of overhead. For RS Max with fixed time division, the required feedbac is: 1) CSI of direct lin users K D K per subchannel per frame, required at the BS ) CSI of relayed users K R K per subchannel per frame, required at the RS 3) user IDs of users in U req, sent from RS to BS Since K D +K R = K, the amount of 1)+) for RS Max with fixed time division is equivalent to the amount of only 1) (or )) for Upper Bound. It is also equal to the amount of CSI required for the usual Max C/I algorithm without relay. Since the number of users in U req is usually small, the amount of 3 will be reasonably small. Thus, the amount of CSI required by RS Max with fixed time division is much lower than required for Upper Bound and slightly higher than for Max C/I. For RS Max with adaptive time division, in addition, the CSI of the K R,alloc allocated relayed users is required at the BS, since the throughput values are needed during time adaptation. But it is still much lower than needed for the Upper Bound, since K R,alloc K R K. Thans to the mechanism of the RS performing its own allocation, the amount of feedbac can be ept much lower than what is usually required. The cost of this overhead is evaluated later in the simulations by the goodput γ in [b/s/hz], defined as n data γ = τ (11) n data + n OH where τ is the cell throughput, n data the number of OFDM symbols in the frame carrying data and n OH the number of symbols carrying the CSI. VI. SIMULATION RESULTS For the evaluation, simulations were made over 0000 sets of channel realizations, each set of channels consisting of independent user channels. The simulations are made in a single cell with a radius of 1000m, with one BS and one RS as depicted in Fig. 1, the RS being placed 800m away from the BS. Users are uniformly distributed. The frame duration is equal to 10ms. Path loss model for urban areas, log-normal shadowing from [7], multipath Rayleigh fading channels and exponential power delay profile from [8] were used. There are 3 subcarriers and 8 subchannels, each composed of 8 contiguous subcarriers, and 5 to 0 users. Power is equally distributed in each subcarrier, with a maximum BS transmit power of 0Watts and 5Watts for the RS. User pacets arrive at the BS queue following a Poisson process. With our path selection algorithm, each user is first attached either to the BS or the RS, next, the time/subchannel allocation is made. The path selection is renewed each time the average user SNRs change. The proposed algorithms are compared with Max C/I algorithm without relay and the Upper Bound algorithm, which gives the performance upper bound. In Fig. 5, the goodput performance is plotted for each algorithm. It is shown that our proposed algorithms outperform the performance of Max C/I algorithm without relaying referred as BS Max, up to 15% for 0 users, even with the additional amount of signaling overhead. Compared to Upper Bound, our algorithms have only around 10% lower goodput, and the gap with the optimal solution would be smaller. RS Max with adaptive time division achieves a better performance than with fixed time division, but they are fairly close. This can be explained by the probability distribution of T BS for RS Max with adaptive time division shown in Fig. 6 where in 65% of the cases, the equal time division gives the best goodput. Thus, it is a good strategy to divide the frame equally since a near optimal performance can be achieved with less overhead and lower complexity. Moreover, the coverage performance was evaluated. Since we consider the coverage after scheduling, we define the system outage probability P out as the probability that the allocated user rates r are lower than a reference rate R, where r is averaged over 10 frames S s=1 P out = K s K S. (1) K s is the number of users in outage for the sample s; S is the total number of samples, equal to the number of iterations

6 divided by 10 since the samples are taen every 10 frames. Fig. 7, where the number of users is fixed to 0, shows that our algorithms outperform the coverage of BS Max, which means that with our algorithms, a higher number of users can support a rate R than in the case without relay. Hence, our algorithms can increase the coverage by increasing the number of supported users. For example, at R = 1 Mbps, the outage probability is reduced by 10% compared to BS-Max, allowing 10% more users to attain a rate higher than 1 Mbps. RS Max with adaptive time division achieves a slightly better coverage than the fixed algorithm. Goodput [b/s/hz] UB BS Max RS Max fixed RS Max adapt Number of users Fig. 5. Cell goodput performance Outage Probability BS Max RS Max fixed 0. RS Max adapt UB Rates [Mbps] Fig. 7. System outage performance results showed that our algorithms performed well compared to the upper bound algorithm, with a much lower complexity and required CSI. Even with an increased overhead, our algorithms outperformed the Max C/I algorithm without relay for both goodput and outage. This was not obvious: without an appropriate algorithm design, the gain from the relay could be erased by the loss due to the increased overhead. In a relayed system, we are mostly interested in reducing the outage probability, which was achieved by our algorithms while increasing the overall goodput. As a future wor, by combining the proposed algorithms with Proportional Fair Scheduling, the system outage is expected to be further minimized RS Max adapt ACKNOWLEDGMENT We would lie to than Samsung Electronics for supporting this wor. Probability of occurrence Length of optimal T in [ms] BS Fig. 6. Probability of occurence of different values of T BS VII. CONCLUSION In this wor, we have proposed practical radio resource algorithms for RS aided cellular system based on OFDMA technology. After a simple path selection procedure, subchannels are allocated for a fixed time division between BS and RS transmissions, followed by the time adaptation. Simulation REFERENCES [1] R. Knopp and P. Humblet, Information capacity and power control in single cell multiuser communications, in IEEE ICC, Seattle, WA, June [] R. Pabst et al., Relay-Based Deployment Concepts for Wireless and Mobile Broadband Radio, IEEE Wireless Comm. Mag., pp , September 004. [3] G. Li and H. Liu, Resource Allocation for OFDMA Relay Networs, in Thirty-eighth Asilomar Conference on Signals, Systems & Computers, Pacific Grove, CA, November 004. [4] M. Herdin, A Chun Based OFDM Amplify-and-Forward Relaying Scheme for 4G Mobile Radio Systems, in IEEE ICC, Turey, June 006. [5] F.-C. R. et al, Recommendation on PMP Mode Compatible TDD Frame Structure, IEEEC80.16MMR-05 07r1, Nov 005. [6] H. Hu and H. Liu, Range Extension without Capacity Penalty in Cellular Networs with Digital Fixed Relays, in IEEE GLOBECOM, Dallas, Texas, December 004. [7] I-K. Fu et al, Reverse Lin Performance of Relay-based Cellular Systems in Manhattan-lie Scenario, IEEEC80.16MMR r1, Jan 006. [8] S. Yoon, et al., Orthogonal frequency division multiple access with an aggregated sub channel structure and statistical channel quality measurements, in Proc. IEEE VTC, vol., Los Angeles, CA, September 004, pp