On the Impact of Physical Layer Awareness on Scheduling and Resource Allocation in Broadband Multi-cellular IEEE Systems

Transcription

1 Submitted to IEEE Wireless Communications On the Impact of Physical Layer Awareness on Scheduling and Resource Allocation in Broadband Multi-cellular IEEE Systems Leonardo Badia, Andrea Baiocchi, Simone Merlin, Silvano Pupolin, Alfredo Todini, Michele Zorzi 1 Abstract Multi-cellular networks based on the IEEE standard appear as very promising systems to provide endusers with broadband wireless access. However, they also pose interesting challenges in terms of radio resource management, where several decisions are left intentionally open by the standards to the implementors. For this reason, we focus in this paper on scheduling and resource allocation, and we investigate how they could be operated in a cross-layer fashion. In particular, we describe the principles of joint scheduling and resource allocation for IEEE operating in AMC mode, and we discuss the critical role played by physical layer considerations, in particular the inter-cell interference estimation and channel-state awareness, on the obtained performance. This leads to identifying key open issues and possible general solutions. I. INTRODUCTION Among the emerging technologies to provide broadband wireless access (BWA), IEEE represents one of the most promising and attractive systems, but also presents very challenging aspects. For example, even though IEEE does not specifically aim at structures with hierarchical tiers, one of the expected areas of application is to provide wireless network coverage where deployment of cables would be too expensive or impractical. For this reason, suggested applications of IEEE include both the creation of a backbone of fixed nodes organized in a wireless mesh network, which in turn offer connectivity to mobile end users, and the implementation of the linkage between mobile terminals and fixed access points, referred in the following to as base stations (BSs), in a multi-cellular setting. In this work, we focus on the latter scenario, however, the conclusions we draw in such a case also provide useful insights on the former one (as well as on the case in which the two scenarios coexist.) The IEEE air interface standard [1] describes in detail the physical (PHY) layer, based on Orthogonal Frequency Division Modulation (OFDM). In particular, we focus on OFDMA with Adaptive Modulation and Coding (AMC) mode of IEEE , which appears to be the most suitable for such a case. L. Badia is with IMT Lucca Institute for Advanced Studies, via San Micheletto 3, Lucca, Italy. A. Baiocchi and A. Todini are with INFOCOM Dept., University of Rome La Sapienza, via Eudossiana 18, Roma, Italy. S. Merlin, S. Pupolin and M. Zorzi are with DEI, University of Padova, via Gradenigo 6B, Padova, Italy.

2 2 This method uses adjacent subcarriers to realize subchannels, resulting in a hybrid FDMA/TDMA medium access scheme. When used with fast feedback channels it can assign a modulation and coding combination per subchannel, enabling water-pouring types of algorithms, and it can be used effectively with an Adaptive Antenna System (AAS) option. Apart from the standard technical specification, in such a scenario several issues about scheduling and resource allocation algorithms are intentionally left open to developers, which stimulates researchers to seek strategies capable of providing good performance in this sense. In this paper we discuss the challenges represented by packet scheduling and resource management, and we explore the realization of a joint scheduler/resource allocator [2]; we avoid investigating the optimization issues, focusing instead on the possible choices for effective and simple implementation. The first contribution of this paper is therefore the outline of a modular scheme where a credit-based scheduler is integrated with an efficient resource allocator based on a low-complexity power-efficient and capacity-driven heuristic criterion. Although these algorithm components are not new themselves, our original contribution is their integration in a modular framework, which represent one of the outcomes of the research project PRIMO [3]. Another contribution of our research is the analysis of interference issues arising in such a scenario. In fact, even though there are studies proposing similar approaches, where packet scheduling and resource management are jointly addressed to have efficient solutions (see, e.g., [4]) for most of them the analysis and the performance evaluation are conducted in a simplified single-cell scenario. In these studies, the impact of other interfering cells is neglected, whereas in the system under examination full frequency reuse is envisioned and therefore the allocation of packets may be greatly affected by the interference conditions in the assigned resource. In this context, we will present and discuss the outcome of simulations obtained with realistic models for the details of OFDM and the physical propagation scenario, and explicitly considering multi-cell interference. As will be clear from the results shown in the following, the same allocation strategy performed regardless of the channel state and of the interference conditions caused by the surrounding IEEE BSs achieves very poor performance compared to the case where the channel state for each link is estimated and taken into account in the resource allocator. We believe that this fact requires a careful design of the system management in a channel-aware manner, in order to properly allocate the resource and to obtain an efficient implementation of IEEE based BWA. For ease of description, in the following we first describe the considered system and subsequently the issues arising from the proposed scheduling and resource management strategy. We finally show numerical results, supporting our general conclusions on the relevance of channel state and interference awareness.

3 3 subchannels matrix subchannels matrix subchannels matrix Neighboring BS Neighboring BS MT Home BWA Serving BS subchannels matrix subchannels matrix Neighboring BS Neighboring BS Fig. 1. System scenario II. SYSTEM FRAMEWORK The reference system framework is described in the following. We consider the forward link of a multicellular system as in Fig. 1, where a complete reuse of the available time-frequency resources among neighboring cells is assumed. This is one of the most challenging aspects, and requires an accurate investigation when performing resource allocation. The resource access in each cell is organized as a hybrid OFDMA-TDMA which corresponds to an IEEE system in the OFDMA-AMC mode. Bandwidth is divided into N s subcarriers and a frame consists of N t subsequent OFDM symbols. In order to reduce the resource addressing space, channel coherence in frequency and time is exploited by grouping adjacent subcarrier, time slot pairs to form a logical subchannel, which is the minimum allocable unit of resource. The addressing space is thus reduced to M s subchannels in frequency and M t slots in time. Each subchannel can be assigned to a different user and can be independently bit and power loaded. In order to make use of this flexibility in the resource management, channel and interference measurements need to be exploited by scheduling and allocation algorithms. We assume that on each frame the BS is able to obtain, through a separate error free channel, the channel status and interference value for each subchannel of each user, as measured in the past frame.

4 4 Traffic flows Scheduler Radio Resource Allocator IEEE OFDMA AMC Fig. 2. Scheduling/allocation framework Such information can be used to estimate the channel and interference conditions in the upcoming frame. The system is loaded with best effort traffic modeled by a constant bit rate source associated to each user. At the beginning of each frame the BS collects the physical layer and traffic information which is passed to the resource management block. In a cross layer perspective, resource management could be pursued by a joint optimization of all the free allocation variables. This would clearly lead to a complex algorithm design, merging physical layer optimization goals with traffic level requirements and involving a large number of variables and parameters. Thus, we split the resource management algorithm into a scheduling part, which deals with traffic related requirements, and a physical resource allocation part, which deals with physical efficiency issues, as shown in Fig. 2. These two layers work together based on a bidirectional information exchange which allows a tight interaction between them. Nonetheless, the layered structure leads to modularity and ease of design in the algorithm development. III. SCHEDULING FRAMEWORK In general, scheduling algorithms for wireless networks are derived from few basic schemes, originally designed for wireline networks, by taking into account the additional feature of a strongly time-varying channel [5]. As in wireless networks one main goal of the scheduler is to obtain high system throughput, while at the same time causing low power consumption, this would result in very unfair allocations, where few users are repeatedly allocated most of the bandwidth, while the rest of them starve. Thus, one additional goal should be to provide fair bandwidth allocation. However, for wireless networks, it is unrealistic to pursue short-term fairness, due to the specific characteristics of the radio medium.

5 5 In fact, if the scheduler aimed at achieving fairness among flows in the short term, the performance of the system would be far from optimal, as we would schedule users experiencing a bad channel state, without any benefits for either their own flow or for the aggregate network throughput. It has long been recognized that a better policy is to allow for some short-term unfairness in order to improve efficiency. The scheduler keeps track of how much each flow is allowed to transmit, and compensates lagging flows when their channel conditions improve, or when they have been starving for too long; thus the goal remains that of attaining fairness, but over a longer time scale. To implement this, the scheduler needs to take into account both channel and traffic state information; hence, we assume that the BS is aware of the queue states of all of its users, and that a sufficiently accurate estimate of the channel state is available, which is easy to achieve if the channel measurement feedback delay is small compared to the rate of variation of the radio channel. Once this information is available, to realize a proper scheduler one should keep into account the channel quality, and some mechanism for keeping track of the service state of each flow must be present. Algorithms based on credits (see, for example, WCFQ [6]) are simple and computationally efficient: the service state of each flow is summed up by a single number, its credit value. A flow gains credits when it is not scheduled, and uses credits when it is scheduled. Channel quality is dealt with in different ways, for example by introducing a cost that depends on the state of the channel; thus the scheduling priority of a flow will depend both on the amount of credits it has cumulated and on its channel quality. A flow experiencing a bad channel is at first prevented from transmitting; it is scheduled again when either its channel quality has improved or it has cumulated enough credits so as to overcome its bad channel quality index. Packet schedulers described in the literature have traditionally been designed for TDMA systems, where the goal of the scheduler is to select one flow at a time for transmission. Thus such solutions did not tackle the problem of simultaneously scheduling packets belonging to different flows and allocating a pool of transmission resources among them. This is the scenario we have to deal with in We propose a modular scheduler-allocator architecture operating with a cross-layered approach, where the packet scheduler passes some packets to the Radio Resource Allocator (RRA), which in turn determines how to transmit them. A suitable scheduler is obtained by letting a simple but efficient credit-based counter be put in feedback loop with the underlying RRA. The additional physical layer information used by our scheduler implies that it can be seen as an opportunistic scheduling mechanism [7]. The interaction between the scheduler and the RRA takes place in two steps: request selection and resource allocation. Both steps require the cooperation of the scheduler and the allocator. First of all, a parameter C req is evaluated, which represents the desired aggregated load of the cell, computed according

6 6 to an estimate of the frame capacity; in order to control the inter-cell interference, C req should be less than the maximum available capacity. In the selection step, the flows to be scheduled in the frame are selected, together with the fraction of transmission resources (number of subchannels) to be allocated to each flow. The selection is made starting from a list of requests produced by a simple credit-based packet scheduling algorithm, taking into account the average throughput achieved by each flow in the past. This list is passed to the resource allocator, which determines which requests should be satisfied, by giving priority to the flows currently experiencing favorable channel conditions relative to their average path gain. Thus, in this step we aim at reducing power consumption (as well as generated interference) by exploiting time diversity. The scheduler parameter C max determines the aggregate size of the requests in the list; thus, a higher value of C req results in a longer request list, provided that each flow has enough backlogged traffic. In general, C max C req ; when C max = C req then the allocator must necessarily select all the requests. Thus the greater the value of C max the greater the degree of freedom left to the allocator in selecting the flows to be scheduled. Once the requests to be allocated have been selected, the resource allocation algorithm determines which subchannel should be assigned to each flow. In this second step, we aim at exploiting frequency diversity, which is one of the features of the OFDMA radio interface. It is important to point out that the way the selection and the allocation of the requests are performed by the allocator has an impact on the performance of the system only in terms of power consumption and achieved throughput; it has no impact on long-term fairness among flows, which is guaranteed by the credit-based scheduling algorithm employed to create the request list. IV. RESOURCE ALLOCATION ISSUES The resource allocation mechanism deals with physical layer issues and its goal is to pursue efficiency in the resource usage. Its behavior depends on the scheduling decision, since the two operations are strictly related. RRA has to jointly set the following allocation variables: subset of data to be transmitted, selected from the request list provided by the scheduler; which subchannel subcarriers, timeslot to allocate to the selected data; which power to use on the selected subchannel; which bitload to use on the selected subchannel. Different users can in fact experience different channel conditions at a given time instant and frequency band. This property is usually referred to as multiuser diversity [8].

7 7 Once the user has been selected, subchannels have to be allocated for the transmission. Depending on the channel model, each user can experience different channel attenuation on each subchannel, which in our case is represented by a subcarrier, time slot couple. This degree of freedom is usually referred to as frequency and time diversity. In general, for any user to allocate, the selection of the channel, the power level (i.e., the Power Control mechanism) and the bitload to assign are strictly interrelated. The AMC option available in the OFDMA mode of the IEEE standard offers high flexibility and efficiency in this sense. The optimum solution of scheduling/allocation problems would clearly require a joint selection of all the optimum parameter values so that the problem turns out to be very challenging. Several optimization algorithms have been proposed to solve this problem, mainly based on heuristic approaches in order to obtain a reasonable computational complexity [9], [10]. It should be noted that in a general multi-cellular system the optimization of all the allocation variables is a network wide operation. An optimal solution in general requires complete network knowledge and a centralized allocation algorithm, and is therefore not feasible from the practical point of view, so that distributed optimization problems have been proposed [11]. In the proposed framework, each cell performs its own resource allocation without explicit control information exchange with neighboring cells. The only used information refers to the channel and interference value measurements and to the traffic request. As described in Section III the scheduling algorithm determines the aggregated throughput to be loaded on the cell and passes to the RRA a tentative list of data requests to be scheduled. The length C max of this list determines the degree of freedom of the RRA in selecting the subset of requests to be transmitted, according to its optimization goals. The aim of the proposed optimization is to maximize the aggregated capacity on each frame by using a low complexity algorithm. Here we briefly describe the used allocation heuristic with the aim of giving an insight on the problem and draw some interesting conclusions on the joint scheduling/allocation algorithms. Let k K indicate the user and s, t the subchannel. Let c k,s,t be the Shannon capacity referred to the specified user and subchannel. At the first step, a maximum power level for each subchannel is fixed, so that the maximum Shannon capacity c k,s,t corresponding to each subchannel for each user can be computed by exploiting the channel and interference information. Each user is associated a metric stating its goodness in terms of available capacity. On each step the request with the highest available capacity among the ones which have not yet reached the minimum rate is selected for a subchannel assignment. Once the request to be served has been chosen, an efficiency

8 8 metric ɛ is computed for each subchannel as: ɛ k,s,t = c k,s,t i K c. i,s,t This index allows us to compare the advantage of allocating subchannel s, t to user k, rather than to any other request. The subchannel with the highest ɛ is associated to the request selected in the previous step. The bitload is set to the highest value below c k,s,t among the available choices of coding-modulation formats and the power is adjusted accordingly. Once all available subchannels have been allocated, if not all the request have been satisfied, the procedure is repeated with a higher value of the maximum per subchannel power. This simple heuristic acts in a greedy way by decoupling the assignment of each allocation variable. Even though, for complexity reasons, we do not search for an optimal solution, the overall framework is very simple to implement, while achieving as shown by the numerical evaluations reported in the next section. V. NUMERICAL RESULTS A first consideration on the behavior of the joint scheduler/rra refers to its ability to exploit the diversity naturally present in the system. In order to test whether the proposed framework is capable of taking advantage of the system multiuser diversity, in Fig. 3 results for a scenario with a variable number of users are reported. A constant global amount of traffic requests, independent of the number of users, has been passed to the allocator. In order to give higher flexibility to the RRA, the list of requests has been filled up to an aggregate size that is twice the target throughput, i.e., C max = 2C req. Fig. 3 shows that the average transmission power decreases as the number of users increases. The proposed algorithm selects the flows to be scheduled by taking into account their channel state, thus exploiting multi-user diversity. As a consequence, the greater the number of users, the greater the chance of being able to schedule a subset of users who are all experiencing good channel conditions, therefore requiring a low transmission power. If we only pursued efficiency, with no regard for fairness, the power reduction with a large number of users would be even more significant: only the best users would be allowed to transmit, while the others would be permanently starved. It is remarkable that our algorithm manages to achieve fairness among flows in the medium term, while succeeding in exploiting diversity with an intelligent time scheduling policy. The effectiveness of exploiting user diversity is confirmed by comparing these results with those obtained using a channel-unaware allocation algorithm, which assigns subchannels without taking the channel state into account, so that no power reduction is achieved as the number of users increases.

9 dbw Channel and interference unaware Channel and interference aware Average number of users per cell Fig. 3. Average cell power consumption 80 Power (W) C max = 3 C req Channel unaware Channel aware C req Fig. 4. Mean transmission power vs Cell load

10 10 In the next set of results, we vary the cell load C req in a 9-cell scenario, with 60 randomly distributed users. The traffic sources are assumed to be in saturation, i.e., each terminal always has packets to transmit. We have set C max = 3C req. Frame duration is 10 ms. In Fig. 4 we plot the transmission power vs. the normalized cell load. It is interesting to compare the behavior of the proposed allocator with that of the channel-unaware allocator. With the former the transmission power linearly increases with the load, implying that the mean power per allocated resource unit remains almost constant; the latter shows a nonlinear power curve, i.e. the power per resource unit grows with the load. The role played by inter-cell interference accounts for this non-linear relation; as the cell load grows, so does the resource occupancy in the nearby cells and therefore the interference. The linear growth observed with our allocator is an indication that an intelligent allocation policy, which also takes interference into account, is able to manage the intercell interference by preferentially allocating less interfered resources. Fig. 5 shows the mean transmission power obtained with different values of C max (normalized to C req.) As expected the power decreases as C max increases. With C max = 1 the allocator can only select which resources to allocate to each flow, but is forced to schedule all the requests selected by the scheduler; as C max increases, it has a higher degree of freedom and can also exploit time diversity. Fig. 6 shows Jain s fairness index [12] for the throughput, computed on a window of W frames, versus the window size W. Again, results obtained by both channel and interference aware and unaware policies are compared. The lower dashed line in Fig. 6 represents the analytical lower bound on the achieved fairness, which is computed without taking into account the allocator policy, and is therefore valid for any resource allocation algorithm. The results obtained by our proposed approach are of course between the lower bound and the channel-unaware scheme, which assigns subchannels to the users with the same probability, regardless of their channel quality. In general, our proposed strategy can indeed be seen as a mechanism to trade lower fairness for lower power expenditure. However, note also that the performance of the proposed strategy is very good if the long-term fairness is considered. We also note that the Jain s fairness index is already around 0.9 by considering a 10 frame window, i.e., 0.1 s, and above 0.95 with 20 frames. This means that the allocation is likely perceived as fair at a subjective level after a relatively short time. VI. CONCLUSIONS In this paper, we have discussed resource management aspects in a multicellular IEEE like network, where challenging optimization issues arise due to the complete resource reuse among neighboring cells. In this context the OFDMA-AMC mode provided by the standard offers a flexible way to manage physical resources allowing for the development of efficient algorithms.

11 Power (W) C /C max req Fig. 5. Average transmission power vs C max, normalized to C req. Jain s fairness index Channel unaware Channel aware Lower bound Window size (frames) Fig. 6. Fairness achieved after each frame

12 12 A framework for a completely distributed and dynamic resource management has been presented which merges traffic and physical level issues in a cross layer perspective. Physical layer information, such as channel and interference measurements, together with traffic information are exploited by the proposed two layer algorithm in order to optimize some performance metrics. In particular, simulations performed in a multicellular scenario have shown that the proposed joint scheduling and resource allocation algorithm is able to trade-off fairness requirements imposed at the flow level with physical efficiency metrics such as power consumption, and to exploit the diversity naturally present in the system, thus confirming that AMC based resource management is an effective and promising approach for future generation wireless systems. REFERENCES [1] IEEE , IEEE standard for local and metropolitan area networks part 16: Air interface for fixed broadband wireless access systems, [2] M. Kodialam and T. Nandagopal, Characterizing achievable rates in multi-hop wireless networks: the joint routing and scheduling problem, in Proc. ACM Mobicom, San Diego, CA, 2003, pp [3] PRIMO project FIRB, available at: [4] T. Kwon, H. Lee, S. Choi, J. Kim, and D.-H. Cho, Design and implementation of a simulator based on a cross-layer protocol between MAC and PHY layers in a WiBro compatible IEEE e OFDMA system, vol. 43, no. 12, pp , Dec [5] H. Fattah and C. Leung, An overview of scheduling algorithms in wireless multimedia networks, vol. 9, no. 5, pp , Oct [6] Y. Liu, S. Gruhl, and E. W. Knightly, WCFQ: an opportunistic wireless scheduler with statistical fairness bounds, vol. 2, no. 5, pp , Sept [7] G. Song and Y. G. Li, Utility-based resource allocation and scheduling in OFDM-based wireless broadband networks, vol. 43, no. 12, pp , [8] P. Viswanath, D. N. C. Tse, and R. Laroia, Opportunistic beamforming using dumb antennas, vol. 48, no. 6, pp , June [9] Y. J. Zhang and K. Ben Letaief, Multiuser adaptive subcarrier-and-bit allocation with adaptive cell selection for OFDM systems, vol. 3, no. 5, pp , Sept [10] J. Cai, X. Shen, and W. J. Mark, Downlink resource management for packet transmission in OFDM wireless comunication system, vol. 4, no. 4, pp , [11] G. Kulkarni, S. Adlakha, and M. Srivastava, Subcarrier allocation and bit loading algorithms for OFDMA-based wireless networks, vol. 4, no. 6, pp , [12] R. Jain, D. Chiu, and W. Hawe, A quantitative measure of fairness and discrimination for resource allocation in shared computer systems, DEC Research Report TR-301, 1984.