QoS-aware Distributed Resource Allocation for Hybrid ...

WPMC06 San Diego, CA,USA Sep.2006

QoS-aware Distributed Resource Allocation for Hybrid FDMA/TDMA Multicellular Networks Simone Merlin, Luca Begnini, Andrea Zanella, Leonardo Badia, Michele Zorzi University of Padova – Department of Information Engineering Via Gradenigo 6/B, 35131 Padova, Italy e–mail: {simone.merlin, andrea.zanella, luca.begnini, leonardo.badia, michele.zorzi}@dei.unipd.it phone: +39 049 827 7656 fax: +39 049 827 7699

Abstract— † This paper considers a multicellular network using a hybrid FDMA/TDMA medium access with a complete reuse of the resources among neighboring cells. The resource allocation mechanism runs in a completely distributed way, tracking the traffic dynamics, adapting itself to the channel and interference condition and providing QoS differentiation among different traffic classes. The proposed algorithm performs both spectral efficiency optimization, by exploiting information on channel and interference, and traffic differentiation at MAC level, by means of a weighted share of the bandwidth among different priority traffic flows. The traffic differentiation mechanism, which is the original contribution presented in this paper, works locally on each cell but exploits values of interference measurements in order to acquire information on the neighboring cells load, with the goal of implementing a collaborative algorithm. Interference measurement is the only form of communication among cells and no explicit signalling is assumed. The aim of the proposed algorithm is to provide traffic differentiation at the MAC level in a global perspective, where the share of the bandwidth inside a cell depends also on the priority of the traffic in the surrounding cells. Finally, the algorithm is tested and validate on a multicellular network with a static and a time–variant channel, thus confirming its behavior.

I. I NTRODUCTION Future generation wireless systems are expected to provide ease of deployment, spectral efficiency, dynamic adaptation to the working condition and QoS mechanism support. The embedding of all such requirements inside innovative systems needs a careful and joint design of different layers. A cross-layer interaction allows the development of efficient and powerful algorithms which can use a complex set of information coming from different functional layers. Such information can be exploited to pursue complex optimization goals, which can involve performance metrics referred to many network aspects. For what concerns physical layer, great attention has been recently devoted to OFDM based systems, since this techniques has been proven to allow for robust transmission in multi–path channel environment obtaining high spectral efficiency. Many recently standardized wireless transmission systems such as IEEE 802.16, IEEE 802.11g, Hyperlan2, DVB-T have adopted OFDM, which seems to be one of the most promising technique for future wireless systems. OFDM modulation can be easily used to realize flexible and efficient OFDMA medium access techniques embedding AMC (Adaptive Modulation and Coding). Such flexibility can be fully exploited only if sufficient information coming from the physical layer is available. Many resource allocation algorithms have been proposed to exploit the frequency and multiuser diversity with the aim to reduce the power consumption [1] or increasing the aggregated rate [2]. Such algorithms mainly exploit information coming from physical layer This work was supported by MIUR within the framework of the ”PRIMO” project FIRB RBNE018RFY (http://primo.ismb.it/firb/index.jsp).

only, in order to assign subchannels to users, setting also the power level and the bit load [1], [3], [4], [5]. Clearly, aggregate throughput and power consumption are not the only performance metric of interest, in particular when the system is required to support traffic with different quality requirements. In this case, admission control and scheduling algorithms have to be deployed, to enforce flow–level QoS requirements or traffic differentiation/prioritization [7]. In particular, traffic differentiation is required when real time and best effort flows coexist in the same system. Pricing policies can also be the reason for allotting different amount of resources to different users [6]. It is important to remark that the resource allocation at the PHY layer is tightly related to the packet scheduling at the data link layer, so that a joint design of these two mechanisms is a very natural approach [8], [9], [10]. Most of the previously cited papers restrict the optimization problem to a single cell. Clearly, this is a simplification since, in a multi– cell environment, the resource allocation and scheduling strategies adopted in a cell determine the interference produced on the surrounding cells, thus impacting on the optimization strategies of those cells. The optimization of the resource allocation and scheduling, hence, should be performed on the entire network. The solution of such a problem would often require a centralized approach, which implies a huge amount of control signalling. On the other hand, fully distributed approached would lead to suboptimal performance. Hybrid solutions, which allow for a limited information exchange among neighboring cells, lie in between and, hence, represent a promising approach to the problem [11]. In this paper a multicellular system is considered, which is based on OFDMA/TDMA medium access and works with a complete resources reuse among cells. No explicit communication is allowed among cells, thus the resource allocation has to be performed in a distributed way and each base station (BS) takes charge of the resource allocation for its mobile terminals (MTs). Traffic flows belong to different classes and to each class a different priority is associated. Note that the inter-cell traffic is centrally managed at the BS, but the class priority still has to be used to regulate contention for the resources with neighboring cells. To this end, a joint PHY resource allocation and scheduling mechanism is presented, which accounts for both PHY layer and flow level information. It jointly assures transmission efficiency and enforces traffic differentiation at MAC level among interfering transmissions. The main goal of the proposed algorithm is to resolve the contention among flows of different cells, competing for the same resources. The mechanism grants higher resource share to the higher priority transmissions among the interfering ones. To clarify this point let us assume two mutually interfering cells, one loaded with a low priority traffic and the other loaded with a high priority traffic. Since

A single logical subchannel, called Physical Basic Unit (PBU), is assumed to consist of a set of adjacent hsubcarrier, time–sloti pairs, which are selected in order to exploit the channel coherence in the time and frequency domains, as shown in Fig. 2. The radio channel is assumed to be flat over a PBU, so that all the hsubcarrier, time–sloti pairs of the same PBU are equally power and bit loaded. Therefore, the resource addressing space can be associated to a N s × N t matrix of PBUs, being N s and N t the number of different frequency and time–slot blocks, respectively. PBU Subcarrier (Ns blocks)

resources are fully reused, low priority traffic can interfere on high priority one. Complying with the priority weighting, the former cell has to relinquish the resources in order to give advantage to the latter. Such mechanism clearly requires each cell to be aware of the interference traffic priority. The interference priority detection is embedded in the proposed algorithm and is based on the interference level measurement only, without any explicit signalling. Resources which are not interfered are allocated with the only aim to maximize the system capacity. A heuristic solution to this last problem is proposed as a part of the whole algorithm. The system is tested in a multicellular scenario both with static and time-variant channels. To the best of our knowledge, such an approach has never been presented before in the literature. The rest of the paper is organized as follows. In Section II the reference scenario is described, together with the framework where the algorithm is developed. Section III describes the proposed algorithm. In Section IV results are shown and conclusions are drawn in Section V. II. S YSTEM DESCRIPTION In this section we describe the system scenario and the framework that the proposed algorithm is applied to.

Time slots (Nt blocks) Fig. 2.

BS functional blocks

A. Scenario The multicellular system [12] is composed by a 3x3 grid of cells. MTs are uniformly scattered on the area and connected to the BSs with the best channel at the moment of system initialization. Topology is assumed to be wrapped at the simulation area limit in order to avoid border effect. Complete frequency reuse among neighboring cells is assumed. This obviously rises the issue of severe interference among cells, so that a smart resource allocation mechanism is required. Since no control communication is allowed among cells, the resource allocation scheme needs to be completely distributed. Two different channel models are considered. The first model assumes a static channel with path–loss exponent α = 4, shadowing σ = 6 dB and delay spread D = 0.5 µs. The second model adds a time variation according to a doppler effect correspondent to a speed of 1 m/s. B. Algorithm framework Fig. 1 represents the framework that the proposed algorithm fits in. Each cell uses a hybrid FDMA/TDMA medium access scheme that, in practice, can be realized upon an OFDM physical layer. Detailed description of the physical layer and the related issues is out of the scope of this paper. Furthermore, at this stage of the work, we have focused only on downlink (from BS to MT) connections. Traffic Sources RRM Request selection and transmission scheduling Interferece estimation Resources classification PHY efficiency optimization

PHY

Fig. 1.

BS functional blocks

According to the complete frequency reuse hypothesis, each PBU can be assigned to multiple users, provided that they belong to different cells. A communication channel is realized by allocating a set of PBUs to a user and assigning the transmission power level and the bit–load to each PBU. Such operations are performed by the joint scheduling/allocation module at the BS. To this end, the BS makes use of PHY–layer information provided each MT in its cell. More in detail, we assume each MT can perfectly estimate the channel gain and interference level experimented in each PBU of the resource matrix. At the beginning of each frame, mobile nodes communicate such PHY–layer information to the BS, by using a separate error-free channel. Each traffic request is marked by the destination user address, the priority q associated to the carried data and the minimum required bit rate. Requests belong to Q different priority classes. Based on this entire information the algorithm on each frame selects the requests to be served and assigns them one or more PBUs, setting also the power and bit-load. Furthermore, the algorithm can decide to deallocate resources to previously accepted flows, depending on the dynamic of the resource status and on the flow priority. The mechanism has two goals: (i) to efficiently allocate the physical resources in terms of adaptation to the channel and interference; (ii) to assure a service differentiation among traffic fluxes of different priorities, even if they belong to different cells. It is worth recalling that no explicit communication is allowed among cells, so, in order to pursue the service differentiation, the allocation/scheduling block needs to estimate the priority of interference source using interference measurements. In the proposed system, the priority differentiation goal can be pursued by acting on several parameters, such as number and quality of PBUs assigned to each flow, time share of the medium, power level and bitload. Traffic differentiation can be obtained by acting on several parameters, such as number and quality of PBU assigned to each flow, time share of the medium, power level and bitload. The proposed algorithm acts only on the PBUs assignment to each traffic class.

III. P ROPOSED ALGORITHM As previously stated, the proposed algorithm realizes the differentiation among priority classes by acting solely on the number of per–priority available resources, and the medium sharing in the time domain. The algorithm can be decomposed in three functional blocks: 1) detection of the interferant priority and classification of the physical resources; 2) frame by frame selection of the admissible pending requests (QoS based admission control); 3) allocation of the physical resources to the accepted flows. Based on the interference level measurements, each PBU is classified according to the estimate priority of the interfering transmissions. In particular, to each PBU hs, ti, with s ∈ {1, . . . , N s}, t ∈ {1, . . . , N t}, is associated an indicator Ps,t , which gives an estimate of the priority of the interference source (the way in which the estimation is performed will be described in the following.) This means that a request k with priority Pk can be transmitted only on PBUs for which Pk ≥ Ps,t . Therefore, new requests can be allocated only on PBUs satisfying the prioritization, while previously allocated requests can be deallocated if an update of Ps,t causes the priority relationship to be violated. Each traffic request k is assumed to consist of a bulk of L bits that requires a minimum transmission rate Rk,min bits/s (in order to meet delay constraints). Therefore, it is possible to determine the maximum number of frames RDk = L/Rk,min required to complete the request service. PBUs allocated to a flow are retained until all the L bits have been transmitted or the priority condition Pk ≥ Ps,t is violated. On each frame, for all allocated PBUs a power adjustment is performed in order to guarantee the already allocated bit load, adapting to the interference and channel changes. Clearly, it can happen that two interfering transmission start to rise their transmission power frame by frame, trying to satisfy the target SNR needed to supply the loaded bits. Such power increase results in an interference increase that can be tracked trough time. This behavior is exploited to estimate the interference priority in the following way. Each PBU hs, ti is assigned a counter Ns,t that is cleared any time the PBU is allocated to a new user. Every time the BS detects an increase greater than I percent in the interfering power, the counter is incremented by 1 with probability pk . The probability pk depends on the priority of the traffic flow k that the PBU is assigned to. If the counter Ns,t exceeds the threshold N thk < RDk , then PBU’s priority Ps,t is set to Ps,t = Pk + 1. In this manner, the k–th flow is deallocated, since it violates the priority relation for that PBU. The counter threshold N thk depends on the priority of flow k. The parameters pk and N thk can be set in order to determine the persistence of the request k on the allocated resources, in the case they are interfered by another transmission. If no interference increase is detected the request is deallocated only at the end of the complete transmission. The rationale behind this behavior is that, if at some time it happens that Nk = N thk it is implicitly assumed that the interfering transmission has a greater priority and thus resources are relinquished. In order to enforce the priority differentiation, once the assignment Ps,t = Pk + 1 is performed (and resources are freed), the value Ps,t is kept for N F frames and then reset to the lower possible value. This allows for a time limited and priority based resource reservation. Values pk and N thk have been set according to the following reasoning. Let us consider consider two mutually interfering transmissions on a given PBU. Both the BSs increase their power on each frame, to overcome the interference increment. In this case, the

Ns Nt Tf BW

= = = =

13 30 10 ms 20 MHz

B L RD Q

= = = =

{320, 640, 960} bits/PBU 1.8 · 106 bits 15 Frames {1, 2, 3}

TABLE I S ALIENT SIMULATION PARAMETERS

probability that the PBU is deallocated to user k is: N thk −1

P [Nk ≥ N thk ] = 1 −

X j=0

!

RDk j pk (1 − pk )(RDk −j) . j

Playing with pk and N thk parameters, we can tune the persistence of the allocation of the contended PBU to user k. Hence, traffic differentiation is pursued giving a greater allocation persistency, i.e., a greater time share, to the higher priority flows. Up to now, we did not mention how the available resources are chosen and loaded. This is because such operations are independent of the QoS mechanism described before, even though it obviously impacts on its behavior. Let us assume a list of requests is eligible for allocation of the available PBUs, according to the previous resource reservation mechanism. The allocation mechanism assigns resources to requests in order to increase the physical transmission efficiency. The allocation, in general, requires to jointly select requests, PBU allocation, bit loading and power loading [2], [4], [3]. Here an optimization problem is defined, which aims at maximizing the allocated capacity. The solution is then approximated by a heuristic algorithm. It is worth noting that this algorithm does not account for traffic priority at all, since it only aims at physical layer efficiency optimization. The algorithm works as follows [13]. At the first step, a maximum power level for each PBU is fixed, so that the maximum Shannon capacity ck,s,t corresponding to each PBU for each user can be computed by exploiting the channel and interference information. The allocation algorithm operates in steps. On each step, one of the available PBU is assigned to a pending request k. Out of all the requests that have not yet reached the minimum required rate, the algorithm selects the request k with highest ηk , which gives the total capacity that the user would get if it were assigned all the available capacity: X ηk = ck,s,t , (s,t)∈F

where F is the set of unallocated PBUs. This strategy should serve first the users that make a better use of the available resources. Then, an efficiency metric ² is computed for each PBU as: ck,s,t ²k,s,t = P . i∈K ci,s,t This index allows us to compare the advantage of allocating the PBU hs, ti to the request k, rather than to any other request. The PBU with the highest ² is associated to the request k selected in the previous step. The bit–loading is set to the highest value below ck,s,t among the available set. Once all available PBUs have been allocated, if not all the request have been satisfied, the procedure is repeated with a higher value of the power per PBU. IV. S IMULATION RESULTS The algorithm has been tested on the scenario described in Section II. Key parameters are reported in Table IV. To test the algorithm, the requested traffic load on the central cell of the cluster has been composed by a fair mix of 3 different priority

Comparison between aggregated throughput of the central cell

Case (a)

20

15

Case (c)

Single Cell

70

70

70

60

60

60

60

50

50

50

50

40

30

40

30

Throughput percentage


Throughput [Mbps]

25

Case (b)

70


30

Case (a) Case (b) Case (c) Single cell


35

40

30

40

30

10 20

20

20

20

10

10

10

10

5

5

10 15 20 25 Average number of mobile station per cell

30

Fig. 3. Aggregated throughput of the central cell varying interference priority (a) : q = 1, (b) : q = 2, (c) : q = 3.

classes Q = {1, 2, 3}. In Fig. 4 it is shown how, in the single cell case (thus no interference is present), also the offered load actually consists of a fair mix of the three traffic classes. In this case the interference based prioritization mechanism is idle; only the capacity maximization block is active and assigns the PBUs according to the efficiency criteria above described. The traffic differentiation mechanism is tested by loading surrounding cells with traffic belonging to a single priority class. Results for three different values of priority are shown in Fig. 3 and Fig. 4, where a direct comparison with the single cell case is possible. Results in Fig. 3 refer to the aggregate throughput of the central cell under different interference condition. Clearly the central cell suffers the interference produced by surrounding cells. Moreover, the higher the priority of the interfering traffic, the lower the aggregated central cell throughput. Fig. 4 shows the share of the throughput of the central cell among traffic classes (q), at the varying of the interference priority (cases a), b) and c)). From this plot it can be deduced how the reduction of the central cell throughput is essentially due to the starvation of the traffic with priority lower than the interfering one. Thus the proposed algorithm is able to enforce a differentiation in the resource usage among different traffic class in an intra–cell perspective. Fig. 5 shows the power consumption per transmitted bit for each traffic class in the central cell, for the three interference cases. It can be noted how the power consumption is higher for lower priority traffic. Moreover, the lowest priority traffic experiences a relevant increase in transmission power as the interference priority increases. Such a behavior can be explained as follows. Let consider the case of two different priority transmissions on two neighboring cells. At the allocation instant, they chooses their most efficient resources. If it happens that the two transmissions select the same resources, the proposed algorithm eventually deallocate the lowest priority one and reserves the resources for the highest priority one. Thus, lowest priority flow is forced to select some other PBUs which physical efficiency can only be less or equal than the one of relinquished PBUs. Fig. 6 and Fig. 7 refer to the time variant channel model. In this case, the interference estimation is more challenging due to the variation of the channel attenuation through time. Hence, the choice

0

1

2 3 q (Priority)

0

1

0

2 3 q (Priority)

1

2 3 q (Priority)

0

1

2 3 q (Priority)

Fig. 4. Share of the throughput among different priority classes of the central cella, at the varying of the interference traffic priority: (a) : q = 1, (b) : q = 2, (c) : q = 3. Average power per bit −44 case (a) case (b) case (c)

−46 −48

q=1

−50 Power/bit [dB]

0 0

−52 −54 −56

q=2

−58 −60 −62 −64

q=3 0

5

10

15 20 25 Average number of MT per cell

30

35

Fig. 5. Average power per bit of the central cell at the varying of interference priority: (a) : q = 1, (b) : q = 2, (c) : q = 3.

of the threshold I is critical. In the proposed scenario it has been tested that a value of I = 5% leads to no priority differentiation at all, whereas a value of I = 10% leads to the results shown in Fig. 6 and Fig. 7 which are very close to the static channel case. In general, the optimum value depends on the speed of channel variation; for this reason its estimation might require a possibly more complex setup. V. C ONCLUSIONS A QoS based and interference aware resource allocation mechanism has been proposed in this paper. The algorithm is applied to a FDMA/TDMA multicellular system with a complete resources reuse. The allocation mechanism is fully distributed among cells since no control traffic is allowed among them. Different priority traffic classes are considered. The main goal of the proposed algorithm is to provide a traffic differentiation at the MAC level in an intra–cell perspective, adapting the resource allocation on a cell to the estimated priority of the

Comparison between aggregated throughput of the central cell 30 Case (a) Case (b) Case (c)

Throughput [Mbps]

25

20

15

10

5

0 0

5

10 15 20 25 Average number of mobile station per cell

30

Fig. 6. Aggregated throughput of the central cell at the varying of interference priority: (a) : q = 1, (b) : q = 2, (c) : q = 3. Time variant channel. Case (b)

Case (c) 70

60

60

60

50

50

50

40

30

20

10

0


70



Case (a) 70

40

30

20

10

1

2 3 Priority

0

40

30

20

10

1

2 3 Priority

0

1

2 3 Priority

Fig. 7. Aggregated throughput of the central cell at the varying of interference priority: (a) : q = 1, (b) : q = 2, (c) : q = 3. Time variant channel.

interfering traffic transmitted by neighboring cells. The mechanism exploits only interference measurements in order to estimate the priority of the interfering traffic and to enforce a collaborative resource allocation. The aim of the collaboration is to provide a greater resource share to the highest priority traffic classes without explicit signalling. The mechanism has been shown to attain the expected behavior in a static channel environment and, under an ad-hoc parameter tuning, also in a slowly time variant channel. R EFERENCES [1] Y. J. Zhang, K. B. Letaief, Multiuser Adaptive Subcarrier-and-Bit Allocation with Adaptive Cell Selection for OFDM Systems, IEEE Transactions on Wireless Communications, vol. 3, no. 5, pp. 1566-1575, September 2004. [2] W. Rhee and J.M. Cioffi, Increase in capacity of multiuser OFDM system using dynamic subcarrier allocation, Proc. of IEEE VTC 2000 Fall, vol. 2, pp. 1085-1089, Sept. 2000. [3] G. Song and Y. Li. Cross layer optimization for OFDM wireless networks part I: theoretical framework. IEEE Trans. on Wireless Commun., 4(2):614, 2005.

[4] G. Song and Y. Li. Cross layer optimization for OFDM wireless networks part II: algorithm development. IEEE Trans. on Wireless Commun., 4(2):625, 2005. [5] M. Chandrashekar and B. Srikrishna A Sub-optimal Joint Subcarrier and Power Allocation Algorithm for Multiuser OFDM IEEE Communication Letters, Vol. 9, No. 8, Aug. 2005. [6] L. Badia, S. Merlin, A. Zanella, M. Zorzi. Pricing VoWLAN Services through a Micro-economic Framework. IEEE Wireless Communications, Vol 13, Feb. 2006, Pages 6-13 [7] S. Hanbyul and B. G. Lee,Proportional-Fair Power Allocation with CDF-Based Scheduling for Fair and Efficient Multiuser OFDM Systems, IEEE Trans. on Wireless Commun., Vol. 5, No. 5, May 2006 [8] M. Ergen, S. Coleri, P. Varaiya, QoS Aware Adaptive Resource Allocation Techniques for Fair Scheduling in OFDMA Based Broadband Wireless Access Systems, IEEE Transactions on Broadcasting, vol. 49, no. 4, pp. 362-370, December 2003. [9] G. Song and Y. (G.) Li. Utility-based resource allocation and scheduling in OFDM-based wireless broadband networks. IEEE Communications Magazine, 43(12):127-134, 2005. [10] J. Cai, Shen X., and W. J. Mark. Downlink resource management for packet transmission in OFDM wireless comunication system, Trans. on Wireless Commun., 4(4):1688, 2005. [11] G. Kulkarni, S. Adlakha and M. Srivastava, Subcarrier Allocation and Bit Loading Algorithms for OFDMA-Based Wireless Networks, IEEE Trans. on mobile computing, vol. 4, no. 6, Nov/Dec 2005. [12] PRIMO project FIRB. http://primo.ismb.it/firb/index.jsp. [13] S. Merlin and A. Zanella. An efficient and adaptive resource allocation scheme for next generation cellular systems In Proc. of WPMC05, Aalborg, DK.