Exploiting Spectral Rese in Resorce Allocation, Schedling, and Roting for IEEE 802.16 Mesh Networks Lien-W Chen 1, Y-Chee Tseng 1, Da-Wei Wang 2, and Jan-Jan W 2 1 Department of Compter Science, National Chiao-Tng University, Hsin-Ch, 30050, Taiwan Email: {lwchen, yctseng}@cs.nct.ed.tw 2 Institte of Information Science, Academia Sinica, Taipei, 11529, Taiwan Email: {wdw, wj}@iis.sinica.ed.tw Abstract The IEEE 802.16 standard for wireless metropolitan area networks (WMAN) has been created to meet the need of wide-range broadband wireless access at low cost. The objective of this paper is to stdy how to exploit spectral rese in an IEEE 802.16 mesh network throgh timeslot allocation, bandwidth adaptation, hierarchical schedling, and roting. To the best of or knowledge, this is the first work which formally qantifies spectral rese in IEEE 802.16 mesh networks and which exploits spectral efficiency nder an integrated framework. Simlation reslts show that the proposed spectral rese schedling and load-aware roting significantly enhance the network throghpt performance in IEEE 802.16 mesh networks. Keywords: IEEE 802.16, WiMax, Mesh Network, Resorce Allocation, Roting, Wireless Network. I. INTRODUCTION The IEEE 802.16 standard for wireless metropolitan area networks (WMAN) is designed for wide-range broadband wireless access at low cost. It is based on a common medim access control (MAC) protocol with different physical layer specifications. The PHY layer can employ the orthogonal freqency division mltiplexing (OFDM) below 11GHz or the single carrier (SC) scheme between 10GHz and 66GHz. The MAC layer of IEEE 802.16 [4] can spport the pointto-mltipoint (PMP) mode and the mesh mode. In the PMP mode, sbscriber stations (SSs) are directly connected to base stations (BSs). So all SSs associated to a BS mst be within the transmission range of the BS. On the other hand, in the mesh mode, each SS can act as an end point or a roter to relay traffics for its neighbors. So there is no need to have a direct link from each SS to its associated BS, and SSs may transmit at higher rates to their parent SSs/BS. Also, a BS can serve wider network coverage with lower deployment cost and higher robstness and flexibility [3]. However, intelligent roting and schedling protocols are needed to flly exploit sch benefits. For IEEE 802.16 mesh networks, efforts have been dedicated to topology design [10], packet schedling [8], and QoS spport [1]. This paper stdies the spectral rese isse in an IEEE 802.16 mesh network throgh mlti-hop roting and schedling. The Y. C. Tseng s research is co-sponsored by Taiwan MoE ATU Program, by NSC grants 93-2752-E-007-001-PAE, 96-2623-7-009-002-ET, 95-2221- E-009-058-MY3, 95-2221-E-009-060-MY3, 95-2219-E-009-007, 95-2218-E- 009-209, and 94-2219-E-007-009, by Realtek Semicondctor Corp., by MOEA nder grant nmber 94-EC-17-A-04-S1-044, by ITRI, Taiwan, by Microsoft Corp., and by Intel Corp. TABLE I COMPARISON OF EXISTING SCHEMES AND OUR RESULTS Schedling Roting Rese Slot Rote Load Featres Qantification Assignment Reconstrction Awareness Wei et al. [2] N/A Yes N/A N/A Tao et al. [5] N/A Yes Yes N/A F et al. [6] N/A N/A N/A N/A Or work Yes Yes Yes Yes proposed framework incldes a load-aware roting algorithm and a centralized two-level schedling scheme, which consider both traffic demands and interference among SSs. Given traffic patterns of SSs, we show how to achieve better spatial rese and ths higher spectral efficiency. Table I compares or work against previos works. Reference [2] proposes an interference-aware rote constrction and a schedling algorithms. However, the algorithm does not flly exploit spectral rese and it is not load-aware (in the sense that the roting tree is a fixed one). How to attach a new SS to a mesh tree is discssed in [6], bt schedling is not addressed in that work. As pointed ot in [5], the network performance highly depends on the order that SSs join the roting tree. Althogh [5] has taken roting tree reconstrction into accont, the traffic demands of SSs are still not considered. Ths, the real traffic bottleneck of the network is not reflected. Compared to existing works, or work is most complete in exploiting spectral rese in IEEE 802.16 mesh networks in the sense that it takes dynamic traffic loads of SSs into accont and integrates not only a hierarchical bandwidth schedling scheme for bandwidth adaptation and timeslot allocation, bt also a roting algorithm with a tree optimization scheme. The rest of the paper is organized as follows. Section II briefly reviews the IEEE 802.16 mesh mode and then formally defines or problem. Section III develops or resorce allocation and schedling framework, followed by or roting and tree constrction algorithms. Performance evalation is given in Section IV. Finally, Section V concldes this paper. II. BACKGROUNDS AND PROBLEM DEFINITION In an IEEE 802.16 mesh network, transmission schedles of SSs can be determined in a distribted manner by individal SSs, or in a centralized manner by the BS. In this work,
to better exploit spectral rese, we will focs on centralized schedling, which is also most commonly sed in the standard for Internet access. In centralized schedling, there are two control messages, MSH-CSCF (Mesh Centralized Schedling Configration) and MSH-CSCH (Mesh Centralized Schedling). The BS can specify the crrent roting tree by sing the last MSH-CSCF message and modify the tree by the last MSH-CSCH pdate. The BS will broadcast MSH-CSCF to all its neighbors, and all the BS neighbors rebroadcast this message to all their neighbors ntil all SSs have received the MSH-CSCF message. As a reslt, all SSs maintain a roting tree whose root is the BS and child nodes are SSs. On the other hand, SSs can transmit MSH-CSCH:Reqest messages to the BS for their traffic demands, which the transmission order is that the SS with the largest hop cont transmits first, and retain the order to join the network for SSs with the same hop cont. After collecting reqests from all SS, the BS can broadcast its flow assignment for all SSs by the MSH-CSCH:Grant message. Since all SS know the crrent roting tree, they can determine the actal schedle from these flow assignments by dividing the frame proportionally. In this work, we consider a mesh network with a gateway BS and a nmber of SSs for Internet access. For centralized schedling, given the roting tree, the bandwidth demand reqested by each SS, and the plink and downlink data rates of each SS, a two-level schedling scheme is designed for the following prposes: (1) dynamically adapt the bandwidths between plink and downlink sbchannels; (2) proportionally allocate frame timeslots among SSs; (3) obtain higher gateway throghpt based on the above two manners. On the other hand, for roting tree constrction, given the traffic demand generated by each SS and the data rate of each link between SSs, a load-aware roting algorithm is developed for constrcting a load-balancing roting tree that can distribte evenly the forwarding data of all SSs and increase concrrent transmissions among SSs so as to get higher timeslot rese ratio. III. THE PROPOSED SPECTRAL REUSE FRAMEWORK A. System Model We propose an integrated spectral rese framework for IEEE 802.16 mesh networks, as illstrated in Fig. 1. There are a roting and a schedling modles. The roting modle collects the channel conditions and bandwidth reqests of all SSs from MSH-CSCH:Reqest messages and comptes a roting tree T for the mesh network. Next, the schedling modle condcts resorce allocation, which contains channellevel schedling (for bandwidth adaptation between plink and downlink sbchannels) and link-level schedling (for timeslot allocation among SSs). Finally, the BS broadcasts the schedling information to all SSs via MSH-CSCH:Grant messages. Below, we will focs on plink traffic schedling, since downlink traffic schedling can be obtained similarly. Fig. 1. The system model at BS B. Resorce Allocation and Schedling Schemes Below, we assme that the roting tree T is known (refer to Sec. 3-3 for the constrction of T ). We will derive or resorce allocation schemes. Let the plink data rate and plink traffic demands of SS i be ri and b i, respectively. From T,we can calclate the aggregated plink traffic demand d i = b i + j child(i) b j for SS i, where child(i) is the set of children of i in T. Ths the demand of transmission time for the plink of SS i is Ti = d i /r i.letc total = i T i be the total plink transmission time of the network, and Ci = j E i Tj be the total plink transmission time of extended neighborhood of SS i, which contains SS i and its one-hop and two-hop neighbors. In the IEEE 802.16 standard, only a portion of Ti /C total is allocated to the plink transmission time of SS i. Clearly, SS i can detect bsy carriers only in Ci /C total portion of time. In the remaining (1 Ci /C total ) portion of time, SS i sees idle carriers. Or scheme is designed to exploit this portion of idle time for additional transmissions by raising the same ratio of allocated transmission time for all SSs. For the fairness of all SSs in E i, the portion of idle time shold be divided proportionally by their transmission time demands. Ths the additional transmission time SS i can obtain is (1 Ci /C total ) T i /C i. So the maximal transmission time with spatial rese for SS i in the mesh network is Ti /C total +(1 C i /C total ) T i /C i = Ti /C i.let Cmax = max{ci, i}. For any SS i sch that C i = Cmax, the SS cold be the bottleneck of the network. Therefore, we propose to assign Ti /C max portion of plink transmission time to each SS i. It is clear that after assigning Ti /C max portion of time to each SS i, the bottleneck SS will see 100% bsy carriers, whereas other SSs sch that Ci <Cmax can see some idle carriers. On the other word, we raise the same ratio of plink transmission time for each SS i from Ti /C total to Ti /C max ntil the bottleneck SS sees 100% bsy carriers. As a reslt, the smaller Cmax the mesh network can rote, the larger transmission time each SS can get. Note that althogh the maximm of Ci among all SS i is sed in the mesh network so that Ti /C max is the lower bond of spectral rese, actally the lower bond is also an pper bond when Cmax is occrred at the one-hop neighborhood of the BS in most reglar mesh networks since all the BS neighbors can not transmit or relay more data for themselves or other child SSs. Continosly, or two-level schedling scheme with spectral rese qantified above will be described in the following
Fig. 2. The timeslots allocated in phase I and phase II sbsections. 1) Channel-Level Schedling: The mesh mode spports only Time Division Dplex (TDD) to share the channel between downlink and plink. The TDD framing is adaptive in that the bandwidth allocated to the downlink verss the plink can vary. The split between plink and downlink is a system parameter and is controlled at higher layers within the system. In or channel-level schedling scheme, the ratio of downlink to plink sbchannel will be set to Cmax/C d max that fits the traffic load distribtion. Therefore, the first F Cmax/(C d max d + Cmax) timeslots in each frame are assigned to downlink sbchannel and the rest timeslots are assigned to plink sbchannel, where F is the nmber of timeslots in a frame. The well-arranged sbchannel bandwidth for plink and downlink cold reslt in that the overall network throghpt is increased significantly, which has been validated by simlation reslts in Section IV. 2) Link-Level Schedling: In IEEE 802.16 mesh networks, SSs notify the BS their data transfer reqirements and the qality of their links to their neighbors. The BS ses the topology information along with the reqirements of each SS to decide the roting and the schedling withot spectral rese. The frame fraction assigned to each SS i is Ti /C total for plink traffic in the IEEE 802.16 mesh mode specification, whereas the fraction is Ti /C max in or schedling with spectral rese as mentioned at the beginning of Section III-B. Note that Cmax is mch smaller than Ctotal in a large IEEE 802.16 mesh network, which implies each SS can obtain mch larger frame fraction from or schedling algorithm. For timeslot assignment, assme that there are N timeslots in a frame for plink sbchannel. We first allocate N (Ti /C total ) timeslots in phase I and then assign N (Ti /C max Ti /C total ) timeslots in phase II, which the total allocated timeslots to SS i is N (Ti /C max). The allocated timeslots in phase I are assigned to each SS i in the mesh network according to its hop cont from the BS, and retain the order to join the network for SSs with the same hop cont. The allocated timeslots in phase II are inserted to the remaining space of frame allocation list for all SS j in E i. As illstrated in Fig. 2, since the forwarding order for all SSs in the mesh network can be hold in phase I and ths the end-to-end delay between the BS and SSs can be minimized, SSs can tilize it by transmitting real-time traffic in order to redce the packet delay. On the other hand, SSs can se the allocated timeslots in phase II withot forwarding order to transmit non-real-time or best effort traffic since the packet delay is not crcial even thogh the end-to-end delay may be the dration of several frames. Note that the sm of the allocated timeslots for the SSs in the extended neighborhood with Cmax eqals to N exactly. Therefore, there are sfficient free timeslots in a frame to insert the allocated timeslots in phase I and phase II for those SSs in the extended neighborhood with Ci that is smaller than Cmax. The link-level schedling algorithm is described as follows. Link-level schedling algorithm Phase I: Allocate N (T i /C total ) timeslots to each SS i according to the transmission order of MSH-CSCH:Reqest ntil all SSs have been allocated. Phase II: (1) Constrct the frame allocation list L i of E i for each SS i in the network. (2) According to the transmission order of MSH-CSCH:Reqest, assign the first N (T i /C max T i /C total ) free timeslots in Li to SS i. (3) Update all frame allocation lists L j that E j incldes SS i. (4) Repeat steps (2) and (3) ntil all SSs have been assigned. C. Roting Tree Constrction The roting tree constrction problem investigated in this section is to find a roting tree with the minimm Cmax in a directed mesh network graph G = (V,E) according to the traffic demand b i reqested by vertex i V and the plink data rate rj of edge j E. We first prove that the roting tree constrction problem is a NP-complete problem, and then propose a load-aware roting algorithm to redce Cmax for spectral efficiency. Below, we show the roting tree constrction is NP-complete by proving that its decision problem is NP-complete. The Problem Given a directed mesh network graph G =(V, E), the traffic demand b i reqested by vertex i V, the plink data rate r j of edge j E, and a real nmber R, determine whether G has a roting tree sch that its C max R. Theorem 1 The roting tree constrction problem is NP-complete. Proof: The roting tree constrction belongs to NP, since we can gess a roting tree and check whether its C max R easily in polynomial time. To prove that the roting tree constrction problem is NP-complete, we have to redce an NP-complete problem to it. We se the partition problem: the inpt is a set X sch that each element x X has an associated size s(x). The problem is to determine whether it is possible to partition the set into two sbsets with exactly the same total size. [7] Consider a special case of mesh networks in Fig. 3. Assme that E a and E b are not overlapped, all plink data rates in E a and E b are the same and low enogh sch that C max is max{c a,c b }, and there are n SSs (x 1,x 2,...,x n ) be
neighbors of SS c and SS d. Let the traffic demands of all SSs in the mesh network except x 1, x 2,..., and x n be zero. Now we start to redce the partition problem to the special case of the roting tree constrction problem. Let X = {x 1,x 2,...,x n }, s(x k ) be the traffic demand of x k for k =1, 2,...,n, and R =5/2 k s(x k)/r slow, where r slow is the data rate of slow link in Fig. 3. The parent node of x k is either vertex c or vertex d. Ths, we can get the smallest Cmax by partitioning X(x 1,x 2,...,x n ) into two sbsets (SS c and SS d) with exactly the same total size. Therefore, if there is a roting tree sch that Cmax = R in G, then there is a partition to divide X into two sbsets with exactly the same total size. This redction can obviosly be performed in polynomial time. Since the special case of the roting tree constrction problem is NP-complete, the general case is also NP-complete. To achieve efficient spectral rese and high throghpt in IEEE 802.16 mesh networks, we propose a load-aware roting algorithm to redce Cmax for plink traffic. In or algorithm, we assme the initial vale of Ci is j E i d j /r j (max) for each SS i in the mesh network, where d j = b j and rj (max) is the highest data rate among links of SS j to its neighbors with less or eqal hop cont. The tree constrction ses a bottom-p fashion that each SS i with the largest hop cont to the BS will be first attached to its neighbors k which have less or eqal hop cont to estimate each new Ck, and then the SS which has minimm C k will be chosen as the parent node of SS i. If there are several SSs with the same minimm Ck, the SS with smaller hop cont has the higher priority. Once each SS with largest hop cont has been attached to its parent node, the remaining SSs withot a parent node repeat the above procedre ntil each SS in the mesh network has a parent node. Note that the step (2) in load-aware roting algorithm is to bild the sbtree with the minimm Ck first, which can balance the distribtion of forwarding traffic and frther redce Cmax. Load-aware roting algorithm (1) Let A be the set of SSs withot a parent node that have the largest hop cont, and B the empty set (2) Estimate each C k for all neighbors k with less or eqal hop cont when SS i in A becomes the child of SS k, and the SS with the smallest C k will be chosen as the parent node of SS i (3) Remove SS i from A, addssi into B, and pdate C l for all SS l E i E k (4) Repeat steps (2) and (3) ntil there is no SS in A (5) Repeat steps (1) (4) ntil each SS has a parent node The analysis of time complexity is as follows. Since each SS only has a parent node, steps (2) and (3) jst repeat n times, where n is the nmber of SSs in the network. The dominant part of steps (2) and (3) is the step (2) that selects the smallest one from at most m d estimated Ck vales, where m is the maximm nmber of SSs with the same hop cont, and d is the maximm degree of SSs. Therefore, the algorithm takes O(nmd) time to bild the roting tree. Fig. 3. Fig. 4. The special case of the roting tree constrction problem The node placement in the reglar mesh topology IV. PERFORMANCE EVALUATION In this section, we provide ns-2 [9] simlation reslts for the spectral rese framework and compare it with the basic 802.16 mesh operation in [4] as well as the concrrent transmission with rote adjstment in [5]. The typical TCP/IP/LL/MAC/PHY stack is sed in or stdy. In addition, we adopt a single channel OFDM PHY and two-ray grond reflection model for radio propagation, and all the SSs are stationary and working in half dplex. In or work, we extend the TDMA MAC modle in ns-2 for timeslot rese in a mltihop environment and se it to stdy the system performance. In or simlation, the node placement in the reglar mesh topology is shown in Fig. 4. There are totally at most 85 nodes which consist of a single BS (node 0) and 84 SSs (node 1 84), and the one-hop neighbors are connected by lines. The channel bandwidth is set to 50 Mb/s and the data rates of all links are the same for simplicity. The extended neighborhood of each SS incldes one-hop and two-hop neighbors. The random roting tree is sed in the basic 802.16 mesh mode and or link-level schedling except that the load-aware roting is marked on the figres. Note that the overall network throghpt has been normalized by the performance of basic 802.16 mesh operation so that the scalability and improvement of or proposed framework are clearly demonstrated in the simlation reslts. Fig. 5 shows the normalized gateway throghpt with different schedling and roting methods, respectively. The
0.5 3 2.5 Basic 802.16 mesh Link-level schedling Load-aware + Link-level Concrrence + Adjstment 50% plink traffic and 50% downlink traffic 4 3.5 Basic 802.16 mesh + 50% fixed plink bandwidth Channel-level schedling + Link-level schedling Load-aware roting + Channel-level + Link-level 50% to 100% plink traffic Normalized overall throghpt 2 1.5 Normalized overall throghpt 3 2.5 2 1.5 1 1 10 20 30 40 50 60 70 80 0.5 10 20 30 40 50 60 70 80 Nmber of SSs Nmber of SSs Fig. 5. The performance comparison for link-level schedling Fig. 6. The performance comparison for channel-level schedling nmber of SSs increases from 4, 12, 24, 40, 60, to 84 and all SSs reqest the same bandwidth for plink and downlink. The throghpt vales are the average of simlation in 100 times with random load distribtion among SSs. As shown in Fig. 5, the proposed link-level schedling scheme otperforms the basic mesh mode significantly. Also, the roting tree generated by the load-aware roting algorithm frther improves the throghpt. It is becase that in the basic 802.16 mesh scheme, the network throghpt drops significantly as the nmber of SSs increases de to the fact that a packet needs to be forwarded several times since the length of relay rote increases with the nmber of SSs in the network, whereas the proposed link-level schedling is mch more scalable than the basic scheme since the degree of spectral rese increases with the network size. In addition, the load-aware roting algorithm prodces better roting paths to distribte the traffic more evenly in the mesh network. Therefore, the scheme with both the load-aware roting and link-level schedling achieves the highest network throghpt. The scheme only sing link-level schedling still has the second best performance. On the other hand, since there is no schedling algorithm provided in [6] and the concrrent transmission scheme in [5] otperforms that withot rote adjstment in [2], we also compare the performance of concrrent transmission with rote adjstment in the simlation. The non-load-aware roting method constrcts a roting tree according to the SS positions, which can not release the traffic bottleneck in the network efficiently. Ths, the benefit of rote adjstment has been limited in the natre nless every SS generates the same traffic load nder the same link data rate. In addition, the concrrent transmission algorithm forces SSs can not transmit data earlier than their child SSs so that the tilization of spectral rese is redced significantly. Therefore, its throghpt improvement is mch lower than or integrated spectral rese framework. Fig. 6 shows the normalized overall throghpt with channel-level and link-level schedling schemes. The configration of simlation is as same as in Fig. 5. However, every SS reqests 50% to 100% plink bandwidth randomly, and ths the downlink bandwidth reqested is 0% to 50% which depends on the plink bandwidth reqested. Note that the basic 802.16 mesh mode allocate the bandwidth eqally for plink and downlink sbchannels. As shown in Fig. 6, the proposed channel-level and link-level schedling scheme otperforms the basic mesh mode more significantly. Again, the combined roting and schedling scheme gets the highest system throghpt. This is becase that channel-level can adapt dynamically the bandwidth between plink and downlink sbchannels based on the traffic load distribtion in the mesh network. Using load-aware roting, the network throghpt can be enhanced as the nmber of SSs increases. As a reslt, the combination of channel-level and link-level schedling as well as load-aware roting can fit more traffic patterns so as to keep high network performance. V. CONCLUSIONS In this paper, we have formally qantified spectral rese in IEEE 802.16 mesh networks. Also, an integrated spectral rese framework for centralized schedling scheme and roting tree constrction is developed. Compared to existing works, or work is most complete in exploiting spectral rese in IEEE 802.16 mesh networks in the sense that it takes dynamic traffic loads of SSs into accont and integrates bandwidth adaptation, timeslot allocation, as well as roting tree constrction nder a framework. Simlation reslts indicate that the spectral rese schedling and load-aware roting significantly increase the overall throghpt in IEEE 802.16 mesh networks. REFERENCES [1] H. Shetiya and V. Sharma. Algorithms for Roting and Centralized Schedling to Provide QoS in IEEE 802.16 Mesh Networks. In WMNeP 05, Oct. 2005. [2] H.-Y. Wei, S. Gangly, R. Izmailov, and Z. Haas. Interference-Aware IEEE 802.16 WiMax Mesh Networks. In VTC Spring 05, May 2005. [3] I. F. Akyildiz, X. Wang, and W. Wang. Wireless Mesh Networks: A Srvey. Compter Networks Jornal (Elsevier), Jan. 2005. [4] IEEE Standard 802.16-2004. IEEE Standard for Local and metropolitan area networks - Part 16: Air Interface for Fixed Broadband Wireless Access Systems. Oct. 2004. [5] J. Tao, F. Li, Z. Zeng, and Z. Lin. Throghpt Enhancement in WiMax Mesh Networks Using Concrrent Transmission. In WCNM 05, volme 2, pages 871 874, Sept. 2005. [6] L. F, Z. Cao, and P. Fan. Spatial Rese in IEEE 802.16 Based Wireless Mesh Networks. In ISCIT 05, volme 2, pages 1358 1361, Oct. 2005. [7] M. Udi. Introdction to Algorithms: A Creative Approach. Addison- Wesley Pblishing Company, 1989. [8] S.-M. Cheng, P. Lin, D.-W. Hang, and S.-R. Yang. A Stdy on Distribted/Centralized Schedling for Wireless Mesh Network. In IWCMC 06, pages 599 604, Jly 2006. [9] The Network Simlator - NS-2. http://www.isi.ed/nsnam/ns/, 1989. [10] V. Gnasekaran and F. C. Harmantzis. Affordable Infrastrctre for Deploying WiMAX Systems: Mesh v. Non Mesh. In VTC Spring 05, volme 5, pages 2979 2983, May 2005.