Chapter 13

Centralized Scheduling Algorithms for 802.16 Mesh Networks

Petar Djukic and Shahrokh Valaee

13.1 Introduction
IEEE 802.16 protocol [1,2] specifies two dierent modes of operation. The first mode of operation is the point-to-multipoint (PMP) mode. In the PMP mode, each 802.16 access point has a dedicated broadband connection to the Internet. Wireless terminals connect to the access points on their first hop and their traffic goes to the Internet through the access points broadband connection. The second mode of operation is the mesh mode, where access points are interconnected with wireless links. In the mesh mode, wireless terminals connect to the access points on their first hop. Then, the wireless mesh carries their traffic to the point-of-presence (POP) where it goes to the Internet (Fig. 13.1). Mesh networks decrease the cost of running the access points, as all access points in the same area share a single broadband connection. Current mesh networks use 802.11a technology to interconnect the mesh backbone [3,4]. However, 802.11a technology is a decade old and was not designed for mesh networks. In particular, 802.11 Medium Access Control (MAC) lacks the
267
2008 by Taylor & Francis Group, LLC

268

WiMAX/MobileFi: Advanced Research and Technology

Mesh Node

Point of Presence High Speed Internet Connection

Mobile Hotspot Hotspot

Wireless Backbone

Internet

Mobile User Mobile User Static Users

Figure 13.1

Mesh networks.

extensions to provide quality of service (QoS) in multihop wireless environments [5]. IEEE 802.16 introduces QoS in the mesh with time division multiple access (TDMA) MAC technology [1]. In 802.16, access points negotiate their end-to-end bandwidth with the POP. IEEE 802.16 MAC enforces end-to-end QoS by scheduling wireless transmissions on the links connecting the access point to the POP. IEEE 802.16 mesh protocol specifies two TDMA scheduling protocols: centralized and decentralized scheduling protocols. The Base Station (POP) uses the centralized scheduling protocol to establish networkwide schedules. In the centralized scheduling protocol, mesh nodes request bandwidth from the Base Station (BS). The BS uses the end-to-end bandwidth requests to assign bandwidth to every link in the network, so that the links form a routing tree terminating at the BS. The BS then multicasts the assignment to all mesh nodes. Mesh nodes use the link bandwidth assignment and a common scheduling algorithm to determine a global transmission schedule. In the decentralized scheduling protocol, mesh nodes negotiate bandwidth assignments with their neighbors without centralized coordination. This chapter reviews the 802.16 centralized scheduling protocol and research on scheduling algorithms for this protocol. Section 13.2 reviews the 802.16 mesh protocol, including the centralized and decentralized scheduling protocols. Section 13.3 gives a detailed description of the 802.16 centralized scheduling protocol. Section 13.4 reviews the current research in centralized scheduling algorithms and compares several published algorithms with each other.

13.2 802.16 Medium Access Control

In this section, we first describe the physical (PHY) layer used in 802.16 and compare it with the PHY layer used in 802.11a. Both PHY layers use orthogonal frequency division multiplexing (OFDM) in the 5-GHz frequency band. We then

2008 by Taylor & Francis Group, LLC

Centralized Scheduling Algorithms for 802.16 Mesh Networks 269

describe the 802.16 MAC and its centralized and decentralized scheduling protocols.

13.2.1 802.16 PHY Layer

IEEE 802.16 is a TDMA-based MAC protocol [1], built on an OFDM PHY layer. OFDM transforms blocks of bits into constant duration symbols carried on a set of frequency orthogonal pilot carriers. In 802.16, OFDM symbols are grouped into frames of equal length and frames repeat over time (Fig. 13.2). The duration of OFDM symbols, Ts, depends on the total bandwidth occupied by the carriers: at 20-MHz bandwidth Ts = 12.5 s and at 10-MHz bandwidth Ts = 25 s. As OFDM symbol duration for 10-MHz bandwidth is twice as long as the OFDM symbol duration for 20-MHz bandwidth, the raw bit-rate at 10-MHz bandwidth is half of the raw bit-rate at 20-MHz bandwidth (Table 13.1). IEEE 802.16 uses OFDM in the 5-GHz frequency band. 802.16 can use hardware with 10-MHz total bandwidth in the licensed frequency bands and hardware with 20-MHz bandwidth in the license-exempt frequency band. IEEE 802.11a PHY layer uses OFDM in the 5-GHz license-exempt frequency band with hardware operating at 20-MHz bandwidth. Table 13.1 compares 802.16 raw bit-rate at 10- and 20-MHz bandwidths to 802.11a raw bit-rate. The two PHY layers have comparable raw bit-rates, however 802.16 provides a TDMA MAC layer, which guarantees QoS [6,7].

13.2.2 802.16 Framing

IEEE 802.16 frames OFDM symbols for two reasons. First, the frame boundaries are used to synchronize the mesh. Second, the frame structure allows the division of control and data traffic into subframes (Fig. 13.2). The control subframe is divided into transmission opportunities, which are seven OFDM symbols long.
Frames Frames

Control Sub-frame OFDM symbols

Data Sub-frame

Figure 13.2 802.16 time division multiple acess (TDMA). OFDM, orthogonal frequency division multiplexing.

2008 by Taylor & Francis Group, LLC

270 WiMAX/MobileFi: Advanced Research and Technology Table 13.1 Comparison of 802.11a and 802.16 Bit-Rates
Data Bits/Symbol Modulation BPSK-1/2 BPSK-3/4 QPSK-1/2 QPSK-3/4 16QAM-1/2 16QAM-3/4 64QAM-2/3 64QAM-3/4 802.11 24 36 48 72 96 144 192 216 802.16 96 X 192 288 384 576 768 864 Bit-Rate Mbits/Second 802.11 6.0 9.0 12.0 18.0 24.0 36.0 48.0 54.0 802.16 10 MHz 3.84 X 7.68 11.52 15.36 23.04 30.72 34.56 802.16 20 MHz 7.68 X 15.36 23.04 30.72 46.08 61.44 69.12

In each control subframe transmission, three OFDM symbols are guard symbols and four OFDM symbols are used for data transmissions at the lowest rate of BPSK-1/2. The size of the 802.16 control subframe is 7 MSHCTRLLEN OFDM symbols, where MSH-CTRL-LEN is a parameter specified by the network operator. Details of the coordination mechanism in the control subframe are available in the 802.16 standard [1], as well as in our summary of the standard [2]. Data transmissions have timing similar to the timing of control packets; each data transmission has two or three OFDM guard symbols and a variable number of data-carrying OFDM symbols. However, 802.16 mandates that the total number of OFDM symbols in each data subframe transmission be a multiple of the number of slots in one data subframe transmission opportunity. The number of OFDM symbols in each data subframe transmission opportunity is determined by dividing the number of OFDM symbols in the data subframe with 256: T /T - 7 MSH -CTRL -LEN DataTxOppSize = f S , 256 (13.1)

where is the ceiling function, Tf is the duration of the 802.16 frame, and Ts is the OFDM symbol duration. However, there may be fewer than 256 transmission opportunities in the data subframe, depending on the frame size and the size of the control subframe. The actual number of transmission opportunities in the data subframe is: T /T - 7 MSH - CTRL - LEN DataTxOppNum = f S . DataTxOppSize (13.2)

2008 by Taylor & Francis Group, LLC

Centralized Scheduling Algorithms for 802.16 Mesh Networks 271

For example, if the frame duration is Tf = 10 ms, the bandwidth of the OFDM is 20 MHz (Ts = 12.5 s) and MSHCTRLLEN 10, the size of a data transmission opportunity is DataTxOppSize = 3 and there are DataTxOppNum = 243 transmission opportunities in the data subframe. The standard restricts the number of transmission opportunities in the data sub-frame to at most 256 because the duration fields in the scheduling control packets are eight bits long. The guard symbols and the transmission opportunity size aect the granularity of bandwidth assignments. For example, if transmission opportunities are three OFDM symbols long and each transmission uses three guard symbols, the smallest data packet size at BPSK-1/2 modulation is 36 bytes in two transmission opportunities (one transmission opportunity is the overhead of three OFDM symbols), the next size for a data packet is 72 bytes in three transmission opportunities and then 108 bytes in four transmission opportunities. If the frame size is Tf = 10 ms, this granularity implies that bandwidth should be assigned in the increments of 28.8 kbps (36 bits = 288 bits).

13.2.3

Assignment of Transmission Opportunities

IEEE 802.16 specifies two dierent protocols used to negotiate assignment of transmission opportunities in the data subframe: centralized and decentralized scheduling protocols. The centralized scheduling protocol assigns transmission opportunities in the first MSHCSCHDATAFRACTION DataTxOppNum transmission opportunities of the data subframe, where MSH-CSCH-DATA-FRACTION is a percentage of data subframe transmission reserved for centralized scheduling, specified by the network operator. The decentralized scheduling protocol assigns transmission opportunities in the last (1 MSHCSCHDATAFRACTION) DataTxOppNum transmission opportunities of the data subframe. In the centralized scheduling protocol, mesh nodes monitor the traffic from their wireless terminals and use this information to request end-to-end bandwidth from the BS. The BS uses the end-to-end bandwidth requests from the mesh nodes and assigns link bandwidths to all links in the network. The assignment of link bandwidths creates a routing tree in the mesh. After assigning link bandwidths, the BS multicasts the assignment through the mesh. The mesh nodes use the link bandwidth assignment to determine all starting times and link durations in the frame (a common, global, schedule). We describe the centralized scheduling protocol in detail in the next section. In the decentralized scheduling protocol, neighboring mesh nodes negotiate local schedules. First, a node wishing to change its allocation of transmission opportunities sends a request for transmission opportunities to its neighbors. One or more of the neighbors correspond with a range of available transmission opportunities. The node chooses a subrange of these transmission opportunities and confirms that it will use them with a third message. The 802.16 standard does not specify

2008 by Taylor & Francis Group, LLC

272

WiMAX/MobileFi: Advanced Research and Technology

any algorithm for the distributed scheduling protocol. We have proposed a distributed scheduling algorithm [8] that can be adapted for this purpose. Our algorithm uses a distributed Bellman-Ford algorithm to iteratively find the TDMA schedule from link bandwidths. The algorithm determines TDMA schedules with a partial knowledge of the network topology, available from 802.16 neighbor tables. The centralized and distributed scheduling give rise to two dierent QoS levels in the mesh network. Links scheduled with the centralized scheduling protocol have guaranteed bandwidth, granted by the BS and known throughout the network. The hop-by-hop bandwidth guarantee in the centralized scheduling routing tree allows end-to-end QoS guarantees for the traffic ows traversing the tree. On the other hand, links scheduled with the decentralized scheduling protocol have transient behavior and their bandwidth depends on the grants from the nodes neighbors. The uncertainty in link bandwidth translates to the best-eort QoS for end-to-end connections using the links scheduled with the decentralized scheduling protocol. We have shown how to take advantage of the QoS provided by 802.16 with network layer solutions in [2]. In that work, we propose a convergence sublayer, connecting the network layer, and the 802.16 MAC layer, that allows the Internet Protocol (IP) Di Serv architecture [9] to take advantage of the two 802.16 QoS service levels. Finding the minimum number of centralized scheduling transmission opportunities in the data subframe is an important provisioning question for 802.16 mesh networks. MSH-CSCH-DATA-FRACTION should be minimized so that as much bandwidth as possible be available for best-eort traffic and enough bandwidth can be allocated for the services requiring guaranteed bandwidth. We have proposed an algorithm that minimizes the number of slots needed to schedule the traffic in the centralized scheduling part of the data frame [10]. The algorithm finds the smallest value of MSHCSCHDATAFRACTION that supports the link bandwidths, subject to the limit on the maximum TDMA delay in the network. TDMA delay occurs when an outgoing link on a mesh node is scheduled to transmit before an incoming link in the path of a packet [11].

13.3 802.16 Centralized Scheduling

In this section, we discuss the 802.16 scheduling protocol in detail. First, we describe the 802.16 mesh control packets of the BS and the mesh nodes use to negotiate endto-end bandwidths. Then, we propose a framework for bandwidth assignment. We compare several scheduling algorithms with each other in the next section.

13.3.1 802.16 Centralized Scheduling Protocol

The centralized scheduling protocol uses MSH-CSCH and MSH-CSCF mesh control packets. The BS uses MSH-CSCF packets to distribute the network topology

2008 by Taylor & Francis Group, LLC

Centralized Scheduling Algorithms for 802.16 Mesh Networks 273

and MSH-CSCH packets to assign end-to-end bandwidth. Mesh nodes use MSHCSCH packets to request end-to-end bandwidth. Centralized scheduling MSH-CSCH and MSH-CSCF packets are transmitted in the control subframe [1,2]. Scheduling of MSH-CSCH and MSH-CSCF control packets is performed by following a breadth-first traversal of a globally known tree topology. The BS starts by broadcasting the global tree topology with MSH-CSCF packets. When a node receives a MSH-CSCF message, it learns the multicast routing tree, as well as which node is broadcasting the MSH-CSCF packet. Given this information, the node calculates when it should transmit its MSH-CSCF packet. After the last node in the tree receives the MSH-CSCF packet from its parent, all nodes know the network topology, as well as the transmission schedule for MSHCSCH and MSH-CSCF packets. If the topology changes, subsequent MSH-CSCF messages notify the mesh nodes of the changes. The nodes request bandwidth from the BS by sending MSH-CSCH messages to their parents in the tree. After all of the requests reach the BS, the BS uses them to calculate the bandwidth for each link in the network and multicasts granted endto-end bandwidths to mesh nodes with new MSH-CSCH messages. If by changing the link bandwidths, the BS also changes the routing tree for the network, it multicasts routing changes with MSH-CSCF messages before it multicasts MSH-CSCH messages. When a mesh node receives a MSH-CSCH message, it uses the routing tree, known from prior MSH-CSCF messages, and the assignment of end-to-end bandwidths to find the assignment of link bandwidths for the entire network. Each mesh node uses the link bandwidth assignment to find a global transmission schedule. The new schedule takes place in the first frame after the last node in the tree receives the MSH-CSCH message from its parent. The 802.16 centralized scheduling protocol forces three requirements on centralized scheduling algorithms. First, the assignment of link bandwidths should result in a tree whose root node is the BS. Second, link bandwidths should be assigned so that end-to-end bandwidth requests are satisfied. If the end-to-end bandwidth requests are not fully satisfied, the assignment of end-to-end bandwidths should be fair. Third, the number of transmissions for each link should be limited to one per frame as the overhead of each transmission is large. The breakdown of algorithm requirements gives a convenient way to split the scheduling problem. First, the BS decides on a tree of links that should carry the traffic. Then, the BS assigns link bandwidths using the routing tree and the end-toend bandwidth requests. The BS multicasts the routing tree with MSH-CSCF messages and end-to-end bandwidth grants with MSH-CSCH messages. After receiving the latest tree topology and end-to-end bandwidth grants, mesh nodes find link bandwidth assignments for the entire network. The BS and the mesh nodes use the same algorithm to assign link bandwidths from end-to-end bandwidths. Finally, each mesh node uses a scheduling algorithm to find a global transmission schedule from the link bandwidths and the network topology (the routing tree). The scheduling algorithm decreases link bandwidths, if necessary, to obtain a valid schedule.

2008 by Taylor & Francis Group, LLC

274

WiMAX/MobileFi: Advanced Research and Technology

Next, we give a simple method for the centralized assignment of link bandwidths. The BS uses this method to find link bandwidths from a routing tree and a set of mesh node bandwidth requests. The mesh nodes use this method to find link bandwidths from the routing tree and the set of mesh node bandwidth grants received in the MSH-CSCH messages. We review several scheduling algorithms in the next section.

13.3.2 Centralized Bandwidth Assignment

IEEE 802.16 standard uses an example to describe how to find link bandwidths from end-to-end bandwidths. In this section, we propose a formal method to find link bandwidth assignments, after the BS receives end-to-end bandwidth requests. We also provide a reverse of the method to find end-to-end bandwidths given a 802.16 TDMA schedule. We use the reverse method in the next section to compare the performance of 802.16 scheduling algorithms. We model the mesh network with a topology graph connecting the nodes in the wireless range of each other. We assume that if the two nodes are in the range of each other, they establish mutual links in the MAC layer, so the mesh can be represented with a connectivity graph G(V, E, f t ), where V = {v1,,vn} is the set of nodes, E = {e1,,em} are directional links between neighboring nodes, and f t: E V V assigns links to pairs of nodes. The connectivity map f t enforces the fact that all links are directional, so for a link ek E, f t(ek ) = (vi, vj ) means that node vi uses link ek to transmit data to node vj. We model channel quality with link bit-rates. Bit-rates depend on the modulation, which depends on signal-to-noise ratio (SNR) for the link. The SNR is divided into several discrete levels and each is associated with its maximum bit-rate. We define link bit-rate as the number of bits transmitted in an OFDM symbol, which is represented with b: E {96, 192, , 864}, where 96 is the number of bits carried in a symbol with the BPSK-1/2 modulation and 864 is the number of bits carried in a symbol with 64 QAM-3/4 (Table 13.1). The routing algorithm chooses a subset of links, forming two spanning trees in the network. The first tree is associated with the uplink traffic and it consists of the links carrying the uplink traffic (to the BS). The second tree is associated with the downlink traffic and contains links carrying downlink traffic (to the mesh nodes). Each mesh node is connected to the BS with two, acyclic, directed paths. The first path is the uplink path directed from the mesh node to the base station and the second path is the downlink path directed from the BS to the mesh node. The uplink up , . . . , P up and the downlink paths are represented with separate sets: P up = {P 2 n } and down up Pdown = {P 2 , , P n } where there are (n 1) paths onthe uplink, one for each node, and (n 1) paths on the downlink, also one for each node. We use the convention that v1 is the BS. There are no paths to the BS as the traffic from its wireless terminals does not traverse the wireless mesh. Each path is a set of links that the packets use to traverse the mesh, so Piup P up, Piup E and Pidown P down, Pidown E.

2008 by Taylor & Francis Group, LLC

Centralized Scheduling Algorithms for 802.16 Mesh Networks 275

Figure 13.3 Simple example of a mesh network.

We represent the node uplink and downlink demands, in bits per second, with g up: V R and g down: V R, respectively.* Because each node corresponds to an uplink and a downlink path, the uplink and downlink requests of each node are associated with the uplink and downlink paths for the node. The required bandwidth on each link, r : E R, can be found by adding up the traffic on each path traversing the link:
n n up i

rj =

g
i =2

I (e j Pi ) +

up

g
i =2

down i

I (e j Pi down ),

(13.3)

where I() is the indicator function, which is 1 when its argument is true and 0 when its argument is false. Because links are unidirectional, uplink and downlink trees are disjoint, hence links never carry both the uplink and the downlink traffic. This means that one of the sums in Equation 13.3 is always zero. For example, for the topology in Figure 13.3, up + g up + g up + g up + g up and r = g down + g down + g down + g down + g down. r4 = g 3 4 5 6 7 3 3 4 5 6 7 Given the required rates on each link, the number of OFDM symbols the link uses in the frame is: rjT f /b j + h dj = DataTxOppSize, DataTxOppSize

(13.4)

* Although we are using the notation that bandwidth demands real numbers, in reality the bandwidths are discrete as they must be rounded to fit into 802.16 MSH-CSCF packets.

2008 by Taylor & Francis Group, LLC

276

WiMAX/MobileFi: Advanced Research and Technology

where h is the number of overhead slots (h = 2 or h = 3), bj is the number of bits in each OFDM symbol, Tf is the frame duration and the formula makes sure the number of slots is divisible by DataTxOppSize. The number of transmission opportunities required by the link in each frame is dj/DataTxOppsize. Ideally, a centralized scheduling algorithm assigns the number of OFDM symbols to each link corresponding to Equation 13.4. However, the scheduling algorithms we review in the following section do not always assign the exact number of required slots to the links. In such cases, it is useful to know the actual end-to-end bandwidth assigned to each mesh node given a specific schedule. We find the endto-end rates in two steps. First, we find the actual rate assigned to link e j, given the number of slots assigned to the link, dj: bj r j = (dj h) __ Tf (13.5)

Second, we use the actual link bandwidth, to find the achievable end-to-end rates. We assume that the nodes use a mechanism such as weighted fair queueing (WFQ) [12]. With WFQ, all end-to-end connections can get a proportional share of bandwidth on the links they traverse. For example, if the end-to-end ows are weighted according to their requested bandwidth, the share of uplink bandwidth that node r j /rj as rj is the total bandwidth of all connections traversing vi gets on link ej is giup the link, Equation 13.3. Similarly, the share of downlink bandwidth that node vi r j /rj. The achievable end-to-end bandwidth is found by congets on link ej is gidown sidering the minimum bandwidth the end-to-end connection gets on all the links on the path:
r j up __ min g g up i = up r i
ejPi
j

(13.6)

and rj down __ = min gi g down i down r

ejPi
j

(13.7)

The reason for taking the minimum over all links in the path is to ensure that all queue lengths on the path are bounded.

13.4 Centralized Scheduling Algorithms

In this section, we compare three dierent 802.16 centralized scheduling algorithms with each other. The first algorithm is suggested in the 802.16 standard [1] and does not use spatial reuse to increase the capacity in the network. We derive the

2008 by Taylor & Francis Group, LLC

Centralized Scheduling Algorithms for 802.16 Mesh Networks 277

second scheduling algorithm from a load-balancing node scheduling algorithm proposed for 802.16 [13]. The new algorithm is equivalent to the original algorithm [13], when the uplink and downlink traffic are scheduled separately. However, our adaptation allows scheduling of uplink and downlink traffic at the same time, thus increasing spatial reuse. We propose the third algorithm based on the BellmanFord TDMA scheduling algorithm [11]. Each algorithm takes the number of OFDM slots each link should transmit in the frame (link durations) as the input and produces a transmission schedule. Link durations are readily available from the routing tree and end-to-end node bandwidths (Equations 13.3 and 13.4). The algorithms follow a common procedure to find a transmission schedule. First, each algorithm finds a link ranking, R: E Z. The ranking determines the order of transmissions in the frame; links with a lower rank transmit before the links with a higher rank. For example, we obtained the schedule in Figure 13.4 with the 802.16 ranking mechanism, which ranks links according to a breadth-first traversal of the tree. The links in the figure have the following ranking: R1 = R 2 = 0, R 3 = R4 = 1, R5 = R 6 = R12 = R7 = 2, and R 8 = R9 = R10 = R11 = 3. We elaborate on this ranking mechanism later in the section. After an algorithm assigns link rankings, it proceeds to assign transmission opportunities to links. In the case of the 802.16 algorithm, the starting transmission opportunity of each link is the first unused transmission opportunity. In the case of the load-balancing algorithm, the starting transmission opportunity is the first transmission opportunity that allows the link to transmit without overlapping with any of its conicting links. In the case of the Bellman-Ford algorithm, the starting transmission opportunity is based on the links minimum distance in the conict graph. The distance in the conict graph is based on the ranking and the link duration, so links with a lower ranking transmit before the links with a higher ranking. The 802.16 algorithm and the Bellman-Ford algorithm also include a final step in which the link transmission times are scaled down to fit in the centralized scheduling part of the data subframe. In the rest of the section, we use the network topology from Figure 13.3. The arrows in the figure correspond to links connecting nodes in the range of each other. The Modulation on all links in the network is BPSK-1/2. We set the frame duration to ten milliseconds, giving a total of 800 OFDM symbols in each frame and MSG CTRLLEN =10, making the number of OFDM symbols in the data sub-frame 730 (there are 7 MSH CTRLLEN = 70 OFDM symbols in the control subframe). With these network parameters, the number of OFDM symbols in each transmission opportunity is DataTxOppSize = 3 and there are a total of DataTxOppNum = 243 transmission opportunities in the data subframe. We set MSHCSCHDATAFRACTION = 80 percent, leaving 194 centralized scheduling transmission opportunities in every frame. We assign link bandwidths so that uplink and downlink requests of each mesh node are 245 kbps. To satisfy the end-to-end connection bandwidths, link rates should be r1 = r 2 = 1470 kbps, r 3 = r4 = 1225 kbps, r5 = r 7 = 490 kbps, and

2008 by Taylor & Francis Group, LLC

278

WiMAX/MobileFi: Advanced Research and Technology

r 6 = r 8 = r 9 = r10 = r11 = r12 = 245 kbps. Using the 802.16 network parameters, we find the total duration required by each link with Equation 13.4 as d1 = d 2 = 156, d3 = d 4 = 129, d5 = d7 = 54, and d 6 = d 8 = d9 = d10 = d11 = d12 = 27 OFDM symbols. Next we give the details of how each scheduling algorithm works and we also make observations about algorithm performance. The numerical results from our simulations are summarized in the tables that follow.

13.4.1

802.16 Algorithm

The 802.16 scheduling algorithm finds a link ranking during a breadth-first traversal of the routing tree. The first visited link, in the traversal of the tree, is assigned the lowest rank. The link traversed next is assigned a higher rank and so on, until all links are assigned a ranking. After ranking the links, the algorithm assigns transmission opportunities. The link with the lowest rank is assigned transmission opportunities at the beginning of the data subframe. The link with the next highest rank is assigned the subsequent transmission opportunities and so on, until all links are scheduled. In case that two links have the same rank, the link with the smaller identifier transmits first. If the total number of assigned transmission opportunities is larger than the number of transmission opportunities reserved for centralized scheduling, the algorithm scales down link durations. After the scaling, the number of transmission opportunities assigned to link ej is:
dj DataTxOppNum MSH CSCH DATA FRACTION ______________________________________ = _______________ D j DataTxOppSize N
g

(13.8)

where dj /DataTxOppSize is the original allocation of transmission opportunities for the link, MSHCSCHDATAFRACTION DataTxOppNum is the number of centralized scheduling transmission opportunities and Ng is the total number of transmission opportunities needed by the schedule. Because the resulting number of transmission opportunities is rational, the 802.16 standard specifies how to round the duration times. First, the links are ordered by the increasing size of the rational part of their scaled duration. Second, for pairs of links, one at the beginning of the ordered list and the other at the end of the ordered list, the algorithm rounds down the link at the beginning of the list and rounds up the link at the end of the list. The algorithm proceeds inward, towards the middle of the list, until all durations are integers. We show the schedule obtained with the 802.16 breadth-first algorithm in Figure 13.4. The schedule corresponds to the network scenario in Figure 13.3. The 802.16 algorithm schedules the links one after another according to their ranking. The algorithm does not use spatial reuse so the total length of the resulting schedule exceeds the number of slots available for centralized scheduling. The links were scaled by about 35 percent so that the schedule can fit in the frame. Tables 13.2

2008 by Taylor & Francis Group, LLC

Centralized Scheduling Algorithms for 802.16 Mesh Networks 279

e1 e2 e3 e4 e5 e6 e7 e8 e9 e10 e11 e12 0 100 200 Control Sub-frame 300 400 500 Centralized Scheduling 600 700 800 Decentralized Scheduling

Figure 13.4 Schedule obtained with the 802.16 scheduling algorithm and breadth-first traversal of the routing tree. (From IEEE standard for local and metropolitan area networks. Part 16; air interface for fixed broadband wireless access systems, 2004.)

and Table 13.3 show the end-to-end bandwidths resulting from this schedule. We observe that the scaling algorithm preserves proportional fairness between end-toend ows. This is a consequence of the linear relationship between link bandwidths and end-to-end bandwidths (Equation 13.3).
Table 13.2 End-to-End Bandwidth Comparison (Uplink)
Bandwidth (kbps) Algorithm 802.16 (breadth first) 802.16 (minimum delay) Load balancing, one iteration Load balancing, two iterations Bandwidth optimal Bandwidth optimal, Dmax = 10 ms Bellman-Ford (minimum delay) v2 153.60 153.60 244.80 244.80 244.80 163.20 163.20 v3 153.60 153.60 244.80 155.52 244.80 161.28 161.28 v4 153.60 153.60 240.00 155.52 240.00 161.28 161.28 v5 144.00 144.00 230.40 155.52 230.40 144.00 144.00 v6 144.00 144.00 230.40 155.52 230.40 144.00 144.00 v7 144.00 144.00 28.80 155.52 230.40 144.00 144.00

2008 by Taylor & Francis Group, LLC

280

WiMAX/MobileFi: Advanced Research and Technology

Table 13.3

End-to-End Bandwidth Comparison (Downlink)

Bandwidth (kbps)

Algorithm 802.16 (breadth f irst) 802.16 (minimum delay) Load balancing, one iteration Load balancing, two iterations Bandwidth optimal Bandwidth optimal, Dmax = 10 ms Bellman-Ford (minimum delay)

v2 153.60 153.60 244.80 244.80 244.80 163.20 163.20

v3 153.60 153.60 244.80 1.92 244.80 161.28 161.28

v4 153.60 153.60 240.00 1.92 240.00 161.28 161.28

v5 115.20 115.20 230.40 1.92 230.40 144.00 144.00

v6 144.00 144.00 230.40 1.92 230.40 144.00 144.00

v7 115.20 115.20 230.40 1.92 230.40 144.00 144.00

The breadth-first ranking does not result in the best transmission order. For example, the return path TDMA delay between v1 and v4 for the schedule in Figure 13.4 is approximately 30 ms (Table 13.4). The TDMA delay occurs when an outgoing link on a mesh node is scheduled to transmit before an incoming link in the path of a packet [11]. For the schedule in Figure 13.4, the delay of approximately 10 ms occurs at nodes v2, v3, and v4.
Table 13.4
Algorithm 802.16 (breadth first) 802.16 (minimum delay) Load balancing, one iteration Load balancing, two iterations Bandwidth optimal Bandwidth optimal, Dmax = 10 ms Bellman-Ford (minimum delay)

Return Trip Propagation Delay

Minimum Delay (msec) 8.76 8.76 8.05 8.05 8.05 8.69 8.69 Maximum Delay (msec) 28.76 8.76 28.05 28.05 18.05 8.69 8.69

2008 by Taylor & Francis Group, LLC

Centralized Scheduling Algorithms for 802.16 Mesh Networks 281

We decrease the TDMA delay resulting from the 802.16 ranking procedure with the minimum TDMA delay ranking algorithm [11]. The minimum delay ranking algorithm ranks the links in a way that guarantees that the total return path TDMA delay is one frame. Initially, the algorithm sets the rank of all the links to zero. The algorithm then examines each return path to the BS, link-bylink, and assigns a rank to each link as a function of the distance from the root of the routing tree. For links in a return path P = {ei,, ek, el ,...ej}, where ei is the starting link at the BS and ej is the terminating link at the BS, the rank is assigned as follows: Rk = max {Rk, Rl + 1}, ek,el P : el ek , (13.9)

ek means that ek follows el in the path. For the example in Figure 13.3, we where el have R1 = 0, R 3 = 1, R5 = R 6 = 2, R 8 = R9 = R12 = 3, R10 = R11 = 4, R7 = 5, R4 = 6, and R 2 = 7. We show the schedule obtained with the 802.16 scheduling algorithm and the minimum delay ranking in Figure 13.5. The total length of the schedule obtained in this way is the same as the total length of the schedule obtained with the 802.16 algorithm, so end-to-end bandwidths resulting from either algorithm are the same (Tables 13.2 and 13.3). The advantage of using the new ordering is that the maximum return path delay is decreased (Table 13.4). The ranking forces the links to transmit in the order that results in TDMA delay on any path of less than ten milliseconds. If we follow links in the topology graph (Fig. 13.3) in the order of transmissions in Figure 13.5, we see that every node in the tree is visited twice before the end of the frame. This is not the case with the schedule in Figure 13.4, where visiting all the nodes twice requires three frames in some cases.

e1 e2 e3 e4 e5 e6 e7 e8 e9 e10 e11 e12 0 100 200 300 400 500 600 700 800 Control Sub-frame Centralized Scheduling Decentralized Scheduling

Figure 13.5 Schedule obtained with the 802.16 scheduling algorithm and minimum delay ranking.

2008 by Taylor & Francis Group, LLC

282

WiMAX/MobileFi: Advanced Research and Technology

13.4.2

Load-Balancing Algorithm

We use the 802.16 node load-balancing algorithm [13] for link scheduling. The link scheduling version of the algorithm uses the same procedure as the node scheduling algorithm, but it schedules links instead of nodes. When the uplink and downlink traffic are scheduled on their own, nonoverlaping, parts of the data subframe, node scheduling is equivalent to link scheduling. For the example in Figure 13.3, the downlink schedule for nodes v2, v3, v4, v5, v6, v 7 corresponds to the schedule for links e1, e3, e5, e6, e 8, e 9. For the uplink, node schedule for v2, v3, v4, v5, v6, v 7 corresponds to the link schedule for e 2, e4, e 7, e10, e11, e12. However, the link scheduling algorithm increases spatial reuse in the network, as it allows the uplink and downlink to be scheduled simultaneously. The load-balancing algorithm works in iterations. At the begining of each iteration, the algorithm calculates the link ranks according to the link satisfaction with the schedule from the previous iteration. The satisfaction is directly proportional to the ratio of the link bandwidth achieved with the schedule and the bandwidth assigned to the link by the BS. For link ej, the satisfaction is:
rj sj = __ rj ,

(13.10)

where r j is obtained from the previous schedule with Equation 13.5 and rj is the required bandwidth on the link. The satisfaction index determines the ranking of the links; links with the lowest satisfaction index have the highest rank and the links with the highest satisfaction index have the lowest rank. If two links have the same satisfaction, the link with the higher link identifier is assigned the higher rank. After establishing the ranking, the load-balancing algorithm schedules the links. First, the algorithm schedules the link with the smallest rank at the beginning of the data subframe. Then, the algorithm schedules the rest of the links in the order of their rank. Each link is scheduled as close to the beginning of the data subframe as possible, so that its transmission does not overlap with transmissions of its conicting links. Scheduling links in this way takes advantage of spatial reuse in the network, however links scheduled toward the end of the frame may get a smaller number of OFDM slots than what they were assigned by the BS. In this case, their transmissions are truncated. The links that were truncated get more OFDM symbols in the next iteration of the load-balancing algorithm. In the next iteration, the algorithm recalculates the satisfaction indexes of the links. The links that were truncated in the previous schedule (previous iteration of the algorithm) have the lowest index, so in the next iteration of the algorithm they are scheduled toward the begining of the frame. Scheduling the links sooner in the frame ensures that their transmissions are longer, making their bandwidths higher than in the previous frame. On the other

2008 by Taylor & Francis Group, LLC

Centralized Scheduling Algorithms for 802.16 Mesh Networks 283

e1 e2 e3 e4 e5 e6 e7 e8 e9 e10 e11 e12 0 100 200 300 400 500 600 700 800 Control Sub-frame Centralized Scheduling Decentralized Scheduling

Figure 13.6 Schedule obtained with the load-balancing algorithm, one iteration.

hand, links that were not truncated in the previous iteration have the highest satisfaction index, so they are scheduled toward the end of the frame where they may be truncated. We show a schedule obtained in the first iteration of the load-balancing algorithm in Figure 13.6. In the first iteration, the satisfaction of all links is the same as they start with the r j = 0, ej E. In this case, the link rank corresponds to link index, so link e12 is scheduled last and its transmission is truncated. Tables 13.2 and 13.3 show the end-to-end bandwidths resulting from this schedule. We see that the uplink bandwidth of node v 7 is lower than its assigned bandwidth as e12 is its bottleneck link. The load-balancing algorithm increases the bandwidth of link e12 in the second iteration. At the beginning of the second iteration, link e12 has the lowest satisfaction, so it is scheduled first in the frame (with no loss of bandwidth). Link e 3 has the highest satisfaction index, so it is scheduled last and is truncated. We show the schedule after the second iteration in Figure 13.7. Tables 13.2 and 13.3 show the end-to-end bandwidths resulting from this schedule. We see that the downlink bandwidths of nodes v3, v4, v5, v6, and v7 are significantly decreased as e3 is their bottleneck link. In the next iteration, the algorithm should increase the bandwidth on e3, so this is only a temporary loss of bandwidth for the end-to-end connections traversing e3. The load balancing does not take into account delay. The maximum return path delay is approximately 30 ms after both iterations (Table 13.4).

13.4.3

Bellman-Ford Algorithm

The Bellman-Ford scheduling algorithm is based on our previous work [11]. The algorithm has two steps. First, the algorithm finds a ranking with good TDMA

2008 by Taylor & Francis Group, LLC

284

WiMAX/MobileFi: Advanced Research and Technology

e1 e2 e3 e4 e5 e6 e7 e8 e9 e10 e11 e12 0 100 200 300 400 500 600 700 800 Control Sub-frame Centralized Scheduling Decentralized Scheduling

Figure 13.7 Schedule obtained with the load-balancing algorithm, two iterations.

delay properties. Finding a ranking with good TDMA delay properties is a computationally hard problem [11]. Nevertheless, we propose a branch-and-bound search method for bandwidth optimal transmission orders, as well as heuristics for good rankings [11]. A ranking is bandwidth optimal if link bandwidths resulting from its schedule cannot be increased with a schedule from another ranking. Second, the algorithm finds a transmission schedule with the Bellman-Ford algorithm. The link transmission times are calculated from the minimum distances in the conict graph for the network. The conict graph has links as vertexes and conicts between the links as arcs and links conict if they cannot transmit at the same time. As the schedule is based on the distances in the conict graph, the conicting link transmissions do not overlap [11]. Figure 13.8 shows a schedule resulting from a bandwidth optimal ranking with no limit on the return path TDMA delay. We use the branch-and-bound method to find the bandwidth optimal ranking [11]. We show the end-to-end bandwidths resulting from the schedule in Tables 13.2 and 13.3. We observe that the end-toend bandwidths resulting from this schedule closely match the required end-to-end bandwidths. The final end-to-end bandwidths are dierent from the required bandwidths as links are allocated OFDM symbols in multiples of DataTxOppSize, which dictates the minimum granulation of bandwidth assignments. As there is no limit on the return path TDMA delay, this schedule causes a maximum return path TDMA delay of approximately 20 ms (Table 13.4). Figure 13.9 shows a schedule resulting from a bandwidth optimal ranking, when the maximum return path TDMA delay is limited to ten milliseconds. We use the branch-and-bound method to find this ranking as well. This schedule is dierent from a schedule that can be obtained by fixing the ranking to the minimum delay ranking with Equation 13.9; it is bandwidth optimal. Tables 13.2 and 13.3 show end-to-end bandwidths resulting from the schedule. We note that end-to-end

2008 by Taylor & Francis Group, LLC

Centralized Scheduling Algorithms for 802.16 Mesh Networks 285

e1 e2 e3 e4 e5 e6 e7 e8 e9 e10 e11 e12 0 100 200 300 400 500 600 700 800 Control Sub-frame Centralized Scheduling Decentralized Scheduling

Figure 13.8

Bandwidth optimal schedule.

bandwidths are significantly lower than the end-to-end bandwidth resulting from the bandwidth optimal schedule with no delay restrictions (30 percent lower), however the bandwidths are also higher than the end-to-end bandwidths resulting from the 802.16 schedule (8 percent higher). The advantage of the schedule in Figure 13.9 is that it limits the return path delay to ten milliseconds, which is half of the maximum delay resulting from the bandwidth optimal schedule with no delay restrictions. Finally, Figure 13.10 shows the schedule resulting from the minimum delay ranking obtained with Equation 13.9. The end-to-end bandwidths resulting from this schedule are the same as the end-to-end bandwidths obtained with the bandwidth optimal schedule with the delay limit (Tables 13.2 and 13.3), showing that the simple heuristic is eective.

e1 e2 e3 e4 e5 e6 e7 e8 e9 e10 e11 e12 0 100 200 300 400 500 600 700 800 Control Sub-frame Centralized Scheduling Decentralized Scheduling

Figure 13.9

Bandwidth optimal schedule, delay limited to Dmax = 10.0 msec.

2008 by Taylor & Francis Group, LLC

286

WiMAX/MobileFi: Advanced Research and Technology

e1 e2 e3 e4 e5 e6 e7 e8 e9 e10 e11 e12 0 100 200 300 400 500 600 700 800 Control Sub-frame Centralized Scheduling Decentralized Scheduling

Figure 13.10

Bellman-Ford algorithm, fixed minimum delay ranking.

13.4.4

Other Approaches

TDMA scheduling has been a research topic for some time, however most of that research is not appropriate for the 802.16 mesh protocol. We identify two main reasons for the disconnect between the previous research in TDMA scheduling protocols and 802.16. First, many of the previous approaches do not model all conicts that exist in TDMA wireless networks [1416]. In these works, the authors assume that only the links sharing a node in common conict (primary conicts). For example, in Figure 13.3 this conict model ignores conicts such as the one between the links e1 and e5. It is possible to resolve secondary conicts by assigning orthogonal frequencies to links in secondary conicts. However, finding this frequency assignment is hard [17,18]. Also, the 802.16 standard does not completely resolve the use of multiple frequencies in 802.16, meaning that any solution using multiple frequencies requires new additions to the standard. Second, many of the algorithms in current literature do not take transmission overhead into account. In the coloring approaches [1416], the scheduling algorithm assigns colors to transmission opportunities in the frame before it begins to color the links. After the coloring, each link has a set of colors dierent from the colors used by the links it conicts with. Because colors correspond to transmission opportunities, the resulting assignment of transmission opportunities is also conictfree. However, this approach allows links to transmit multiple times in the frame. For example, if a link assigned two colors for transmissions and these colors correspond to nonconsecutive transmission opportunities, the link transmits twice in the frame. The link loses 28.8 kbps of bandwidth with the second transmission because of the three guard OFDM symbols (36 bytes = 288 bits every 10 ms). We now summarize another scheduling algorithm [19] that has the same problem as the coloring algorithms. The algorithm assigns transmission opportunities

2008 by Taylor & Francis Group, LLC

Centralized Scheduling Algorithms for 802.16 Mesh Networks 287

in rounds. In each round, the algorithm assigns one transmission opportunity to links that need it to satisfy their bandwidth requests. Links may not receive any transmission opportunities in the round if the assignment introduces conicts with other links in the round. The reason why we could not implement this scheduling algorithm without significant changes is that we cannot limit the number of times each link transmits in the frame. This can have a very negative impact on the resulting end-to-end bandwidths. For example, the algorithm may schedule a link to transmit all of its data in nonconsecutive data transmissions, thus allocating no bandwidth to the link, if DataTxOppSize 3.

13.5 Conclusion
We have given a summary of the 802.16 mesh protocol, with focus on centralized scheduling algorithms. We have compared three centralized link scheduling algorithms with each other. The first algorithm is proposed in the 802.16 standard [1]. We have obtained the second algorithm from a load-balancing node scheduling algorithm [13]. The third algorithm is based on the Bellman-Ford algorithm and was proposed for TDMA scheduling [11]. The Bellman-Ford algorithm has the best properties. Unlike the 802.16 algorithm, it takes advantage of spatial reuse and unlike the load-balancing algorithm it always allocates bandwidth proportional to assigned link bandwidths. The Bellman-Ford algorithm also takes TDMA delay into account, so it can find link schedules that result in small end-to-end TDMA delay.

References
1. IEEE standard for local and metropolitan area networks. Part 16: air interface for fixed broadband wireless access systems (2004). 2. P. Djukic and S. Valaee. 802.16 mesh networking. In eds. S. Ahson and M. Ilysa, Handbook of Wimax, Chapter 8, pp. 149177, CRC Press, Boca Raton (2007). 3. J. Camp, J. Robinson, C. Steger, and E. Knightly. Measurement driven deployment of a two-tier urban mesh access network. Technical Report TREE0505, Rice University (December 2005). 4. Nortel Network. Wireless mesh networkextending the reach of wireless LAN, securely and cost-eectively. http://www.nortelnetworks.com/solutions/wlan/ (November 2003). 5. S. Xu and T. Saadawi. Does the IEEE 802.11 MAC protocol work well in multihop wireless ad hoc network, IEEE Communication Magazine, 39(6), 130137 (June 2001). 6. P. Djukic and S. Valaee. 802.16 MCF for 802.11a based mesh networks: a case for standards re-use. In 23rd Queens Biennial Symposium on Communications (2006). 7. P. Djukic and S. Valaee. Towards guaranteed QoS in mesh networks: Emulating WiMAX mesh over WiFi hardware. In The Fourth Workshop on Wireless Ad hoc and Sensor Networks WWASN (2007).

2008 by Taylor & Francis Group, LLC

288

WiMAX/MobileFi: Advanced Research and Technology

8. P. Djukic and S. Valaee. Distributed link scheduling for TDMA mesh networks. In Proceedings of ICC (2007). 9. Y. Bernet, P. Ford, R. Yavatkar, F. Baker, L. Zhang, M. Speer, R. Braden, B. Davie, J. Wroclawski, and E. Felstaine. A framework for integrated services operation over Diserv networks. RFC 2998 (Informational) (November 2000). URL http://www. ietf.org/rfc/rfc2998.txt. 10. P. Djukic and S. Valaee. Quality-of-service provisioning in multi-service TDMA mesh networks. In Proceedings of 20th International Teletraffic Congress (2007). 11. P. Djukic and S. Valaee. Link scheduling for minimum delay in spatial re-use TDMA. In INFOCOM (2007). 12. S. Keshav. An Engineering Approach to Computer Networking (Probability, Random Variables and Stochastic Processes, Addison-Wesley, Boston, MA, 1997). 13. D. Kim and A. Ganz. Fair and efficient multihop scheduling algorithm of IEEE 802.16 BWA systems. In Broadnets, pp. 895901 (2005). 14. B. Hajek and G. Sasaki. Link scheduling in polynomial time. IEEE Transactions on Information Theory, 34(5), 910917 (September 1988). 15. T. Salonidis and L. Tassiulas. Distributed dynamic scheduling for end-to-end rate guarantees in wireless ad hoc networks. In MobiHoc, pp. 145156 (2005). 16. M. Kodialam and T. Nandagopal. Characterizing achievable rates in multi-hop wireless mesh networks with orthogonal channels. IEEE/ACM Transactions on Networking, 13(4), 868880 (2005). 17. W. K. Hale, Frequency assignment: theory and applications. Proceedings of the IEEE , 68(12), 14971514 (December 1980). 18. M. Alicherry, R. Ghatia, and L. Li. Joint channel assignment and routing for throughput optimization in multi-radio wireless mesh networks. In Mobicom (2006). 19. H.-Y. Wei, S. Ganguly, R. Izamailov, and Z. Haas. Interference-aware IEEE 802.16 WiMax mesh networks. In VTC Spring05 (2005).

2008 by Taylor & Francis Group, LLC