S09 S02 P04

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the ICC 2008 proceedings.

CaDAR: an Efficient Routing Algorithm for
Wireless-Optical Broadband Access Network
Abu (Sayeem) Reaz1 , Vishwanath Ramamurthi1 , Suman Sarkar1 ,
Dipak Ghosal1 , Sudhir Dixit2 , and Biswanath Mukherjee1
1 University of California, Davis, USA
2 Nokia Research Center, Palo Alto, CA USA

Email: {asreaz,rama,sumsarkar,dghosal,bmukherjee}@ucdavis.edu, sudhir.dixit@nsn.com

Abstract—Hybrid Wireless-Optical Broadband Access Net- I. I NTRODUCTION
work (WOBAN) is a combination of wireless and optical networks
to optimize the cost and performance of an access network.
A Wireless-Optical Broadband Access Network (WOBAN)
Wireless nodes collect traffic from end users and carry them to has a wireless mesh network (WMN) to connect to end users
the optical part of a WOBAN using multiple hops, accumulating and an optical network to carry the aggregated traffic collected
delay at each wireless node. Moreover, the radio capacity on over a WMN [1], [2]. In this way, a WOBAN can have the
each wireless link limits the capacity on each outgoing link from cost-effective deployment of a WMN while having higher
the node in a single-radio Wireless Mesh Network (WMN) of a
WOBAN. Thus, delay and capacity limitation in the WMN of a
performance due to the optical network. At the back end of the
WOBAN is a major bottleneck. We design a capacity and delay network, Optical Line Terminal (OLT) resides in the Central
aware routing scheme, CaDAR, to minimize the delay and increase Office (CO) and is connected via optical fiber to multiple
network support in the WMN of a WOBAN. Our analysis shows Optical Network Units (ONU). At the front end, a set of
that CaDAR is an efficient routing scheme for a single-radio wireless nodes (routers) form a WMN. End users, both mobile
WMN for a WOBAN that can support much higher load and
has lower system delay than other approaches because of better
and stationary, connect to the network through these nodes,
load balanced routing. whose locations are fixed. A selected set of these nodes, called
Keywords: Wireless-optical broadband access network, gateways, are connected to the optical part of the network.
routing, delay, capacity assignment Usually, gateways are attached with of one of the ONUs [1].
Figure 1 shows the architecture of a WOBAN.

Fig. 1. Architecture of a WOBAN.

978-1-4244-2075-9/08/$25.00 ©2008 IEEE


An end user sends packets to a nearby wireless node called CaDAR, which exploits the work by Fratta et al. [10]
of the WOBAN for upstream communication. These packets and Kleinrock [11] on capacity assignment (CA) and flow
travel through the WMN, possibly over multiple hops, and assignment (FA), which were originally designed for general
reach the OLT via the gateways. (From the OLT, the packets packet networks using optimal capacity assignment and flow
can be routed to the rest of the Internet.) Similarly, for deviation on links to minimize delay. The combined problem
downstream communication, packets travel from the OLT of capacity and flow assignment (CFA), however, has an
through the WMN to the end users. As a packet may need optimal solution for a specific input flow [11] and can be
to travel several hops through the WMN, delay in the access considered as ‘local optimal’.
network can be significant. Moreover, some links in the WMN Our objective in this work is to design and investigate the
may be over-utilized and others may be under-utilized under properties of an efficient routing scheme based on capacity
inefficient routing. If the transmission capacity of a wireless assignment and delay on the links of a WOBAN. Our con-
node (i.e., its radio) is not properly distributed among its tributions are: (i) proposing a routing scheme, CaDAR, with
transmission links, the link delays may become high. Thus, capacity assignment and delay awareness; and (ii) comparative
capacity assignment and packet delay are important factors to performance analysis of CaDAR with DARA and CFA.
consider for communication between the wireless nodes. (As
an aside, note that communication between an end user and II. C APACITY AND D ELAY AWARE ROUTING
its nearby wireless node takes a single hop which is always CaDAR is a routing algorithm for the WMN of a WOBAN that
there, so it is ignored in our problem formulation.) We propose minimizes the network delay by assigning the radio capacities
a Capacity and Delay Aware Routing (CaDAR) algorithm that and link weights based on link states. Two wireless nodes
routes packets in the WMN to reduce the WMN-wide average have a link between them if their distance is less than their
packet delay in the WMN using optimal capacity assignment respective transmission range. Each node has one radio; so a
on the links and efficient routing. node’s limited radio capacity needs to be distributed among
In [3], the authors propose a link-activation framework for its outgoing links. Each node advertises the states of all of its
scheduling packets in a wireless network. They show that the outgoing links using Link-State Advertisement (LSA). Based
packet delay for wireline schedulers, Weighted Fair Queu- on LSA, capacity is assigned to the links and shortest-delay
ing (WFQ) and Coordinated Earliest Deadline First (CEDF), paths between the wireless nodes and gateways are calculated.
when implemented over the wireless multi-hop network, are
guaranteed to achieve approximately twice the delay of the A. Capacity Assignment in CaDAR
corresponding wireline topology. In [4], the authors provide Time-Division Multiple Access (TDMA) [12] is a popular
admission control schemes for a multi-hop wireless network multiplexing scheme in telecom networks, so it is a good
for packets with QoS requirements. In [5], the authors develop choice for the WMN in a WOBAN for communication between
an on-demand Capacity Aware Routing (CAR) protocol that the wireless nodes. Number of time slots in a TDMA frame
makes use of Bottleneck Link Capacity (BLC) as the link depends on the number of links induced from a wireless node.
metric for wireless networks. CAR discovers paths on a per- There are 3 nodes originated from node u in Fig. 2; so the
flow basis assuming a Carrier-Sense Multiple Access (CSMA) radio at node u operates on 3 time slots. Similarly, node w
based Media Access Control (MAC) protocol [6]. In [7], the has 2 time slots.
authors present a new metric for routing in a multi-radio, Assignment of different time-slot durations for different
multi-hop WMN. The metric assigns weights to individual links induced from a node translates to assignment of different
links based on the Expected Transmission Time (ETT) of capacity on each link. Noting the current flow on each wireless
a packet over the link. Then, the metric chooses a high- link from the LSAs, we determine the fraction of a node’s
throughput path between the source and the destination. An radio capacity that should be assigned to each link originated
interference-aware routing metric, iAWARE, for a multi-radio by itself such that the WMN-wide average delay is minimized.
WMN is proposed in [8]. iAWARE aids in finding paths Sum of capacities on each outgoing link from a wireless node
that are better in terms of reduced inter-flow and intra-flow must be equal to the capacity of the radio at the node. Capacity
interference. assignment is optimal for the flow given by LSA. If a link
Reference [2] proposes and investigate the characteristics does not carry any flow on it, capacity assignment of CaDAR
of “Delay-Aware Routing Algorithm (DARA)” that minimizes automatically treats that link as “non-existent” and the time-
the average packet delay in the wireless front end of a WOBAN. slot duration for that link is assigned as zero for that LSA
Reference [2] shows that DARA achieves better load balancing period.
and less congestion compared to traditional approaches such Figure 2 shows an example of capacity assignment where
as minimum-hop routing algorithm (MHRA) and shortest-path node u has three neighbors v, w, and x. In general, the loads
routing algorithm (SPRA). In addition to minimizing the delay, on node u’s outgoing links, λuv , λuw , and λux could be
DARA also improves the average hop count compared to the different. To minimize delay, radio capacity at node u should
predictive throughput routing algorithm (PTRA), a popular be distributed properly among the links (u, v), (u, w), and
protocol used in several deployments [9] for the wireless (u, x) based on their respective flows. This can be achieved
front end of a WOBAN. We design an improved algorithm, through assigning different time-slot durations for the different


uw the packet is injected into a WOBAN. The delay from the
uv
w end user to the wireless node is unavoidable and relatively
u wu small compared to the delay incurred in the mesh because a
v packet usually travels through several hops in the WMN before
ux wx reaching a gateway.
The packet delay in the WMN part of a WOBAN has four
contributors:
x 1) Transmission delay: As described in Sec. II-A, each
link can have different capacity. Transmission delay
Fig. 2. Wireless links and their flows (not all links are shown). depends on the capacity on each link. If the capacity
of the link is higher, the transmission delay is lower.
1 1
Transmission delay on a link is µCuv where µ is the
links based on their flows. Hence, capacities on these links, average packet size.
Cuv , Cuw , and Cux need to be assigned based on λuv , λuw , 2) Slot synchronization delay: Slot synchronization delay
and λux . Similarly, for the radio at node w, Cwu and Cwx is associated with TDMA-based operation of a wireless
depend on λwu and λwx . Note that Cuw is assigned from the channel. Each router needs to send a packet to its
radio at node u and Cwu is assigned from the radio at node neighbor in its designated time slot. This delay occurs
w. Moreover, because upstream and downstream traffic flows because an arriving packet needs to synchronize to the
could be different, λuw = λwu . Thus, links (u, w) and (w, u) time-slots for communication with the neighbor of the
are treated differently. wireless node. Average slot synchronization delay in a
Usually, in the WMN, packets are exchanged between 1
TDMA system is 2µCuv .
wireless nodes and the gateways. So, we refer to destinations
as g and sources as s. For any source-destination pair (s, g), 3) Queuing delay: Queuing delay depends on the service
we denote the average traffic intensity between them as γsg . rate and packet arrival rate at a wireless node. Higher
If the number of nodes in the WMN of a WOBAN is N and capacity and lower arrival rate leads to lower delay.
total traffic is γ, then: Because queuing delay is cumulative, when a packet
travels through several wireless nodes before reaching
N N the destination nodes, then at each node it accumulates
γ= γsg ; s = g (1) queuing delay. Delay in a queue can be approximated
s=1 g=1 as µCuv1 uv if the arrivals are independent and packet
−λ
Load on a link is defined as the amount of traffic on a lengths are exponentially distributed.
link. In Fig. 2, if γvw and γuw goes through link (u, w), then 4) Propagation delay: In a WMN, the routers are close to
λuw = γvw + γuw . We denote ω(N ) to be the set of N nodes one another; hence the propagation is delay is negligible.
of the wireless mesh of a WOBAN. Then, system delay, T , can
We approximate the packet arrivals at the wireless nodes
be defined as follows [11]:
to be independent. Noting that the propagation delay on a
wireless link is negligibly small compared to other delay
N N
1 λuv components for typical settings, the average packet transfer
T = ; u, v ∈ ω(N ); u = v, (2)
γ µCuv − λuv delay, or link delay, on any link (u, v), depends on the
u=1 v=1
transmission delay from node u to node v, slot synchronization
1
where µ is the average packet size. delay (for TDMA), and queueing delay at node u.
We want to assign Cuv such that T is minimized for a given So, delay on any link (u, v) is given by:
γ (Eqn. (1)). If ζu is the capacity of wireless node u, Cuv can
be derived as [13]: 1 1 1
duv = + + (4)
µCuv 2µCuv µCuv − λuv
λuv
√
λuv (ζu − v µ ) λuv The total delay for packet transmission between two nodes
Cuv = + √ ; u ∈ ω(N ); ∀v ∈ η(u) equals the sum of the delays on each link in the path between
µ v λuv
(3) the two nodes [2]. We assign the link delays as weights of
(using the approaches in [10] and [11]). corresponding links [2] and compute shortest paths to the
We obtain λuv from LSA messages. We adjust the time-slot gateways from each node and vice versa. This gives shortest
duration τuv for link (u, v) based on Cuv [13]. delay path for each node to the gateways to incorporate delay
awareness in CaDAR. Though this delay awareness can be
B. Delay Awareness in CaDAR extended to the delay on the optical backhaul of a WOBAN
In the WMN of a WOBAN, an end user connects to a (Fig. 1), to determine the shortest-delay path in WOBAN, in
wireless node and start sending packets to the wireless node it this work we focus on delay to the gateways from wireless
is connected to. When the packet reaches the ingress node, nodes and vice-versa.


As shown in Fig. 2, if the shortest delay path from node v to for a specific input flow. The results in [2] show that DARA
w is {v, u, w}, then path delay from v to w can be calculated outperforms routing approaches based on shortest path, min-
from Eqn. (4) as: imum hop, and throughput; hence performance improvement
of CaDAR over DARA is a very significant result.
dvw = dvu + duw (5)
In a multi-hop WMN of a WOBAN, a packet waits in queue
and uses a time slot for transmission at each of its hops. If a
packet has a delay requirement, we calculate the delay from
a wireless node to the gateway, dsg , as shown in Eqn. (5).
If dsg is greater than the minimum delay, then the packet
will not reach the destination within the required time. In that
case, admission control can be incorporated with CaDAR to
save bandwidth for other packets by assigning the time slot to
other packets and the packet will not be injected to WOBAN.
Otherwise, the packet will be routed by CaDAR as described in
Section II-C. Admission control is an open problem for future
study.
C. CaDAR Algorithm
CaDAR operates on current flows on different wireless links
in a WOBAN. Each node sends out periodic LSA messages to
its neighbors, notifying the states of its link. We assign the
Fig. 3. Network topology used in this study.
capacities and the weights on the link for routing. When a
packet arrives at the wireless part of a WOBAN, CaDAR finds Each node has a load, which is the rate of traffic between
the shortest-delay path based on the flows and capacities on the optical network of a WOBAN and that wireless node. The
the links. CaDAR algorithm is described in Algorithm 1. upstream load is two-third of the downstream load (Section
II-A). So, in Fig. 4, load at each node indicates 60% down-
Algorithm 1 CaDAR Algorithm stream and 40% upstream traffic.
1: while true do
2: Receive kth LSA from neighbors 1.6
3: Determine the flow λk on all link (u, v) for kth LSA
uv
4: Assign capacity Cuv based on λk
k
uv
1.4
5: Assign weight Wuv = duv = µC k + 2µC k + µC k 1 k
k 1 1
uv −λuv
uv uv
DARA
1.2
6: For all source s (Wireless Nodes) and destination g CFA
System delay (msec)

CaDAR
(Gateways), calculate the shortest path and dsg using 1
k
Wuv
k
7: Route packet from s to g through WOBAN based on Wuv 0.8
8: end while
0.6

If the link states vary frequently, it means that the traffic 0.4
is changing very rapidly. In that case, it is not feasible to
change the capacity and the link weight based on the flow 0.2
obtained from LSAs; instead estimate the flows on the links,
eλk , by maintaining a weighted moving average (WMA). We
uv
0
0 1 2 3 4 5 6 7 8
can calculate eλk = α × eλk + (1 − α) × λk where α is the
uv uv uv
Load at each node (Mbps)
the decaying index; and change Cuv based on eλk . Similarly,
k
uv
Fig. 4. System delay for different loads at the wireless nodes.
we calculate Wuv with eλk instead of λk to perform routing.
k
uv uv
This is also an open problem for future study.
Figure 4 shows the system delay, which is the total delay
III. I LLUSTRATIVE E XAMPLES over all the links weighted by their respective flows, for
We apply CaDAR on the 25-node wireless network with DARA, CaDAR, and CFA for different traffic loads at the
two gateways as shown in Fig. 3. Each node is equipped with wireless nodes. We observe that CaDAR can support nearly
one radio with a capacity of 54 Mbps, as in IEEE 802.11g three times the maximum load DARA can handle while
[6]. We compare the performance of CaDAR with DARA [2] still maintaining low system delay. We also see that CaDAR
and a static algorithm, CFA [11], which gives optimal solution performs almost as well as the local optimum (CFA) for a


2.3 to decrease the delay which in turn may increase the average
number of hops than CaDAR.
2.2
DARA Figure 6 shows the load balancing of CaDAR, DARA, and
2.1 CFA CFA. We plot the traffic difference, which is the difference
CaDAR
between the maximum and the minimum packet intensities
Avg. number of hops

2 for links in the WMN. Smaller the difference, better will be
the load balancing (or less will be the link congestion) and
1.9
vice versa. Though CaDAR and DARA perform delay-aware
routing, CaDAR has lower link congestion. For a higher load,
1.8
CaDAR has lower flow and capacity on the links than CFA,
1.7 which is reflected in Fig. 6. We conclude that CaDAR performs
better load balancing irrespective of the load.
1.6
IV. C ONCLUSION
1.5 A WOBAN tries to combine the cost effectiveness of a WMN
0 1 2 3 4 5 6 7 8
Load at each node (Mbps) and high capacity of an optical network in the access network.
But a WMN introduces delay and capacity bottleneck. To gain
Fig. 5. Avg. number of hops for different loads at the wireless nodes.
the desirable performance from a WOBAN, we proposed an
efficient routing scheme, called Capacity and Delay Aware
4 Routing (CaDAR), for a WOBAN to support higher load in the
x 10
Difference (max − min) traffic on a single link (msg/sec)

2.5 network and minimize packet delay. CaDAR distributes the
radio capacity of a single-radio wireless node optimally among
DARA its outgoing links and performs delay-aware routing in a
2 CFA WOBAN. Our illustrative examples show that CaDAR performs
CaDAR
very close to the local optimal solution while minimizing
average number of hops in the WMN of a WOBAN.
1.5
R EFERENCES
[1] S. Sarkar, S. Dixit, and B. Mukherjee, “Hybrid wireless-optical broad-
1 band access network (WOBAN): A review of relevant challenges,”
IEEE/OSA Journal of Lightwave Technology, vol. 25, no. 11, Nov. 2007.
[2] S. Sarkar, H. Yen, S. Dixit, and B. Mukherjee, “DARA: Delay-Aware
0.5 Routing Algorithm in a Hybrid Wireless-Optical Broadband Access
Network (WOBAN),” in Proc. IEEE ICC, Glasgow, Scotland, Jun. 2007.
[3] G. Narlikar, G. Wilfong, and L. Zhang, “Designing multihop wireless
backhaul networks with delay guarantees,” in Proc. IEEE INFOCOM,
0 Barcelona, Spain, Apr. 2006.
0 1 2 3 4 5 6 7 8 [4] S. Lee, G. Narlikar, M. Pal, G. Wilfong, and L. Zhang, “Admission
Load at each node (Mbps)
control for multihop wireless backhaul networks with QoS support,”
in Proc. IEEE Wireless Communications and Networking Conference
Fig. 6. Load balancing (or link congestion) for different loads at the wireless (WCNC), Las Vegas, NV, Apr. 2006.
nodes. [5] T. Liu and W. Liao, “Capacity-aware routing with multi-channel multi-
rate wireless mesh networks,” in Proc. IEEE ICC, Istanbul, Turkey, Jun.
2006.
[6] http://standards.ieee.org/getieee802/802.11.html.
[7] R. Draves, J. Padhye, and B. Zill, “Routing in multi-radio, multi-hop
load at each node up to 2 Mbps and follows very closely for wireless mesh networks,” in Proc. ACM MobiCom, Philadelphia, PA,
higher loads till about 6 Mbps. Sep. 2004.
CFA outperforms CaDAR at higher load because CFA does [8] A. Subramanian, M. Buddhikot, and S. Miller, “Interference aware
routing in multi-radio wireless mesh networks,” in Proc. IEEE WiMesh,
flow distribution on the links iteratively while adjusting capac- Reston, VA, Sep. 2006.
ities on the link unless local optimal is reached. On the other [9] Tropos Networks. http://www.tropos.com.
hand, CaDAR performs capacity assignment and does fixed [10] L. Fratta, M. Gerla, and L. Kleinrock, “The flow deviation method:
An approach to store-and-forward communication network design,”
routing (hence, assign flows to links) based on current LSA. Networks, vol. 3, no. 2, pp. 97 – 133, Mar. 2007.
So, at higher loads, CFA readjusts the flows and capacities to [11] L. Kleinrock, Queueing Systems, Volume II: Computer Applications,
obtain better throughput and delay on links than CaDAR. Wiley-Interscience, 1976.
[12] http://en.wikipedia.org/wiki/Time division multiple access.
Figure 5 shows the average number of hops for the three [13] A. S. Reaz, “Capacity and delay aware routing in hybrid wireless-optical
algorithms. As a WMN is the major contributor to contention broadband access networks WOBANs.” PhD Qualifying Examination
Report, UC Davis, Sep. 2007.
in a WOBAN, fewer hops in a WMN translates to lower con-
tention. We observe that CaDAR and DARA have overlapping
average number hops as both of them perform delay-aware
routing. On the other hand, CFA deviates flows on the links

S09 S02 P04

More Related Content

What's hot (20)

Viewers also liked (7)

Similar to S09 S02 P04 (20)

More from Shahab Shahid (16)

S09 S02 P04