Towards Green Broadband Access Networks - IEEE Xplore

Towards Green Broadband Access Networks Pulak Chowdhury, Massimo Tornatore, Suman Sarkar, and Biswanath Mukherjee University of California, Davis, CA 95616 USA Email:{pchowdhury, mtornatore, sumsarkar, bmukherjee}@ucdavis.edu Abstract—Energy consumption in next-generation access networks is rapidly increasing with the increase of bandwidth demands. In this paper, we devise design techniques aimed to build “green” access networks to reduce energy consumption. We apply these techniques on a novel access network paradigm named WirelessOptical Broadband Access Network (WOBAN). We present a Mixed Integer Linear Program (MILP) model to investigate the power consumption in WOBAN over dynamic traffic profiles. We analyze the impact of energy-aware design on the performance of WOBAN. Our results indicate that there are large scopes of power savings if we incorporate energy-aware design and routing in access networks.

I. I NTRODUCTION Access network is the “last mile” of the communication network that connects the telecom Central Office (CO) to the residential and business customers. With the proliferation of the Internet, customer demands for bandwidth-intensive services (such as Video-on-Demand, online Gaming, HDTV, etc.) are rapidly increasing. Therefore, today’s access network should exhibit higher transport capacity. There are several access technologies proposed and deployed in the market – Digital Subscriber Line (xDSL), Cable Modem (CM), Wireless and Cellular networks, fiber-to-the-x (FTTx), Hybrid Wireless-Optical Broadband Access Network (WOBAN), etc. Access technologies such as xDSL, CM, Wireless, and Cellular networks do not live up to satisfying future broadband Internet demands. FTTx technologies can provide higher bandwidth but still remain cost-prohibitive. WOBAN – a novel hybrid access network paradigm with the combination of high-capacity optical backhaul and wireless front-end – can provide higher bandwidth in a cost-effective manner [1]. Performance of access technologies is improving over time, reducing the cost per byte of traffic and making the broadband Internet affordable to more users. This refuels the tremendous growth of the Internet and scales up the size of broadband access networks. For example, let us consider FTTx which has different underlying technologies, such as direct fiber, shared fiber, and the most dominant one – Passive Optical Network (PON) (e.g., EPON, GPON, etc.). Infonetics Research [2] – a leading market research firm – estimates that the worldwide PON equipment market grew 17% during the second quarter of 2008. With the increase of size and bandwidth demand in the Internet, power consumption also increases. An estimation shows that the Internet electricity consumption at US is in billions of dollars [3], and there is significant wastage of electricity (several TWh/year) due to inefficient network and system design [4]. A good portion of this energy is consumed by idle network elements [3]. Therefore, energy can be conserved by reducing the power consumption of these idle network elements. Access network comprises a large part of the Internet. It is also a major energy consumer in the Internet due to the presence of huge number of active elements [5]. If we can reduce energy consumption in the access network, it will automatically provide a significant reduction of overall Internet power consumption. Access network power-consumption reduction not only has the potential of enormous cost savings, this will also lead us to develop environment-friendly technologies, thereby achieving the ultimate goal of “green” Internet [6]. Therefore, future access

networks should be green featuring efficient energy management schemes to reduce “carbon footprint.” Till now, most of the power management approaches in networking research focus on increasing energy efficiency of equipments. There are two other directions of network energy management – energy-aware network design and energy-aware protocol design [3], [7]. In this paper, we develop energy-aware design techniques and routing protocol for “green” WOBAN. Although these techniques are developed for WOBAN, we believe that they will be proven to be more general and can be adapted to other access network technologies like wireless networks and different PON variants. To the best of our knowledge, this is the first paper to devise energy-consumption reduction techniques in hybrid wireless-optical access networks. The remainder of this paper is organized as follows. In Section II, we briefly discuss related work on network energy management. Section III describes the WOBAN architecture, techniques for energy-aware WOBAN design, and energy-aware WOBAN routing protocol. In Section IV, we present a case study to illustrate the effectiveness of our energy-aware design on WOBAN. Section V includes illustrative numerical examples on the case study and gives insights on better power management. Finally, we conclude in Section VI. II. R ELATED W ORK Many research efforts in network energy management focus on reducing system-level power consumption. In [8], the authors advocate for the use of optics to increase router’s scaling capacity and reduce power consumption. Dynamic power management techniques for system components are presented in [9]. There has been significant amount of research efforts for power management in mobile adhoc networks and sensor networks. Mobile adhoc networks along with sensor networks are constrained with limited power capabilities. Hence, these networks and their protocol design should emphasize on power management techniques. Reference [10] provides a nice overview of these techniques at both MAC and network layers. At MAC layer, some proposed protocols focus on algorithms to put the radios of the idle wireless nodes in coordinated sleeping. Other proposals include TDM link scheduling, protocols to minimize unnecessary transmission, etc. [4]. The network layer protocols focus mainly on selecting routes to minimize network energy consumption. Data centers and server clusters consume a great deal of electricity power for searching, cooling, and other purposes. A large number of scholarly papers address the power management issues of data centers and individual servers (for example, [11]). Researchers are also arguing to relocate data centers near the renewable energy sources to build “green” grid technology. In [6], the authors make a proposal for designing “follow-the-sun, follow-the-wind” grids to build green data centers. Authors of [7] advocate power-aware design and configuration of core networks. They use optimization techniques to find minimal configuration of the router chassis for supporting different traffic loads, thereby reducing network power consumption. An initial proposal on the required architectural change of the network to support selective connectivity, a state where a host

978-1-4244-4148-8/09/$25.00 ©2009 This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE "GLOBECOM" 2009 proceedings.

can choose whether to stay connected or disconnected, is given in [12]. Selective connectivity state happens when hosts go to sleep to reduce network power consumption. Methods which enable energy consumption reduction of networked desktop computers while these computers are idle are presented in [3]. Authors of [4] suggest how coordinated and uncoordinated sleeping mechanisms can put the Internet nodes to sleep (energy saving state) with some changes in Internet protocols. Despite all these efforts, there remain significant challenges to deploy energy-efficient schemes over the Internet, especially over the access networks. III. G REEN WOBAN In this section, we present the architecture of WOBAN, a mathematical model of energy-aware WOBAN design, and an energy-aware routing algorithm for WOBAN. A. WOBAN Architecture Hybrid Wireless-Optical Broadband Access Network (WOBAN) is a novel access network architecture with an optimal combination of optical backhaul (e.g., a Passive Optical Network (PON)) and a wireless front-end (e.g., WiFi and/or WiMAX). Figure 1 presents the architecture of WOBAN. WOBAN optimizes the deployment cost due to less-expensive wireless front-end and maximizes the bandwidth performance of a broadband access network [1].

Fig. 1.

WOBAN architecture

In WOBAN (Fig. 1), a PON segment starts from the Optical Line Terminal (OLT) at the telecom CO and ends at multiple Optical Network Units (ONU). Multiple wireless routers form the front-end of WOBAN. A selected set of these routers are called gateways. The front-end of WOBAN is essentially a multi-hop Wireless Mesh Network (WMN) with several wireless routers and a few gateways. These gateways are connected to the PON backhaul through the ONUs. Each ONU can support several wireless gateways. End users (both mobile and stationary) connect to WOBAN through the wireless routers. B. Energy-Aware WOBAN Design WOBAN represents a hierarchical access architecture with gateways as the initial traffic aggregation points. ONUs are the next aggregator level in the hierarchy, while OLT is the highest aggregation point and connects the access network with the metro/core network. As gateways are part of the WOBAN frontend, they can also collect their nearest end-user traffic. There are several important aspects of WOBAN we need to consider for energy-aware design. Current WOBAN design, deployment, and management idealogies provide infrastructure that showcases fault tolerance, reliability, and robustness while

providing high level of performance. Thanks to the mesh frontend, traffic can be rerouted through alternate paths in case of failures such as fiber cut, wireless router or gateway or ONU failures. Moreover, there is a capacity mismatch between the wireless front-end and optical backhaul. The redundant capacity in the optical backhaul will provide extra reliability during the failure so that traffic can be rerouted through alternate paths. At a specific instant, it is possible to find several WOBAN topologies that can satisfy the required capacity and reliability objectives. All these are possible due to the densely interconnected wireless mesh front-end which has many redundant paths to route traffic. The flexibility provided by the wireless front-end of WOBAN can be exploited to enable energy savings in the optical part. There is another important observation on access network traffic profile. The traffic load on access network comes directly from customers, and it is well known that there are daily fluctuations of this load. During WOBAN (as well as other access networks) deployment, the common practice is to deploy network equipments so that it can support maximum traffic load. Consequently, during low-load hours, some parts of the network may be under-utilized. Hence, to design WOBAN topologies with reduced power consumption, we need to consider the following points – (a) deployed network has some extra capacity, and (b) traffic load varies during different hours of the day. Thus, we can selectively put some nodes to a low-power (sleep) state during low-load hours, thereby reducing network-wide power consumption. In the wireless front-end of WOBAN, we can adopt coordinated sleeping techniques [4] from mobile adhoc networks research to reduce wireless router energy consumption. Therefore, in this paper, we mainly focus on how to put optical components of WOBAN into sleep state. We will not consider putting OLT into sleep state as it connects the WOBAN to the rest-of-the-Internet. However, for protection purposes, in a PON segment, it is possible to have several OLTs in a ring setup. In that case, a low-load OLT can be put into sleep state while rerouting its traffic through other OLTs. In this paper, we intend to reduce ONU power consumption in WOBAN by putting low-load ONUs into sleep state. Now, how can we put an ONU to sleep state? Current IEEE 802.3ah/ 802.3av standards do not define any low power state for ONU [13]. However, proposals have been made to IEEE 802.3av task force to include low-power states for ONU so that it can go to sleep during idle periods [14]. Typical power consumption by an ONU during active state is approximately 10 W [15]. It is also estimated that during sleep state, power consumed by an ONU is less than 1 W [14]. Existing ONU boxes in the market include a T X DISABLE input which disables the transceiver of an ONU [15]. Disabling the transceiver can reduce ONU power consumption several fold. In WOBAN, OLT can manage a centralized sleeping mechanism to put low-load ONUs into sleep. The mechanism works as follows. An OLT maintains two watermarks for the traffic load at ONUs – low and high watermark. At OLT, there are input queues for each ONUs connected to it. During different hours of the day, OLT will observe the traffic load at different ONUs by measuring the length of corresponding input queues. If an ONU is operating under low watermark, OLT can put that ONU into sleep. The wireless mesh front-end of WOBAN will reroute the affected traffic due to ONU shut-down to alternate paths. OLT can put back sleeping ONUs into active state when traffic load increases above high watermark in the currently active ONUs. Now, we aim to find out the optimal number of ONUs needed


to support a given amount of traffic load. This is a form of multicommodity flow problem where each commodity represents the traffic flow between a source-destination pair. We can formulate the problem as a Mixed Integer Linear Program (MILP). In the following, we present the mathematical model in MILP form. 1) Mathematical Model: Our model takes as input a WOBAN with preassigned link capacities and a traffic matrix based on the dynamic daily traffic profile of a service area. The model generates output which determines minimum number of ONUs that need to be kept active to route the given traffic load. The other ONUs can be put to sleep state so that network-wide power consumption is minimized. The model also finds the routing path for each source-destination (s, d) pair in the traffic matrix. In WOBAN, for the upstream traffic, the destination is always OLT where the source can be one of the wireless routers, and for the downstream traffic, the source is the OLT and the destination is one of the wireless routers. To describe the model, we introduce some notations for the parameters and variables as follows. • WOBAN topology: denoted by a weighted and directed graph G = {V, E} where V is the set of nodes and E is the set of links. V has three subsets – Vw is the set of wireless nodes, Vonu is the set of ONUs, and OLT represents the OLT. If nodes u and v have a link, the link is denoted by (u, v). E has several subsets – Ew is the set of wireless links, EOG is the set of ONU-to-Gateway links, EGO is the set of Gateway-to-ONU links, ET O is the set of OLT-toONU links, and EOT is the set of ONU-to-OLT links. • (s, d): identifies source-destination pair of the upstream/downstream traffic in the traffic matrix. • Xu : Binary variable denoting ONU state, Xu ∈ {0, 1}. 0 denotes ONU is asleep, and 1 denotes ON U is active. s,d • λu,v : Binary variable denoting download flow on link (u, v) for a (s, d) (s is OLT, d denotes routers) pair, λs,d u,v ∈ {0, 1}. s,d • γu,v : Binary variable denoting upload flow on link (u, v) for s,d a (s, d) (s denotes routers and d is OLT) pair, γu,v ∈ {0, 1}. • Cu,v : Variable expressing the capacity over a wireless link (u, v). This variable can assume non-integral values. • COG : Capacity of ONU-to-Gateway link. • CGO : Capacity of Gateway-to-ONU link. • COT : Capacity of ONU-to-OLT link. • CT O : Capacity of OLT-to-ONU link. • T : Input traffic matrix with two different types of traffic values - (1) Vs,d : Download traffic between a (s, d) pair and (2) Zs,d : Upload traffic between a (s, d) pair. For each (s, d) pair, we assume the upload traffic is a fraction of V the download traffic, i.e., Zs,d = fd,s , ∀s,d where ft is t a constant value. Now, the objective function can be written as: minimize

Xu

(1)

s,d (u,v) Zs,d γu,v

s,d (u,v) Vs,d λu,v

−

s,d (v,u) Vs,d λv,u

=

∀u ∈ V, ∀(s, d) ∈ T

−Vs,d , u = d +Vs,d , u = OLT 0, otherwise

=

∀u ∈ V, ∀(s, d) ∈ T

+Zs,d , u = s −Zs,d , u = OLT 0, otherwise

s,d (Vs,d λs,d u,v + Zs,d γu,v ) ≤ Cu,v ,

(3)

∀(u, v) ∈ Ew

(4)

(s,d)

Wireless Constraints: The dual-threshold interference model [16] is used to find the set of all interfering links at each wireless node of WOBAN. The wireless interference constraints are translated to constraints that allocate capacity on each wireless link. A wireless node divides its capacity to all its incoming and outgoing links as it can not transmit and receive at the same time. So, the interference-free radio capacity available (Cu ) at each node u is shared between all the outgoing links from u (first term in Eqn. 5) and all the incoming links to u (second term in Eqn. 5).

Cu,v +

v

Cv,u ≤ Cu ,

∀u ∈ Vw

(5)

v

Equation 6 forms the secondary interference constraint of the wireless mesh of WOBAN. This constraint states that a node cannot receive any signal from any other node when an interfering link is active. The first term of Eqn. 6 is same as Eqn. 5, representing the shared capacity among incoming links to node u. The second term represents all the links which interfere with node u (Iu,v ) and which do not have node u as one of their end points.

Cv,u +

Cp,q ≤ Cu ,

∀u ∈ Vw

(6)

(p,q)∈Iu,v

v

Wired Capacity Constraints: Downstream traffic flows are limited by the capacity of the ONU-to-GW (Eqn. 7), and OLT-toONU (Eqn. 8) links. Similarly, upstream traffic flows are limited by the capacity of the GW-to-ONU (Eqn. 9), and ONU-to-OLT (Eqn. 10) links.

Vs,d λs,d u,v ≤ COG ,

∀(u, v) ∈ EOG

(7)

Vs,d λs,d u,v ≤ CT O ,

∀(u, v) ∈ ET O

(8)

s,d Zs,d γu,v ≤ CGO ,

∀(u, v) ∈ EGO

(9)

s,d Zs,d γu,v ≤ COT ,

∀(u, v) ∈ EOT

(10)

(s,d)

(s,d)

(s,d)

(s,d)

Wired Directionality Constraints: These constraints ensure that no upstream traffic is flowing in the downstream direction in the wired part of WOBAN and vice versa. λs,d u,v = 0,

∀(u, v) ∈ EOT ∪ EGO , ∀(s, d) ∈ T

(11)

s,d γu,v

∀(u, v) ∈ ET O ∪ EOG , ∀(s, d) ∈ T

(12)

= 0,

ONU State Constraint: This constraint determines the ONU s,d state. If some traffic (upstream (γu,v ) or downstream (λs,d u,v )) flows through an ONU u, it should be active (Xu = 1), otherwise it should be in sleep state (Xu = 0). This condition can be expressed by the following constraint,

subject to the following constraints: Flow Constraints: Equation 2 captures the fact that, in all nodes of WOBAN, total outgoing downstream traffic should be equal to total incoming downstream traffic except for the source (OLT) and the destination nodes (wireless routers). Similar argument holds for upstream traffic (Eqn. 3) except for the source nodes (wireless routers) and the destination (OLT).

s,d (v,u) Zs,d γv,u

Wireless Capacity Constraint: A wireless link in WOBAN can carry both upstream and downstream traffic. Equation 4 states that the summation of all traffic through a wireless link (u, v) should not exceed the capacity (Cu,v ) of the link.

u∈Vonu

−

Xu ≥

v

(s,d)

λs,d u,v +

v

M

(s,d)

s,d γu,v

,

∀u ∈ Vonu

(13)

where M is a very large value used to map the flow variables into a binary variable (Xu ). Path Length Constraints: These constraints put a limit on path length. As Eqns. 14 and 15 shows, each upstream or downstream (s, d) path should not be longer than H hops.

(2)

λs,d u,v ≤ H,

∀(s, d) ∈ T

(14)

u,v


s,d γu,v ≤ H,

∀(s, d) ∈ T

3

(15)

B

u,v

C

3

This model turns out to be a MILP formulation as some variables (such as Cu,v ) can take non-integral values. Generally, MILP problems are known to be NP-hard, but sophisticated heuristics can be used to solve the problems in a reasonable amount of time. The number of design variables in this problem is not exceptionally large. But for a very large number of nodes, the problem becomes quite large. We try to solve our model with smaller networks to verify the correctness of our formulation. For solving the MILP, we use ILOG CPLEX [17] on a Intel Core 2 Duo machine with 1 Gigabyte RAM and Ubuntu Linux OS. CPLEX uses state-of-theart branch-and-cut for solving MILP problems. We use a small 43 node WOBAN with 30 wireless routers, 8 gateways, 4 ONUs, and 1 OLT. Each ONU can drive 2 gateways. In this network, the MILP can be solved in a reasonable amount of time depending on the size of the traffic matrix. With various combinations of the load size and path length constraints, the results show that we have to keep at least one ONU alive and in the extreme case, all ONUs are on. However, for WOBAN with a larger number of nodes and more traffic load, we need to build some heuristics to solve this problem. We already have a heuristic on deciding which ONUs to shut down “put ONUs with load less than low watermark to sleep.” For routing, we need to find an energy-aware heuristic which is described below. C. Energy-Aware Routing There are several routing protocols proposed for WOBANlike architectures. Two of them are – (a) Delay Aware Routing Algorithm (DARA) [18], (b) Capacity and Delay Aware Routing (CaDAR) [18]. These routing algorithms are Link State (LS) protocols where a node periodically transmits its link-state information to the network by Link-State Advertisement (LSA). Upon receiving the LSAs from all the nodes, each node finds a map of the network and can build a routing table (generally by using some variant of Dijkstra’s algorithm) to route traffic to other nodes in the network. LS protocols generally vary on how they assign link weights in the LSA. For example, DARA uses predicted link delay metric to assign link weights. Based on link weight assignment, these protocols try to achieve several performance objectives. One such objective is load balancing which balances the traffic load in all parts of the network [18]. Load balancing is a good performance objective as it tries to fairly utilize all parts of the network. But, it may lead to under-utilization of some segments of the network during lowload hours. During low-load hours, traffic can be supported using small number of devices in the network. Our routing algorithm is an energy-aware LS protocol with the objective to reduce network-wide energy consumption by putting underutilized nodes (mainly ONUs) of the network into sleep. When routing traffic, the objective will be “use the already-used paths.” The routing algorithm tries to route traffic through the ONUs which are already used so that probabilities of other ONUs being unused increase. In that way, we can put zero-load ONUs to sleep. Moreover, we may find some other ONUs with very low load (ONUs with loads under low watermark). By being more aggressive, we can put these ONUs to sleep and let the wireless mesh reroute the traffic through these ONUs. When traffic load increases, sleeping ONUs can be activated to carry increased traffic. [Note: Due to space constraints, we only present the logic behind the routing algorithm]

3 7

S

6

H

2

G E

F 2

Fig. 2.

D 1

2

Residual capacity as link weights.

To achieve this, we can modify the LS routing algorithm in WOBAN so that link weights are assigned to satisfy our energy objective. So, we use residual capacity as the link weight. Initially, each link will have its initial capacity as the weight. When traffic flows through a link, its next link weight will be the capacity left (original capacity minus traffic flow) on that link. To route traffic from source to destination, we find the lowest residual capacity path. For example, let us consider the small network in Fig. 2. The links are annotated with their weights (residual capacities). To send traffic from S to D, energy-aware routing algorithm will route traffic through the path S-E-F-G-D which has the lowest residual capacity (2 + 2 + 2 + 1 = 7). This approach, however, has its own shortcomings as shown in Fig. 2. The algorithm selects the path with 4 hops although that is not the shortest path while using other metrics (such as hop length or delay). This will increase the average path length and path delay in the network. To deal with this problem, we can introduce a term named hop offset – the purpose of this term is to reduce average path length. If we have a hop offset m, we add m to the path cost for each hop, i.e., for a path of n hops, the cost of the path will be residual capacity of the path +n × m. For example, as in Fig. 2, for a hop offset 1, the path costs are 7 + 4 × 1 = 11, 9 + 3 × 1 = 12, and 13 + 2 × 1 = 15 for paths SE-F-G-D, S-B-C-D, and S-H-D respectively, and S-E-F-G-D will be the selected path. But, for a hop offset 3, S-B-C-D will be the chosen path in our algorithm, and for a hop offset 5, S-H-D will be the chosen path. Selecting the optimal hop offset depends on how much delay the network connections can tolerate. There is another important item to consider. We should select hop offset in such a way that average path length does not increase unproportionately from regular shortest path routing. If average path length increases too much, that means more wireless hops per path, i.e., more wireless transmissions and receptions. Each wireless transmission/reception requires power. So, out of proportion average path length may diminish the power savings that we gain from putting the ONUs to sleep. IV. C ASE S TUDY Figure 3 shows a hypothetical WOBAN deployment scenario in Davis, which is a small city in Northern California near Sacramento. Davis is the home of the University of California, Davis. The selected part of Davis has three different areas: (a) Downtown, (b) UC Davis Campus, and (c) Part of residential area. These areas are selected as they have a very nice blend of technology-savvy users, and we can showcase how traffic profile varies across different parts of the network, and also depending of the traffic profile during different time of the day, how we can put nodes to sleep. The telecom CO is situated in the downtown area. Total 140 wireless routers, 24 gateways, 12 ONUs, and 1 OLT are deployed in this hypothetical deployment. The OLT drives 12 ONUs, and 1 ONU drives 2 gateways. Downtown has 41 routers, 8 gateways, and 4 ONUs; Campus area has 37 routers, 6 gateways, and 3 ONUs; and Residence area has 62 routers, 10 gateways,


(c) Part of Davis Residential Area

Map Courtesy: Google Maps 1

Legends: 0.9

Telecom CO

Downtown Area Campus Area Residence Area

Percentage of Active Routers

0.8

OLT Splitter ONU Wireless Gateway Wireless Router Optical Fiber CAT-5 Cable

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

01−03 1

2 04−06

3 07−09

4 10−12

5 13−15

6 16−18

7 19−21

8 22−24

7 19−21

8 22−24

Hours

(a) Percentage of active routers

3 Downtown Area Campus Area Residence Area

(a) Davis Downtown

Hypothetical WOBAN deployment in Davis.

and 5 ONUs deployed. The routers are equipped with one radio with a capacity of 54 Mbps (IEEE 802.11g). Capacity allocation among wireless links is accomplished by TDM link scheduling. The OLT and ONU have capacities of 1 Gbps and 100 Mbps respectively. The low watermark is set to 5%, i.e., the OLT puts ONUs with load less than 5% of the total traffic to sleep. The aforementioned MILP model is not able to produce optimization results for a large WOBAN like the hypothetical deployment in Davis in a reasonable amount of time. Therefore, we use two heuristic approaches in our simulation to achieve near-optimal solution. They are - (a) put ONUs with load less then low watermark to sleep, (b) energy-aware routing algorithm. A. Traffic Modelling To show numerical examples on the performance of energyaware WOBAN design, we need to develop a reasonable traffic profile of the deployment areas during different hours of the day. Access networks deal directly with the user-generated Internet traffic. So, the behavior of end users has a significant impact on the performance of the access network. Based on different sources [19] and our observations, we develop a traffic profile for each of the three different areas of Davis. Each of these areas has different types of Internet users with different behavior patterns and different peak usage periods. We divide the 24 hours of a day into 8 periods: (1) Period 1 (01-03 hours), (2) Period 2 - (04-06 hours), (3) Period 3 - (07-09 hours), (4) Period 4 - (10-12 hours), (5) Period 5 - (13-15 hours), (6) Period 6 - (16-18 hours), (7) Period 7 - (19-21 hours), and (8) Period 8 - (22-24 hours). Period 1 begins at 00.01 AM in the morning. The granularity of these periods can be modified as necessary. Figure 4 shows such a traffic profile for downtown, campus, and residential areas of Davis. There are two parts of the traffic profile – Fig. 4(a) presents the percentage of active routers (which is directly proportional to the active users) in these areas, and Fig. 4(b) shows the average (poisson distributed) load

2

1.5

1

0.5

0

1 01−03

2 04−06

3 07−09

4 10−12

5 13−15

6 16−18

Hours

(b) Average load on active routers Fig. 4.

Traffic profile of Davis Internet users. 60

Regular Mode Power Save Mode Energy−Efficient Routing

50

% of ONUs in Sleep State

Fig. 3.

Average Load at Active Routers (Mbps)

(b) UC Davis Campus

2.5

40

30

20

10

0

1 01−03

2 04−06

3 07−09

4 10−12

5 13−15

6 16−18

7 19−21

8 22−24

Hours

Fig. 5.

Power savings in energy-aware WOBAN.

on these active routers. Both of these data together generate the traffic profile of these areas. We assume for each active user, the upload traffic is approximately 14 of the download traffic. V. I LLUSTRATIVE N UMERICAL E XAMPLES We apply our energy-aware design and routing protocol on the hypothetical WOBAN deployment in Davis (Fig. 3). We collect


5 4.5


4

Average Path Length

3.5 3 2.5 2 1.5 1 0.5 0

1 01−03

2 04−06

3 07−09

4 10−12

5 13−15

6 16−18

7 19−21

8 22−24

Hours

(a) Average path length

1 0.9


Average Path Delay (msec)

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

1 01−03

2 04−06

3 07−09

4 10−12

5 13−15

6 16−18

7 19−21

8 22−24

Hours

(b) Average path delay Fig. 6.

Energy-aware WOBAN performance.

results for three different setups: (a) WOBAN in regular mode with no energy-saving techniques, (b) WOBAN in power-save mode (with regular shortest path routing) where low-load ONUs are put to sleep, and (c) WOBAN in energy-aware routing mode where energy-aware routing is deployed on top of power-save mode configurations. We compare the energy savings and measure the impact of ONU shutdown on the performance of the network. Figure 5 shows the power savings (in terms of percentage of ONUs in sleep state) during different periods of the day. Obviously, there is no energy saving in the regular mode. Interestingly, on an average, we can put 50% of the ONUs to sleep state in our hypothetical deployment by using the other two setups. One may argue that if we can put 50% of the ONUs to sleep, why do we deploy them in the first place? The answer is that, at high load, when all the routers are active, we will not be able to put any ONU to sleep. The power-saving opportunity lies somewhere else – specifically in the traffic profile. When one part of the network is at high load, the other parts may be in low load. We can save energy by putting ONUs in those low-load parts to sleep. Figure 6(a) presents the average path length in three different setups of WOBAN. In WOBAN, all (s, d) paths have two wired (OLT⇔ONU, ONU⇔GW) hops, and the rest are wireless hops.

The average path lengths in the energy-aware routing mode are comparable with the regular mode. Hop Offset tries to reduce the average path length in energy-aware routing mode. Otherwise, higher average path length could diminish our power conserving benefits with the extra wireless transmission/reception power consumed. We can see that putting ONUs to sleep does not significantly increase the average number of wireless hops, thanks to the availability of various similar-cost paths in the wireless mesh of WOBAN. Hence, the energy usage in the wireless part does not increase significantly when we put ONUs to sleep. Figure 6(b) provides the average transmission delay of the paths for different setups. Again, the transmission delays are very much comparable in all these setups. Consequently, we can state from the results that we can save a good portion of energy consumption in WOBAN by careful design and energy-aware routing without compromising the performance. VI. C ONCLUSION In this paper, we showed that energy consumption of WOBAN can be reduced by efficient design and energy-aware routing. We examined the impact of these energy-aware design decisions on the performance of the network. It appears that with suitable design decision parameters, we can achieve comparable performance (of WOBAN) with the energy-aware mode. The energysaving in the optical part of WOBAN also does not increase the energy usage in the wireless part. We conclude that the energyaware design techniques applied on WOBAN can be generalized so that they are also applicable to other access networks. R EFERENCES [1] S. Sarkar, S. Dixit, and B. Mukherjee, “Hybrid Wireless-Optical Broadband Access Network (WOBAN): A review of relevant challenges,” Journal of Lightwave Technology (JLT), vol. 25, no. 11, pp. 3329–3340, Nov. 2007. [2] “Infonetics research.” http://www.infonetics.com/, 2009. [3] K. J. Christensen, C. Gunaratne, B. Nordman, and A. D. George, “The next frontier for communications networks: power management,” Elsevier Computer Communications, vol. 27, pp. 1758–1770, Aug. 2004. [4] M. Gupta and S. Singh, “Greening of the Internet,” ACM SIGCOMM’03, Karlsruhe, Germany, p. 1926, 25-29 Aug. 2003. [5] C. Lange and A. Gladisch, “On the energy consumption of FTTH access networks,” OFC/NFOEC’09, San Diego, Mar. 2009. [6] B. St. Arnaud, “CANARIE.” http://www.canarie.ca/, 2009. [7] J. Chabarek, J. Sommers, P. Barford, C. Estan, D. Tsiang, and S. Wright, “Power awareness in network design and routing,” IEEE INFOCOM’08, Phoenix, Arizona, pp. 1130–1138, 2008. [8] I. Keslassy, S. Chuang, K. Yu, D. Miller, M. Horowitz, and O. Solgaard et al., “Scaling Internet routers using optics,” ACM SIGCOMM’03, Karlsruhe, Germany, Aug. 2003. [9] L. Benini, A. Bogliolo, and G. D. Micheli, “A survey of design techniques for system level dynamic power management,” IEEE Transactions on VLSI Systems, vol. 8, no. 3, pp. 299–316, Jun. 2000. [10] C. Jones, K. Sivalingam, P. Agrawal, and J. Chen, “A survey of energy efficient network protocols for wireless networks,” Wireless Networks, vol. 7, no. 4, no. 4, pp. 343–358, 2001. [11] J. Chase and R. Doyle, “Balance of power: energy management for server clusters,” 8th Workshop on Hot Topics in Operating Systems, May 2001. [12] M. Allman, K. Christensen, B. Nordman, and V. Paxson, “Enabling an energy-efficient future internet through selectively connected end systems,” ACM SIGCOMM HotNets ’07, Atlanta, Georgia, Nov. 2007. [13] G. Kramer. Personal Communication, Mar. 2009. [14] J. Mandin, “EPON powersaving via sleep mode.” IEEE P802.3av 10GEPON Task Force Meeting, www.ieee802.org/3/av/public/2008-09/3av0809-mandin-4.pdf, Sept. 2008. [15] “NEC GE-PON data sheets.” http://www.nec.co.jp/, 2009. [16] V. Ramamurthi, A. S. Reaz, and B. Mukherjee, “Optimal capacity allocation in wireless mesh networks,” IEEE Globecom’08, New Orleans, Dec. 2008. [17] “ILOG CPLEX: High-performance software for mathematical programming and optimization.” http://www.ilog.com/products/cplex/, 2009. [18] S. Sarkar, P. Chowdhury, S. Dixit, and B. Mukherjee, Broadband Access Networks: Technologies and Deployment, ch. Hybrid Wireless-Optical Broadband Access Network (WOBAN). Springer, 2009. [19] J. Farber, S. Bodamer, and J. Charzinski, “Measurement and modelling of Internet traffic at access networks,” EUNICE’98, Munich, pp. 196–203, Aug. 1998.
