THE Information and Communication Technology (ICT)

Transcription

1 2094 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 11, JUNE 1, 2014 Follow the Sun, Follow the Wind Lightpath Virtual Topology Reconfiguration in IP Over WDM Network Gangxiang Shen, Senior Member, IEEE, Yunlei Lui, and Sanjay Kumar Bose, Senior Member, IEEE Abstract Green House Gas (GHG) emissions mainly come from the consumption of non-renewable energy. To reduce GHG emissions of IP over WDM networks, we propose to maximize renewable energy usage at each network node location so as to reduce the consumption of non-renewable energy. A Follow the Sun, Follow the Wind strategy is proposed for the IP over WDM network to periodically reconfigure the lightpath virtual topology to enable more lightpaths to start or end at nodes where maximum renewable energy is available. We develop a mixed integer linear programming model to design new lightpath virtual topologies. Since the computational complexity of the optimization model is excessive, we also propose a simple but efficient heuristic algorithm to tackle this. Our results indicate that a network operated in this way can significantly reduce non-renewable energy consumption as illustrated in the example network scenarios considered. Index Terms Follow the sun, follow the wind (FTSFTW), green network operation, mixed integer linear programming (MILP) model, renewable energy. I. INTRODUCTION THE Information and Communication Technology (ICT) sector worldwide is guilty of contributing around 2% of the man-made CO 2 emitted each year [1]. If nothing is done to reduce this, ICT s contribution to Green House Gas (GHG) emissions is projected to nearly double by 2020 [2], [3]. This would be clearly unsustainable and therefore reducing GHG emissions in ICT applications such as networks and communications has become an urgent and challenging area of research. The current approach of dealing with this issue is to improve the overall energy efficiency of the network [4] [6]. An approach commonly used for this is to do lightpath bypass, which reduces as much as possible the number of optical-electricaloptical conversions in the network since these tend to consume the most energy. For example, the authors in [7] design an Manuscript received November 22, 2013; revised February 25, 2014 and April 2, 2014; accepted April 2, Date of publication April 15, 2014; date of current version May 21, Part of this paper was presented at the 11th International Conference on Optical Internet [32]. This work was supported in part by the National 863 Plans Project of China under Grant 2012AA011302, the National Natural Science Foundation of China under Grants and , the Research Fund for the Doctoral Program of Higher Education of China under Grant , and the Natural Science Foundation of Jiangsu Province under Grants BK and BK G. Shen and Y. Lui are with the School of Electronic and Information Engineering, Soochow University, Suzhou , China ( shengx@suda.edu.cn; yunlei198938@163.com). S. K. Bose is with the Indian Institute of Technology, Guwahati , India ( skbose@iitg.ernet.in). Color versions of one or more of the figures in this paper are available online at Digital Object Identifier /JLT energy-minimized IP over WDM network by employing the lightpath bypass strategy to reduce the number of router ports used. Another interesting method is to turn off or to put some network devices or components into sleep mode when they are free of traffic. In [8], a realistic IP network topology was considered and evaluated for the amount of energy that can be potentially saved when nodes and links which are free of traffic are turned off during off-peak periods. Compared with the single-line-rate technology implemented currently, mixed-line-rate technology with more flexibility in provisioning bandwidth on demand is another feasible way to achieve a more energy-efficient network [9], [10]. In addition, other technologies such as traffic grooming [11] [13] and energy-aware routing and wavelength assignment (RWA) [14] [16] have been proposed to improve the overall energy efficiency of an optical network. Alternatively, GHG emissions can also be reduced by promoting greater use of renewable energy, such as solar or wind energy, and this has always been a recognized goal of the ICT industry [17] [20]. For example, [17] and [18] suggest that cloud computing services should be preferably handled by data centers which have sufficient renewable energy for processing, if the energy consumption for processing is more than the energy used to transport the data. Y. Zhang et al. [21] also focus on the problem of dynamically distributing service requests among data centers in different geographical locations, based on the local weather conditions, so as to maximize the use of renewable energy. Using an energy-minimized design to further reduce the GHG emission in IP over WDM networks has been studied in [22]. However, designing a robust IP over WDM network with renewable energy sources such as solar and wind power is a challenge, as these are far less reliable than hydro-electric or traditional fossil fuel power plants [18]. The solar and wind power available at any given location will change dynamically depending on the local weather conditions, i.e., changing sunlight and wind speeds. Fortunately, it is possible to predict the solar and wind energy generation during each time slot or period based on the weather forecast [23], [24]. For example, a time slot of one day has been chosen in this paper (to be more accurate, it is also possible to consider an hourly slot for the prediction of green energy generation. However, we lack accurate hourly weather forecast data). This implies that the amount of renewable energy available at a network node in a certain period can be a known parameter. This can be used in operating an IP over WDM network which aims to consume the least amount of non-renewable energy by following the available state of renewable energy in that time slot/period. In a subsequent time slot/period (e.g., next day), the network can be reconfigured based on the availability IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See standards/publications/rights/index.html for more information.

2 SHEN et al.: FOLLOW THE SUN, FOLLOW THE WIND LIGHTPATH VIRTUAL TOPOLOGY RECONFIGURATION IN IP OVER WDM NETWORK 2095 Fig. 1. Architecture of IP over WDM network. of solar and wind power in that time slot/period. We call the network based on the above operational strategy Follow the Sun, Follow the Wind IP over WDM networks. The key contributions of this paper are summarized as follows: (1) develop a Follow the Sun, Follow the Wind IP over WDM network operational strategy and associated resource allocation method which will periodically reconfigure the lightpath virtual topology (e.g., reconfigure the lightpath virtual topology every day at midnight) based on the amount of solar and wind energy predicted at each network node; (2) minimize the energy consumption of the whole network using an optimization model and propose a simple but efficient heuristic for the configuration of the Follow the Sun, Follow the Wind type of IP over WDM network. In addition, the key differences of this study from existing works like [22] are (1) the work in [22] designs an IP over WDM network which aims to minimize the non-renewable energy consumption based on the average output of solar energy from historical data, while we operate an IP over WDM network based on the daily forecasted weather conditions and outputs of renewable energy at each node location; (2) the work in [22] does not consider the reconfiguration of virtual topology because it performs a singleiteration design based on yearly average output of solar energy, while we periodically reconfigure the lightpath virtual topology to ensure the minimum consumption of non-renewable energy for each day. The rest of this paper is organized as follows. In Section II, we introduce the network model and the mechanism of Follow the Sun, Follow the Wind. In Section III, we present the research problem and the optimization model. A simple but efficient heuristic is proposed in Section IV. In Section V, we present the models to predict the renewable energy generation and the test conditions of this study. In Section VI, we present and discuss the simulation results for the study. The paper is concluded in Section VII. II. NETWORK MODEL AND FOLLOW THE SUN, FOLLOW THE WIND APPROACH A. Network Model We consider a transparent IP over WDM network optical network [10] as shown in Fig. 1. The IP over WDM network consists of two layers, i.e., the optical layer and the IP layer. Nodes in the optical layer are optical cross-connects (OXCs), which are interconnected by physical fiber links. Associated with each fiber link, a pair of wavelength multiplexer and demultiplexer is deployed to combine and separate wavelengths at the two ends Fig. 2. An example of virtual topology establishment with renewable energy: (a) physical topology; (b) virtual topology with non-renewable energy; (c) virtual topology with sufficient renewable energy at node N1; (d) virtual topology with sufficient renewable energy at node N2. of the link. To enable optical signals to travel long distances, EDFAs may be deployed along the fiber links. In addition, there are many end-to-end lightpaths or optical channels connecting source and destination nodes. For each end-to-end optical channel, a pair of transponders is required for data transmission. These end-to-end optical channels form a virtual topology for the upper IP layer, in which each virtual link can contain multiple end-to-end optical channels that connect the same pair of source and destination (router) nodes. In the IP layer, a core router connects to transponders via short-reach (SR) interfaces and the core routers connect to each other via virtual links, which correspond to end-to-end lightpaths in the optical layer. A hybrid-power node architecture where power is supplied by mix of renewable energy and non-renewable energy [22] is considered in this study. A non-renewable energy source is needed as a backup if the amount of renewable energy available at the node is not sufficient for operation during a certain period (e.g., a day); this is expected to happen because of the unreliable nature of the renewable energy sources. We assume that the consumption of non-renewable energy causes much higher GHG emissions compared with the consumption of renewable energy as in [18] and therefore the GHG emission caused by consumption of renewable energy is neglected. Thus, the total GHG emissions of an IP over WDM network will be decided by the consumption of non-renewable energy and it reduces if a portion of the non-renewable energy consumption or possibly all of it is replaced by renewable energy. Therefore, the current problem of reducing the GHG emission of an IP over WDM network becomes one of minimizing the total non-renewable energy consumption in the network. B. Mechanism of Follow the Sun, Follow the Wind Fig. 2 demonstrates the mechanism of Follow the Sun, Follow the Wind IP over WDM networks. Fig. 2(a) shows an example network containing three nodes connected to each other through fiber links. We assume that the node pair (N0, N1) has 25-Gb/s IP traffic and the node pair (N0, N2) has 10-Gb/s IP traffic. In addition, the capacity of each wavelength is assumed to be 40 Gb/s.

3 2096 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 11, JUNE 1, 2014 To serve the IP traffic, direct lightpaths (N0, N1) and (N0, N2) are established to form the virtual topology for the IP layer as shown in Fig. 2 (b). This minimizes the total energy consumption and thus reduces the GHG emissions when no renewable energy is available in the network. However, assume that the location at node N1 has a very sunny and windy day and thus has sufficient renewable energy available. To reduce the non-renewable energy consumption, we can establish lightpaths (N0, N1) and (N1, N2) for the IP layer as shown in Fig. 2(c). IP traffic between node pair (N0, N1) is served by lightpath (N0, N1) and IP traffic between node pair (N0, N2) is carried through two consecutive lightpaths (N0, N1) and (N1, N2). Note that the two traffic flows are groomed on the common lightpath (N0, N1). Here, although the number of lightpaths in the network does not decrease, the consumption of non-renewable energy is reduced. This is because there are just two router ports at nodes N0 and N2 consuming non-renewable energy in Fig. 2(c) compared to three router ports at those nodes consuming non-renewable energy as in Fig. 2(b). For the same reason, when the node location N2 has a sunny and windy day and the others do not, we can establish the virtual topology as in Fig. 2(d) to reduce the consumption of non-renewable energy. When this approach is followed, IP over WDM networks which establish virtual topologies according to the state of sun and wind so as to minimize non-renewable energy consumption are called Follow the Sun, Follow the Wind networks. III. MILP OPTIMIZATION MODEL A. Problem Statement Our objective is to build a lightpath virtual topology for the Follow the Sun, Follow the Wind IP over WDM network that consumes the least amount of non-renewable energy. The following inputs are given. 1) A physical topology G = (N, E), which consists of a set of nodes N and links E. The node set corresponds to IP routers and OXCs. Within a single node, an IP router is connected to an OXC via SR interfaces. The link set consists of the physical fiber links in the network. 2) A forecast traffic matrix Λ sd between node pairs (s, d). 3) The amount of renewable energy (sum of solar and wind energy) ES i available in a day at node i. We assume that the renewable energy outputs can be predicted based on weather reports (i.e., based on the state of the sun and the wind on that day). These inputs are provided as the given parameters of the optimization problem. The optimization objective is to minimize the total non-renewable energy consumption of an IP over WDM network on that day. The constraints of the problem include (1) fully serving all the traffic demands between node pairs on that day, (2) the total energy supply (renewable energy and non-renewable energy) at each node is sufficient to support its energy consumption on that day, (3) a limited maximum number of router ports/transponders at each node so as to keep the capital expenditure (CAPEX) at an acceptable level, and (4) a limited transmission capacity of each transponder/router port. The problem aims to find (1) an optimal virtual topology which consumes the least non-renewable energy in the optical layer and IP layer on that day, (2) the number of used router ports/transponders required at each network node, and (3) the amount of non-renewable energy consumed at each node. Note that the power consumption of optical amplifiers and optical cross-connects is much lower than that of router ports/transponders. Typically, it contributes only about 3 5% of total network energy consumption [7]. Therefore, in this study we ignore the energy consumption of these components so as to make the design problem more tractable. B. Other Terms We define additional terms as follows. Sets and Parameters: In addition to all the parameters introduced in Section III-A, we define other sets and parameters as follows: B capacity of each wavelength channel in Gb/s; E r total energy consumption of each router port in a day; E t total energy consumption of each transponder in a day. Note that a router port corresponds to a transponder in the IP over WDM network and the capacity of each router port and transponder are both B Gb/s; Δ i the maximum number of router ports/transponders at node i Variables: λ sd ij traffic demand between node pair (s, d) that traverses virtual link (i, j) (real); v ij number of lightpaths (i.e., optical channels) between virtual link (i, j) (integer); tr i number of router ports/transponders at node i (integer); ens i non-renewable energy consumption at node i in a day (real). Note that the minimum value of non-renewable energy consumption at node i on a certain day can be zero if the renewable energy sources can supply the entire energy needed at that node without consuming any nonrenewable energy. Note that all the above variables can only take non-negative values. C. Objective Minimize i N ens i (i.e., minimize the total non-renewable energy consumption of the network in a day). D. Constraints λ sd ij j N :i j j N :i j s N d N :s d λ sd ij Λ sd, i = s λ sd ji = Λ sd, i = d 0, otherwise s, d, i N : s d (1) λ sd ij Bv ij i, j N : i j (2) = λ ds ji i, j, s, d N : i j, s d (3) v ij = v ji i, j N : i j (4) v ij = tr i i N (5) j N :i j

4 SHEN et al.: FOLLOW THE SUN, FOLLOW THE WIND LIGHTPATH VIRTUAL TOPOLOGY RECONFIGURATION IN IP OVER WDM NETWORK 2097 tr i Δ i i N (6) ens i + ES i tr i (E r + E t ) i N. (7) Constraint (1) ensures the flow conservation constraint in the IP layer. Constraint (2) ensures that there is sufficient capacity on each lightpath virtual link to carry traffic flows. Constraint (3) says that the traffic flows are bi-directional and are co-routed in the IP layer. Constraint (4) ensures that the lightpath virtual links for the IP layer are bi-directional. Because any unidirectional lightpath starting from a node needs a transmitter which is provided by a transponder, constraint (5) ensures that the total number of unidirectional lightpaths starting from node i is equal to the total number of transmitters (transponders) at node i. Constraint (6) ensures that the number of router ports/transponders deployed at each node does not exceed the maximum number set in this study so as to keep the CAPEX acceptable. The last constraint ensures that the total energy supply including renewable energy and non-renewable energy at each node can satisfy the total energy consumption of the network node. Note that in this paper, we take a simplified model to distribute the overhead energy consumption of the node. This is done by dividing the chassis and router backplane energy consumption evenly onto each router port. The power consumption of a router port is assumed to have incorporated the power consumption of both router processing/grooming on the router backplane and the router chassis. For example, if a router port consumes x-w power, and a router with p ports has y-w overhead power consumption. We evenly distribute the overhead power consumption onto each router port to find its power consumption to be x + y/p W. Thus, as in [7], to measure the total energy consumption of the IP over WDM network, we simply sum the energy consumption of all the router ports and transponders that are incorporated with the overhead energy cost. Also, to achieve the overall lowest total energy consumption of the entire network, we do not consider the constraints of traffic load balance on different virtual links and the maximal allowed physical distance of lightpaths. IV. HEURISTIC APPROACH We estimate the computational complexity of the above optimization as follows: the MILP model has a total of O( N 4 ) variables (due to variable λ sd ij ) and O( N 4 ) constraints [due to constraint (3)], where N is the total number of nodes in the network. For a large network which contains many networks nodes, there would be a huge number of variables and constraints. For example, if N = 100, there are a total of 10 8 variables and constraints, which makes the problem intractable. Therefore, for building the virtual topology of a large-size network, an efficient heuristic algorithm is required for a fast solution. To plan a lightpath virtual topology for an IP over WDM network, the multi-hop bypass heuristic algorithm proposed in [7] can reduce the total energy consumption for the entire network. However, the algorithm does not take into account the source of the energy, i.e., whether it comes from a renewable or nonrenewable source. In order to minimize the non-renewable energy consumption, we propose a new heuristic algorithm where the traffic flows preferentially traverse intermediate nodes which have more renewable energy. This avoids unnecessary consumption of non-renewable energy at those intermediate nodes which have less renewable energy. Traffic aggregated at the nodes which have a lot of renewable energy would make these nodes act as preferred central transit nodes through which lightpaths would be established to most of the other nodes. The detailed steps of the heuristic algorithm are as follows: Step 1: We first divide the whole IP traffic demand matrix into two matrices with the first containing an integral number of optical channels of capacity B Gb/s and the second containing all the remaining traffic whose amount is smaller than B Gb/s. The traffic demands in the first matrix will occupy the full bandwidth B Gb/s and cannot be groomed with others; these would also need at least one router port/transponder respectively at the source and destination nodes. For these, we employ the singlehop bypass strategy [7] to directly establish lightpath virtual links between the node pairs and add these lightpaths into the virtual topology. Each lightpath corresponds to a pair of router ports/transponders at the source and destination nodes. We then update the network state information which would include the remaining capacity for each lightpath, the number of router ports/transponders deployed at each node, and the amount of renewable energy used at each node. Step 2: To handle the traffic demands in the second matrix, we put all the node pairs into a node pair list NPL and select M network nodes which have the highest amounts of renewable energy 1 and put them into a node list NL. Note that this is done because we want the nodes which have the highest amounts of renewable energy to act as aggregator nodes. Then, for the node pairs whose source or destination node is in the node list NL,we establish direct lightpaths between these node pairs and add the lightpaths into the virtual topology. We then update the network state information. Step 3: For the remaining node pairs which still have not been served, we get each one from the NPL in sequence, serve them on the lightpath virtual topology (excluding the links which do not have sufficient remaining capacity) by employing the shortest path algorithm, and update the network state information, if that is possible. Otherwise, we find a shortest route based on the current lightpath virtual topology excluding the nodes which do not have enough renewable energy or where the number of router ports/transponders used exceeds the limit set for the node. This will ensure the selected route to traverse the intermediate nodes that have sufficient renewable energy so as to reduce the overall non-renewable energy consumption in the network. If this succeeds, we establish some or all lightpaths for the chosen route as needed to serve the traffic flow and update the corresponding network state information. Otherwise, we establish a direct lightpath between the node pair and update the network state information. The flowchart of Step 2 and Step 3 is given in Fig. 3. Step 4: After we finish establishing the lightpath virtual topology for the IP over WDM network, we can count the total number of router ports/transponders used and calculate the total amount of energy consumed at each node. We then further find 1 We order all the nodes based on the amounts of their renewable energy from the highest to the lowest, and choose top M nodes which have the highest amounts of renewable energy.

5 2098 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 11, JUNE 1, 2014 Fig. 3. Flowchart of Step 2 and Step 3 in the heuristic algorithm. Fig. 5. Test networks. (a) NSFNET, (b) US backbone network (USNET). Fig. 4. An example with shuffle strategy: (a) traffic matrix with serving order 1; (b) established virtual topology for serving order 1; (c) traffic matrix with serving order 2; (d) established virtual topology for serving order 2. the non-renewable energy consumption of the entire network by calculating the total energy consumption and the renewable energy available at each node on that day. Note that non-renewable energy consumption at each node can be zero if there is sufficient renewable energy available at the node which can supply all the energy required on that day. In the process of establishing the lightpath virtual topology for an IP over WDM network as above, we found that the serving sequence for the node pairs (i.e., the node pair list NPL) can significantly affect the optimization result. For example, consider a network with three nodes N0, N1, and N2 where the traffic demands of node pairs (N0, N1), (N0, N2), and (N1, N2) are 25, 25, and 10 Gb/s, respectively, and the capacity of each lightpath is 40 Gb/s. If the node pair list is ordered as (N1, N2), (N0, N1), and (N0, N2) [see Fig. 4(a)] and are served in sequence, then no traffic demand would share a common lightpath with others and three lightpaths are established to form the virtual topology [as shown in Fig. 4(b)]. In contrast, if the node pair list is ordered as (N0, N1), (N0, N2), and (N1, N2) [as shown in Fig. 4(c)], we establish two lightpaths (i.e., lightpaths (N0-N1) and (N0-N2)) to first serve the traffic demands between node pairs (N0, N1) and (N0, N2) [as shown in Fig. 4(d)]. For the traffic demand of node pair (N1, N2), the traffic grooming process can be applied to serve it by sharing the capacity on the common lightpaths (N0-N1) and (N0-N2). In this way, only two lightpaths are required to serve the same traffic demand matrix. Based on this observation, in order to further reduce non-renewable energy consumption, we apply a shuffle strategy to change the order of service sequence [25] in this heuristic algorithm. Specifically, we shuffle the node pairs in a list to change their order before Step 3, and then recalculate the non-renewable energy consumption by following the flowchart (i.e., Step 3 and Step 4). We repeat the process of shuffling node pairs and calculating non-renewable energy consumption T times, compare the results of non-renewable energy consumption, and select the one that consumes the least non-renewable energy as our final result. In the heuristic algorithm, we employ the Dijkstra s shortest path first algorithm to find the shortest route. The algorithm therefore has an O( N 2 ) complexity which is much lower than that of the MILP model. If we also consider the shuffling process, the computation complexity of the algorithm changes to be O(T N 2 ), where T is the times of node pair shuffling. V. RENEWABLE ENERGY MODELS AND TEST CONDITIONS Without losing generality, each network node location is assumed to be configured with the same numbers of solar panels and wind turbines to generate renewable energy whenever the

6 SHEN et al.: FOLLOW THE SUN, FOLLOW THE WIND LIGHTPATH VIRTUAL TOPOLOGY RECONFIGURATION IN IP OVER WDM NETWORK 2099 weather conditions permit [26]. The metrics of sky cover (i.e., cloud cover) and wind speed are considered as the most important factors in the renewable energy harvest from solar panels and wind turbines, respectively [27]. These metrics are assumed to be approximately predicted by the weather reports and can be used to estimate the output of renewable energy. Note that it is important to properly configure the numbers of solar panels and wind turbines. The outputs of renewable energy at each node cannot be too little; otherwise, no nodes can act as central (aggregator) nodes that can provide extra renewable energy in addition to its own energy consumption for establishing lightpath virtual links to relay the communications between other node pairs. On the other hand, the configuration also becomes meaningless if each network node is over-provisioned with renewable energy for its equipment power supply, which leads to an always-zero non-renewal energy consumption. Thus, in this study we deploy solar panels and wind turbines at each node that would provide just sufficient renewable energy if it is a sunny and windy day at the node location. More specifically, we deploy solar panels and wind turbines at each node that would provide renewable energy just sufficient for the node to operate with all (maximally 64) of the ports on and without consuming any non-renewable energy. According to a previous study [22], we assume that one square meter silicon solar cell can produce daily average 0.28 kw of power on a completely sunny day (i.e., sky cover equal to 0%), and generates no solar power on a completely cloudy day (i.e., sky cover equal to 100%). Based on this, we assume that the solar energy harvested in a day is proportional to the sky cover conditions on that day (i.e., P =0.28 (1 sky cover)). For example, if the sky cover is 50% on a certain day, then one square meter silicon solar cell on average will produce 0.28 (1 50%) = 0.14 kw power on that day. The wind power harvested is given by P = 1 2 αaρv3, where α is an efficiency factor (because the factor is generally smaller than 1.0 based on different conditions, in this study we typically set α =0.5), A is the effective windward area, ρ is the density of air, and v is wind speed [24]. In addition, the wind speed might differ considerably at different heights. However, in order to keep the model simple, we have ignored this factor in this study. Finally, we assume that the solar plants can work 12 hours with stable solar energy outputs in a day and the wind turbines can operate all day, i.e., for 24 hours. In order to provide just sufficient renewable energy on a sunny and windy day, we assume that the area of silicon solar panels configured at each node is 100 square meters and the effective windward area configured at each node are 200 square meters. Therefore, according to the weather reports for average sky cover and average wind speed available at the location of each network node from National Weather Service forecasts [28], we can estimate the solar and wind energy that would be generated at each node on that day (i.e., the parameter ES i ). The 14-node 21-link NSFNET network and 24-node 43-link US backbone network (USNET) (shown in Fig. 5) are considered as our test networks. The network nodes and their geographic locations in the two networks are also shown in Table I and Table II, respectively. Table I shows the weather conditions TABLE I WEATHER CONDITIONS AND RENEWABLE ENERGY OUTPUTS ON MAY 1 st, 2013 OF THE NSFNET NETWORK TABLE II WEATHER CONDITIONS AND RENEWABLE ENERGY OUTPUTS ON MAY 7 th, 2013 OF THE USNET NETWORK

7 2100 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 11, JUNE 1, 2014 TABLE III WEATHER CONDITIONS AND RENEWABLE ENERGY OUTPUTS FROM MAY 1 st, 2013 TO MAY 7 th, 2013 AT NODE N1 OF THE NSFNET NETWORK (SEATTLE, WA) (i.e., sky cover and wind speed) and the renewable energy outputs (i.e., sum of the solar and wind energy) on May 1st, 2013 at each node location of the NSFNET network. Table II shows this information for the USNET network on May 7th, From these tables, we can see that the outputs of solar and wind energy on a day vary significantly at different nodes due to different weather condition at each node location. For example, the solar panels and wind turbines at N2 of the NSFNET network which is located at San Diego, CA outputs kWh renewable energy on that day, while N6 which is located at Lincoln, NE generates 1560-kWh renewable energy on that day. Table III shows the weather conditions and renewable energy outputs from May 1st to May 7th, 2013 at node N1 of the NSFNET network (i.e., Seattle, WA). It indicates that the outputs of solar and wind energy vary across different days due to changes in the weather. The traffic demand between each node pair is a uniformly distributed random variable over a given range, i.e., centered at its average as in [7]. For example, given an average demand intensity X = 40 Gb/s, the actual demand between each node pair is generated by a random function uniformly distributed over the range [10, 2X-10], i.e., [10, 70] Gb/s. In this study, we set the average demand intensity for the NSFNET network as X = 40 Gb/s and for the USNET network as X = 20 Gb/s. 2 The capacity of each wavelength is assumed to be 40 Gb/s (i.e., B = 40 Gb/s). According to the Cisco CRS-1 8-slot single-shelf system data sheet [29], the power consumption of a 40-Gb/s router port is 600 W 3 and its energy consumption in a day is 14.4 kwh (i.e., E r =14.4 kwh). The power consumption of a 40-Gb/s transponder is assumed to be 125 W [30] and its energy consumption in a day is 3 kwh (i.e., E t =3kWh). In addition, the maximum number of router ports/transponders deployed at each node (i.e., Δ i ) is set to be 64 in this study 4. 2 The NSFNET network has 14 nodes, which corresponds to 91 node pairs. In contrast, the USNET network has 24 nodes, which corresponds to 276 node pairs. If two networks have the same traffic intensity X, then the total traffic demand on USNET will be much higher than that of NSFNET (i.e., three times). Thus, in order to make the total traffic demands of the two networks comparable (not significantly different), we have set X of USNET to be half of that of NSFNET. 3 Note that this is a simplification of incorporating power consumed by chassis and router backplane into power consumed by router ports. For example, a CRS- 1 fully configured with eight line cards consumes about 4834-W power, so on an average each router port is considered to consume about 600-W power. 4 Note that if the maximum number of router ports/transponders per node is small, some IP traffic would not get served. Meanwhile, it is not cost-efficient to have this number very large. The number 64 was chosen as a practical compromise to ensure that the IP traffic can just be served in the case study. Fig. 6. Total non-renewable energy consumption on different days. (a) NSFNET, (b) USNET. We employed the AMPL/Gurobi software package (version 5.0.0) to solve all the MILP models on a 64-bit server with 2.4-GHz CPU and 8-G memory. The MIPGAP of all the MILP models is 1%. We employed the Java Eclipse platform installed on a regular desktop to implement the heuristic algorithms. VI. NUMERICAL RESULTS A. Total Non-Renewable Energy Consumption Fig. 6(a) shows the total non-renewable energy consumption in the network from May 1st to May 7th for the NSFNET network. MILP-Optimal in the legend indicates the optimal virtual topology configuration based on the Follow the Sun, Follow the Wind strategy. MILP-Energymin is the optimal design approach of minimizing the total energy consumption in the entire network in [7]. Multihop-Bypass is the heuristic applied to minimize the total energy consumption in [7]. Note that the latter two cases do not consider the geographic distribution of renewable energy and also ignore the energy consumption of optical amplifiers and OXCs due to their low contribution to the total energy consumption of the entire network. The heuristic algorithm developed in this paper which is based on the Follow the Sun, Follow the Wind strategy is denoted as Follow- Shuffle-1000 in the legend whose result is the best one selected from 1000 shuffles of the node pair list. Finally, the total nonrenewable energy consumption in the network is measured by summing the non-renewable energy consumption at each node. Note that the non-renewable energy consumption at a node is zero if the node has sufficient renewable energy to support node equipment.

8 SHEN et al.: FOLLOW THE SUN, FOLLOW THE WIND LIGHTPATH VIRTUAL TOPOLOGY RECONFIGURATION IN IP OVER WDM NETWORK 2101 In Fig. 6(a), we can see that the MILP virtual topology configuration based on the Follow the Sun, Follow the Wind strategy always provides a lower bound and the Multihop-Bypass heuristic algorithm gives an upper bound on non-renewable energy consumption of the entire network. Compared to the traditional purely energy-minimized design approach, the Follow the Sun, Follow the Wind method always performs better requiring lower non-renewable energy for both the MILP optimal scheme and the heuristic algorithm. This is evident because the Follow the Sun, Follow the Wind approach considers the state of renewable energy available at each node when establishing optical channels along virtual links. In addition, for the strategy of Follow the Sun, Follow the Wind, the heuristic approach performs close to the corresponding MILP case for the period from May 1st to May 4th, while on the other days, their performance differences seem large. This is because more renewable energy is available in the period from May 1st to May 4th, which makes it easier to plan a relatively efficient lightpath virtual topology than on the other days. For example, consider the two topologies shown in Figs. 4(b) and (d). Although virtual topology (b) has more lightpaths and consumes more total energy, they will consume the same amount of non-renewable energy if there are sufficient renewable energy available at nodes N1 and N2 on that day. In this example, either of the virtual topologies is suitable as both will have the same amount of non-renewable energy consumption. However, virtual topology (b) will certainly consume more non-renewable energy than (d) if sufficient renewable energy is not available at nodes N1 or N2. In this case, only virtual topology (d) can be chosen for lower non-renewable energy consumption. This therefore explains why it is more difficult to plan the virtual topology to achieve the least non-renewable energy consumption from May 5th to May 7th when lower renewable energy is available at each network node on those days. We also compare the performance of the two heuristic algorithms. As shown in Fig. 7(a), the heuristic algorithm based on the Follow the Sun, Follow the Wind strategy can save up to 75% non-renewable energy consumption on a given day compared to the traditional multihop-bypass approach. It can also achieve over 20% less non-renewable energy consumption even on the days which have little renewable energy making it more difficult to plan the lightpath virtual topology. In addition, we can see that the total non-renewable energy consumption and the percentages of non-renewable energy saving on different days vary greatly due to the fluctuation of renewable energy available at different node locations on different days. Similar observations can be made for the larger USNET network as shown in Fig. 6(b) and Fig. 7(b). The result of the heuristic approach is also chosen from 1000 shuffles as in the case of NSFNET. Note that it is generally more difficult to plan the lightpath virtual topology for a larger network. Because of the larger number of nodes, the results of the heuristic algorithm are not as close to the MILP optimal case as in the NSFNET network. Nonetheless, comparing the results of the two heuristic algorithms, we can see that the heuristic algorithm based on the Follow the Sun, Follow the Wind strategy can still lower the non-renewable energy consumption about 68% on Fig. 7. Percentages of non-renewable energy saving by the Follow the Sun, Follow the Wind approach compared with traditional multihop-bypass approach. (a) NSFNET, (b) USNET. May 1st compared to the traditional multihop-bypass heuristic algorithm. In addition, comparing the percentage saving by the heuristic algorithm based on the Follow the Sun, Follow the Wind strategy over the traditional multihop-bypass heuristic approach for the two test networks [see Fig. 7(a) and Fig. 7(b)], it seems that there is no direct association between the percentages of non-renewable energy reduction and network size. The percentages of non-renewable energy saving vary only according to the outputs of renewable energy at different nodes. B. Distribution of Energy Consumption at Each Node Fig. 8(a) shows the relationship between the renewable energy available and the total energy consumption at each node on May 1st 2013 for the NSFNET network. The plot with the dotted line shows the renewable energy outputs at each node on that day. The plot with the solid line shows the total energy consumption at each node for the different approaches. We can see that the Follow the Sun, Follow the Wind approach can serve the network traffic in a good coincidence with renewable energy distribution, i.e., it establishes more lightpaths at a node if it has more renewable energy so as to reduce non-renewable energy consumption at other nodes. For example, node N6 has the highest amount of renewable energy available and therefore establishes the largest number of lightpaths. As expected, the nodes with sufficient renewable energy act as central (aggregator) transit nodes of the network so as to reduce the consumption of non-renewable energy at the other nodes. In contrast to this, the approach which merely minimizes the total energy consumption

9 2102 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 11, JUNE 1, 2014 Fig. 9. Total non-renewable energy consumption with different number of aggregator nodes. (a) NSFNET on May 1 st, (b) USNET on May 7 th. cannot provide this feature as it establishes lightpaths only according to the traffic demand matrix, completely ignoring the state of renewable energy distribution in the network. Fig. 8(b) shows the case of May 7th 2013 which was a day when less renewable energy was available in the NSFNET example. On this day, almost all the nodes need non-renewable energy to satisfy the total energy consumption requirements. However, the plot for the heuristic still follows the plot of the renewable energy output quite closely. Figs. 8(c) and (d) show the corresponding results of the US- NET network on May 1st and May 7th, respectively, from which observations similar to those for the NSFNET network can once again be made. Fig. 8. Relationship between renewable energy available and the total energy consumption at each node.(a) NSFNET on May 1 st, (b) NSFNET on May 7 th, (c) USNET on May 1 st, (d) USNET on May 7 th. C. Impact of Number of Aggregator Nodes The proposed heuristic algorithm (i.e., Follow-Shuffle-X heuristic algorithm) in Section IV contains a step to select M nodes which have the highest amounts of renewable energy and establishes direct lightpaths involving these M nodes to make them act as aggregator nodes. This choice must be judiciously made. Choosing too few aggregator nodes may lead to a situation where full use of the available renewable energy cannot be made. On the other hand, choosing too many such nodes will lead to a highly connected virtual topology which also cannot make full use of the renewable energy available at these nodes. Thus, it is generally difficult to predict how many such aggregator nodes would be needed. Fig. 9(a) shows the total non-renewable energy consumption versus the number of aggregator nodes chosen on May 1st 2013

10 SHEN et al.: FOLLOW THE SUN, FOLLOW THE WIND LIGHTPATH VIRTUAL TOPOLOGY RECONFIGURATION IN IP OVER WDM NETWORK 2103 Fig. 11. The average numbers of IP hops traversed by IP traffic in different virtual topology design cases. Fig. 10. The number of aggregator nodes which can achieve the lowest nonrenewable energy consumption on different days. (a) NSFNET, (b) USNET. for the NSFNET network. The results were obtained by the heuristic algorithm with the Follow the Sun, Follow the Wind strategy and considering 1000 node pair shuffles. We can see that when the number of aggregator nodes is 14 (which would imply a fully connected virtual topology), the non-renewable energy consumption is the highest. On the other hand, the consumption of non-renewable energy is also high if there are no aggregator nodes in the network (i.e., the number of aggregator nodes is zero). In Fig. 9(a), the lowest non-renewable energy consumption is achieved at the point where four nodes are chosen as the aggregator nodes. Similar observations can be made for the USNET network on May 7th as in Fig. 9(b). In this case, the non-renewable energy consumption is the lowest when the number of aggregator nodes is five. Fig. 10 shows the number of aggregator nodes which can achieve the lowest non-renewable energy consumption on different days for the two test networks. For the NSFNET network, we can see that the number of aggregator nodes varies from 1 to 4 to achieve the lowest non-renewable energy consumption on different days. Similarly, this number varies between 5 and 6 for the USNET network. Note that the USNET network needs more aggregator nodes as it is bigger. We can also conclude that the exact number of aggregator nodes which can achieve the lowest non-renewable energy consumption cannot be predicted, but an approximate range may be found based on the historical data available. We also consider the influence of using the nodes with renewable energy as aggregator nodes on the average number of IP hops traversed by IP traffic demand. As a cost paid for reducing non-renewable energy consumption, IP traffic tends to traverse the aggregator nodes for maximal use of renewable energy. This can cause detours of the traffic flows, thereby increasing the number of traversed IP hops. Fig. 11 compares the average number of IP hops traversed by the IP traffic in the different network cases. The heuristic algorithm based on the Follow the Sun, Follow the Wind strategy is denoted as Follow-Shuffle-1000 (M aggregator nodes) in the legend, which means that each of the results is the best one chosen from 1000 node-pair shuffle results. M is the number of aggregator nodes, under which the lowest non-renewable energy consumption is achieved for the entire network. According to Fig. 9, M is four for the NSFNET network and five for the USNET network. We can see that IP traffic routed by the Follow the Sun, Follow the Wind heuristic algorithm on average traverses a larger number of IP hops than that of the traditional Multihop-Bypass design (i.e., 30% and 10% more hops respectively for the NSFNET and USNET networks). This therefore verifies the cost paid for reducing non-renewable energy consumption through using the aggregator nodes. D. Impact of Node Pair Shuffle Times As discussed before, the number of node pair shuffles would significantly affect the performance of the Follow the Sun, Follow the Wind heuristic algorithm. In general, more shuffles can lead to a better result, and for a larger network size, a larger number of shuffles is required for a reasonable performance. Fig. 12(a) shows how the total number of shuffles affects the total non-renewable energy consumption on May 7th for the NSFNET network. Obviously, the non-renewable energy consumption decreases with the increase of shuffle times. However, because the computation time is proportional to the total number of shuffles, it is important to consider the tradeoff between the performance of the heuristic algorithm and the shuffle times when implementing this strategy. Similar observations can also be made from the result of the USNET network on May 1st as shown in Fig. 12(b). As a reference of the computational complexity, a typical desktop takes about half an hour to run the Follow the Sun, Follow the Wind heuristic algorithm when the shuffle times are 10,000 for the USNET network. Thus, the computation

11 2104 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 11, JUNE 1, 2014 Fig. 12. Total non-renewable energy consumption with different shuffle times. (a) NSFNET on May 7 th, (b) USNET on May 1 st. complexity could become an issue when implementing this type of shuffle strategy. To reduce the computation time for an even larger number of shuffles, we may consider a parallel computing approach based on the cloud computing technique to distribute a certain number of shuffles evenly onto a set of parallel servers and let them run the shuffles independently and finally collect their outputs to select the best result. The detail on the cloud-computing-based shuffle optimization process can be found in [31]. VII. CONCLUSION AND FUTURE WORK Renewable energy such as solar and wind energy is introduced at each node location to reduce non-renewable energy consumption. Also, a Follow the Sun, Follow the Wind operational strategy, under which the lightpath virtual topology is periodically reconfigured so that more lightpaths start or end at nodes where maximum renewable energy is available, is proposed. Based on such a strategy and the daily weather forecast data, we developed an MILP model to optimally design the lightpath virtual topology for the IP over WDM network by considering the available state of renewable energy at each network node location. We also develop a simple but efficient heuristic that incorporates the shuffle of node pair lists to design the lightpath virtual topology. The simulation results indicate that the proposed Follow the Sun, Follow the Wind strategy can save up to 75% or more non-renewable energy in the considered example scenarios compared with the traditional pure energy-minimized design approach. Also, it is found that the Follow the Sun, Follow the Wind approach can serve network traffic according to the renewable energy distribution, i.e., it establishes more lightpaths at a node with more renewable energy so as to reduce non-renewable energy consumption at other nodes. In addition, an optimal number of aggregator nodes which can ensure the lowest non-renewable energy consumption in the entire network is observed for a network operated under the Follow the Sun, Follow the Wind strategy. Finally, for the heuristic algorithm based on the Follow the Sun, Follow the Wind strategy, we find that its performance is dependent on the number of node pair shuffles. A larger number of shuffles are expected to achieve better performance, but a saturated trend is also observed that the performance improvement becomes marginal if a sufficient number of shuffles have been evaluated. As the key limitations of the current research, the Follow the Sun, Follow the Wind strategy was verified based only on the case studies of a single week and two test networks. The study also made some simplified assumptions such as identical, daily average solar capacity and ideal wind turbine efficiency factor. In the future works, we will extend the current approach to incorporate more realistic conditions and to evaluate the impacts of the aspects such as different sizes of renewable installations, geographic differences in insolation, different efficiency factors of wind turbines, etc. Some sensitivity analyses can also be performed to investigate the impacts of these aspects. We may also extend the current study from one week to one year (or even several years) to obtain yearly average savings. Finally, because manufacturing of network node equipment also contributes to GHG emissions, it is meaningful to employ the Life Cycle Assessment (LCA) process to evaluate the performance of the Follow the Sun, Follow the Wind strategy. ACKNOWLEDGMENT The authors would like to thank Prof. Mukherjee and Dr. Dikbiyik from UC Davis for providing the geographic information of the USNET test network. They would also like to thank the reviewers for their valuable comments that helped improve this paper. REFERENCES [1] An inefficient truth. (accessed Apr. 1, 2014). [Online]. Available: %20Full%20Report.pdf [2] ICT industry combats climate change. (accessed Apr. 1, 2014). [Online]. Available: [3] J. Malmodin, A. Moberg, D. Lunden, G. Finnveden, and N. Lovehagen, Greenhouse gas emissions and operational electricity use in the ICT and entertaiment & media sectors, J. Ind. Ecol., vol. 14, no. 5, pp , Oct [4] C. Stewart and K. Shen, Some joules are more precious than others: Managing renewable energy in the datacenter, presented at the Workshop Power Aware Comput. Syst., Big Sky, MT, USA, Oct [5] K. K. Nguyen, M. Cheriet, B. S. Arnaud, V. Reijs, A. Mackarel, P. Minoves, A. Pastrama, and W. Van Heddeghem, Renewable energy provisioning for ICT services in a future internet, in The Future Internet (Lecture Notes in Computer Science). New York, NY, USA: Springer, 2011, pp [6] M. Gattulli, M. Tornatore, R. Fiandra, and A. Pattavina, Low-carbon routing algorithms for cloud computing services in IP-over-WDM networks, in Proc. IEEE Int. Conf. Commun., 2012, pp