RECENT statistics showed that the usage of the Internet, a

Transcription

1 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 20, OCTOBER 15, Towards Green Optical/Wireless In-Building Networks: Radio-Over-Fiber Apurva S. Gowda, Ahmad R. Dhaini, Leonid G. Kazovsky, Hejie Yang, Solomon T. Abraha, and Anthony Ng oma Abstract Energy efficiency has become a major paradigm in the design and operation of future telecommunication networks. Recent studies show that the aggregate power consumption of inbuilding IT networks (residential and office) is massive and comparable with that of data centers due to the large number of buildings. In this paper, we analyze the energy efficiency of next-generation in-building IT networks to deliver high-speed mobile access to end users via integrated optical/wireless networks using Radio-over- Fiber (RoF) technology. Based on a validated energy efficiency model, our results show that although individual point-to-point RoF links are not as energy efficient as legacy Baseband-over-Fiber links, RoF networks may actually be more energy-efficient when designed keenly with small cells sizes and when the static energy consumption of the remote units is above a particular threshold. Under the assumptions used in this paper, we show that the DRoFbased architectures can be designed to more energy efficient for cell sizes <17 m. Index Terms Energy efficiency, in-building networks, optical communication, radio-over-fiber, wireless LAN. I. INTRODUCTION RECENT statistics showed that the usage of the Internet, a reliable communication tool in today s society, increases by 40% per year in North America and by about 20% worldwide [1]. Recent studies also showed that the Internet currently consumes times more energy than the minimum required [2]. The design of current networks aims at meeting strict performance requirements such as bandwidth and deployment cost constraints with little regard to energy efficiency. However, with the increasing demand for bandwidth due to the proliferation of bandwidth-hungry multimedia applications, by year 2025, the amount of energy required to keep telecom networks running is expected to increase by two orders of magnitude [3]. Recognition of the increase in network energy consumption, along with the rising cost of energy and the harmful effect Manuscript received January 15, 2014; revised March 10, 2014; accepted April 3, Date of publication April 6, 2014; date of current version September 1, This work was supported by Corning, Inc.; National Science Foundation (NSF) under Grants CNS and CNS ; and the Natural Sciences and Engineering Research Council of Canada (NSERC). A. S. Gowda, A. R. Dhaini, and L. G. Kazovsky are part of the Photonics and Networking Research Lab, Stanford University, Stanford, CA USA ( asgowda@stanford.edu; adhaini@stanford.edu; l.kazovsky@ stanford.edu). H. Yang, S. T. Abraha, and A. Ng oma are with Corning Incorporated, Corning, NY 14831, USA ( YangH4@corning.com; abrahast@corning.com; NgomaA@corning.com). Color versions of one or more of the figures in this paper are available online at Digital Object Identifier /JLT Fig. 1. Daily energy consumption of global optical transport, global core and metro optical access networks [3], and in-building networks in the U.S. only. of CO 2 emissions, has led to extensive research in academia and industry to realize energy-efficient (green) communication networks. Research efforts that address the foregoing problem mainly focus on reducing the energy consumption of data centers since they are considered the major energy hog in the Internet. However, recent studies reveal that the power consumption of data centers and that of in-building networks (residential and office) are on a comparable scale due to the large number of buildings and the spreading Internet literacy of their users [4]. According to [5], by year 2017, 55% of all IP traffic will originate from wireless and mobile devices. It has been predicted that by year 2017, two-thirds of all IP traffic will be consumer traffic from households, university populations and Internet cafes. This means that the source of extensive wireless Internet traffic will be from in-building IT networks. Thus, it is important to explore ways for reducing the energy consumption of office and residential building networks. Due to the large number of these types of networks, even a small decrease in the energy consumption in each of the individual networks will lead to a significant improvement in the overall energy consumption. As depicted in Fig. 1, the energy consumption of in-building networks in the U.S. alone is currently higher than the energy consumption of both the global optical transport and wide/metro area networks worldwide. This is due to the large number of office and residential buildings in the U.S. about 17% of the total commercial floor space in the U.S. is office buildings [6]. Although it is commonly believed that the energy consumed by space heating, lighting, etc., in office and residential buildings far trumps that of IT networks, the design of current IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See standards/publications/rights/index.html for more information.

2 3546 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 20, OCTOBER 15, 2014 Fig. 3. In-building network architecture (with three floors and a basement). Fig. 2. Energy Usage in U.S. in-buildings [6]. in-building networks aims at meeting capacity and performance requirements with little regard to energy efficiency. Thus, as the demand for bandwidth increases, the network design would require more complex technologies. This drastically increases the energy consumption. Fig. 2 shows how with the absence of energy efficient solutions, by year 2020, the energy consumption of IT networks would surpass all other power electronics in U.S. buildings. In this paper, we investigated radio-over-fiber technologies for achieving energy-efficient optical/wireless converged in-building networks. We proposed and analyzed three different fiber-based in-building architectures; ARCH1 is the legacy architecture (i.e., using baseband-over-fiber); ARCH2 and ARCH3 centralize the DSP processing but have dedicated links for each cell. We developed extensive theoretical and simulation models and used experimentally measured data to validate key parts of the models. Our results show that although individual point-to-point RoF links may not be as energy efficient as legacy baseband-over-fiber links, there are some RoF network architectures, which may actually be more energy efficient when designed keenly with small cell sizes and when the static energy consumption of the remote units is above a particular threshold. The remainder of the paper is organized as follows. In Section II, we provide an overview of in-building networks and widely employed technologies. Section III presents the energy efficiency analysis based on the modeling of point-to-point links. In Section IV, we present the network architectural analysis. In Section V, we present our results. We finally conclude the paper in Section VI. II. OVERVIEW ON IN-BUILDING NETWORKS AND TECHNOLOGIES In general, communication networks can be divided into outdoor and indoor networks. Outdoor networks comprise core, metro/edge and access networks, while indoor networks generally consist of data centers, and residential and office building IT networks. In this section, we present an overview of in-building IT networks. Specifically, we describe the network architecture of in-building networks, their energy consumption and the related art for making them energy efficient. A. In-Building IT Network Architecture As illustrated in Fig. 3, a typical in-building IT network consists of a main distribution frame (MDF) that connects the inbuilding network to the outside world (typically the access network), and an intermediate distribution frame (IDF) that connects the networks on various floors to the MDF. The MDF and IDFs are inter-connected via optical Ethernet connections, and on each floor wireless access points (WAPs) provide the users wireless access; whereas wired (copper Ethernet) links provide the users wired access to the Internet. A residential network can be considered a simplified version of this network, e.g., the network on a single floor. B. In-Building Network Technologies: Baseband-Over-Fiber and Radio-Over-Fiber In recent years, the integration of fiber and wireless networks has been presented as the evolutionary next step toward realizing high-speed, wireless and cost-effective broadband access networks. The legacy architecture utilizes baseband-over-fiber (BoF), which consists of sending the baseband data over an optical link and subsequently over the wireless link. Typically, it is realized using separate protocols for the optical link and wireless link. With the proliferation of mobile bandwidth intensive in-building networking devices, there is an unprecedented need for pushing this integration closer to the user. Radio-overfiber (RoF) is a promising technology that has been actively researched as a strong candidate for achieving optical-wireless convergence. RoF enables the wireless signal to directly ride on the optical carrier. Owing to the high bandwidth available in the optical link, multiple wireless services can use the same backhaul optical network. Besides high bandwidth, RoF allows for centralized network control and simple remote antenna units, which facilitates easier maintenance, upgrade, and sharing of resources.

3 GOWDA et al.: TOWARDS GREEN OPTICAL/WIRELESS IN-BUILDING NETWORKS: RADIO-OVER-FIBER 3547 There are several technologies for sending the wireless signal over the optical fiber link. Namely, analog RoF (ARoF), digitized RoF (DRoF), intermediate frequency-over-fiber and digitized intermediate frequency-over fiber [9]. In an ARoF link, the remote unit has a very simple structure; however, when sending an analog signal on the optical carrier, the optical link will have a limited dynamic range and will be affected by fiber nonlinearity, noise and chromatic dispersion. These limitations are relaxed when sending a digital signal over the optical link. Thus, DRoF is more reliable on the optical link. However, digitizing a radio frequency signal requires band-pass sampling and more analog-to-digital and digital-to-analog converters (ADC/DAC), which introduces quantization, aliasing and jitter noise. Also, the remote unit will have more circuitry. C. Literature Review Several papers investigated the power consumption of ARoF and DRoF systems [10] [13]. The authors of [9], [13], [15] studied the energy consumption of ARoF and DRoF as backhaul for mobile networks. The models assume that the energy consumption of the analog and digital circuitry is given and estimate the power consumption of the power amplifier using a path loss model and the drain efficiency. The authors of [11], [12] also considered a similar energy consumption model. These models estimate the change in power consumption with higher path loss effectively but do not take into account the effect of power losses and noise on the link. To account for both power losses and noise on the link, a model based on the signal-to-noise ratio (SNR) or bit-error-rate (BER) is required.the authors of [14] analyzed the SNR and spurious free dynamic range (SFDR) of an ARoF link. The work in [15] provided a comparison between ARoF and DRoF based on the SNR on the optical link; it studied the different noise sources on the two links and the higher effect noise has on an ARoF link. Finally, [16]also showed the effect of optical-electrical-optical (OEO) losses and noise on the performance of an ARoF link. In a RoF link, the radio frequency (RF) signal travels through the optical link and is directly transmitted over the air and demodulated at the wireless receiver. To quantify the effect of noise and losses on the energy consumption, a more comprehensive SNR model for the whole link, from the wireless transmitter to the wireless receiver, is required. The SNR model can then be used for receiver sensitivity calculations and subsequently energy consumption calculations. In this paper, we provide a comprehensive energy consumption model based on the SNR of the entire link that considers losses such as electricaloptical-electrical (EOE) losses, loss due to sinc behavior of DAC (DRoF), etc., for both ARoF and DRoF. We use the model to compare in-building architectures based on BoF, ARoF, and DRoF. III. POINT-TO-POINT LINK ANALYSIS To analyze the energy consumption of the green architectures, we perform a point-to-point link analysis on the technologies used in each of the architectures, i.e., ARoF and DRoF. In our previous work [17], we use a preliminary model to perform the Fig. 4. Generic point-to-point block diagram. analysis. However, the model does not account for the effect of: 1) optical link loss on the downlink on ARoF; 2) digital to analog conversion power loss due to sample-hold circuit in DRoF. In this paper, we extend the model to include the effect of such losses. Fig. 4 shows a generic point-to-point link. The RF transmitter (TRx) and the head end unit (HEU) are located at a central location (i.e., at the MDF in the proposed architectures). The RF TRx modulates/demodulates the RF signal and performs the necessary signal conditioning. The HEU contains the EOE conversion circuitries required to transmit/receive the RF signal on the optical link, and the RAU contains the EOE circuitries and the analog signal conditioning circuitries. In this section, we present a model to estimate the power consumption of point-to-point analog links. We consider two cases: 1) Estimate Transmitter s Power Amplifier Gain Based on BER: The BER at the RAU is set to a certain value and the required power of the signal is estimated at different points in the transmitter, thus determining the gain of the power amplifier. Using the gain, the power consumption of the power amplifier is estimated. Energy per bit is used as the comparison metric. 2) Fix Transmitter Design and Estimate Wireless Transmission Distance Based on BER: The power at the transmitter is known and based on the transmitter design, the power at different points from the transmitter to the receiver is estimated. The energy efficiency metric used is energy per bit per unit area to normalize with respect to transmission distance. In this case, we focus on the downlink scenario. The uplink scenario is left for our future work. The analysis of the link and architectures are done by estimating the power amplifier gain (Case 1), since we can obtain the effect of cell size directly. However, experimentally validating the loss equations using this approach is challenging. Thus, we validate the use of the loss equations experimentally by fixing the transmitter design and estimating the wireless transmission distance (Case 2). Since the same equations are used in both cases, the related assumptions are valid for the first approach as well (Case 1). In this study, we focus on carrier frequencies in the UHF and lower part of the SHF range (300-6 GHz), since most current wireless standards specify protocols in this range. The maximum frequency depends on the maximum bandwidth of the available optical components. However, generally, for protocols in the EHF range more complex optical link modulation/demodulation techniques are required; this is outside the scope of this work. By applying the analysis on the network level, we can then measure the power consumption of the proposed green network architectures. A summary of the most relevant notations used in the analysis is given in Table I. The rest are defined in the

4 3548 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 20, OCTOBER 15, 2014 TABLE I SUMMARY OF IMPORTANT NOTATIONS Parameter Power of the received RF signal Responsivity of the photo-detector Slope efficiency of the laser MZM bias voltage MZM half-wave voltage Load resistance Optical carrier power Length of the fiber Fiber loss per km Power amplifier gain Path loss Gain of the antenna Noise factor Boltzman s constant Relative intensity noise Charge of an electron Bandwidth of the signal Room Temperature Wavelength of carrier frequency Wireless transmission distance Low noise amplifier gain Sampling Frequency Constellation Size Upper frequency bound Root mean square of clock jitter Abbreviation P Rx (W) R (A/W) s l (W/A) V b (V) (V) R l (Ω) P opt (W) L (km) α G pa G pl G a F k (J/K) RIN (db/hz) q (C) B (Hz) T (K) λ (m) d (m) G lna f s (Hz) M f h (Hz) σ τ (s) Fig. 5. Analog RoF, ARoF. (a) Downlink. (b) Uplink. TABLE II POWER CONSUMPTION OFVARIOUSCOMPONENTS Component Power Consumption (mw) RF Analog Transceiver [18] [20] 400 Ethernet Transceiver [4] 1000 Laser Driver [21] 10 Transimpedance Amplifier [22] 250 Band-pass sampling ADCs+DACs [23] 80 V si bias (Laser source) [21] 7.8 Wireless DSP (modulation+demodulation) 60.2 paper. Table II lists the power consumptions assumed for key components throughout this paper. A. Analog Radio-Over-Fiber: Typical ARoF downlink and uplink systems are shown in Fig. 5. Each consists of a RF analog transceiver chain that generates the RF signal. The receiver on the downlink is the same as that of the BoF link (i.e., user s receiver). However, on the uplink, detection of the wireless signal occurs after the optical link. Hence, the SNR on the uplink differs from that on the downlink. 1) ARoF Estimating Power Amplifier Gain: The power consumption of the power amplifier is estimated given a certain drain efficiency and the required transmitting power to maintain the desired BER at the receiver. On the downlink, a high frequency signal is directly modulated onto an optical carrier, which is received at the RAU. After O/E conversion and bandpass filtering, the signal is transmitted over the air to the enduser. Due to the optical link, the RF signal is subject to 1) power loss due to modulation onto the optical carrier, 2) thermal noise due to analog components at the transmitter and receiver, and 3) relative intensity noise (RIN) and shot noise due to the EOE conversion components. The optical transmission can give rise to higher-order harmonics in the signal due to inherent nonlinear effects. However, since the focus of this study is limited to in-building networks where the length of optical links is typically less than 1 km, with low bit-rates and low launch power [24], we do not consider dispersion or non-linearity on the optical link. In this study, we assume direct modulation onto the optical carrier. On the downlink, the average SNR of the received RF signal at the end-user is given by SNR down = P Rx, (1) N tot,down N tot,down = FkTB where N tot,down is the thermal noise due to the analog components in the RAU. We use a simplified path loss model with exponent l which is given as follows: ( ) 2 Ga λ G pl =. (2) 4πd l/2 Throughout this paper, we assume isotropic antennas, thus G a =1. The required average SNR to maintain a certain BER can be calculated using the BER equations for an additive white Gaussian noise channel (AWGN) [25]. Using direct detection of small signal model of an impedance matched system [26], the loss in RF power due to the optical link can be approximated as m 2 = e αl (s l R) 2. The required received power P Rx can be calculated as P Rx = G pl m 2 (G pa P in ) (3) where P in is the power of the RF signal after performing analog signal conditioning at the MDF (shown in Fig. 4).

5 GOWDA et al.: TOWARDS GREEN OPTICAL/WIRELESS IN-BUILDING NETWORKS: RADIO-OVER-FIBER 3549 Using (1) and (3), we can solve for the required output power of the power amplifier, P Tx = G pa P in, that compensates the power loss due to the optical link and the path loss. For a given drain efficiency, ρ, and back-off, η, ofapower amplifier, the transmission energy consumption per bit (power amplifier + over the air energy loss) of the wireless link, E Tx can be calculated as ( ) η PTx E Tx = ρ B. (4) On the uplink, the RF signal received from the end-user is modulated onto the optical carrier at the RAU. The average SNR of the RF signal at the receiver is given by SNR up = m2 G lna P Rx, (5) N tot,up N tot,up =4FkTB+ ( R 2 ) R l P 2 opt 10 RIN/10 B +(2qRP opt )B. For the uplink, P Rx represents the required received power from which we can calculate the required transmit power P Tx as follows: P Tx = P Rx. (6) G pl For a given drain efficiency, ρ, and back-off, η, of the power amplifier, the transmission energy consumption per bit of the wireless link can be calculated using (4). The power consumption of the laser can be calculated using the dc voltage, V s, and the dc current, I bias plus the laser driver power consumption, P LD, as follows: P laser = V s I bias + P LD. (7) The power consumption of the photodetector can be calculated from the drain voltage, V DD, incident average optical power, P opt, and the responsivity, R PD, as follows: P PD = V DD P opt R PD. (8) The energy consumption per bit of the EOE conversion circuitry is given by E OEO = P laser + P PD + P TIA. (9) R b where R b is the wireless bit rate. Finally, the total energy consumption per bit is given by E ARoF = E Tx + E RFTX + E RFRX + E DSPTX + E DSPRX + E OEO. (10) where E RFTX, E RFRX, E DSPTX and E DSPRX are the analog circuit energy and digital circuit energy consumptions of the wireless transmitter and receiver, respectively. To estimate the energy consumption of DSP circuits we first identify and estimate the necessary DSP operations and computational complexity of each block in the DSP circuitry. From the energy consumption of a logic transition in a metaloxide-semiconductor field-effect transistor (MOSFET), the energy consumption of various gates such as an inverter, AND, NOR, etc., can be calculated (using the transistor level design diagrams). We assume an activity factor as calculated in [27]. Using these values, the energy consumption of basic digital circuits such as a multiplier, an adder, etc., can be calculated from simple logic circuit diagrams. The energy consumption of a complex DSP operation is assumed to be the linear sum of energy consumptions of smaller DSP operations, such as multipliers, adders, AND/OR logic, etc. The energy consumption of a CMOS transistor implemented with 32 nm CMOS process is estimated as 3 fj [28]. The DSP blocks considered include the FFT/IFFT block, precoder, encoder/decoder and scrambler/descrambler. The model estimates the dynamic power consumption of the DSP and is as given in Table II. 2) ARoF Estimating Wireless Transmission Distance: We assume that the power of the signal modulated on to the optical carrier, P in, is known. In the case of direct modulation of the laser, the radiated signal power is given by P Tx = G HEU m 2 G RAU P in (11) where G HEU and G RAU represent the gain due to the components in the HEU and RAU, respectively. We also consider the case of externally modulating the intensity of the laser using an MZM with balanced drives. When modulating a sinusoid of amplitude V in, it can be shown that the output power after the photodetector is ( Popt e αl ( πvb P optrf = sin 4π ) ( ) 2 2πVin J 1 R) R l (12) where J n () is the nth-order Bessel function of the first kind. The detailed derivation is given in Appendix. Thus, P Tx = G RAU P optrf. (13) Assuming a certain receiver sensitivity, P Rx, and a path loss model, the maximum transmission distance, d, can be calculated using a path loss model [say (2) and (6)]. The energy efficiency metric is then given by: E b/m 2 = P DC R b πd 2 (14) where P DC is the dc power consumption of the components used in the system and, R b is the wireless bit-rate. B. Digitized Radio-Over-Fiber Typical digitized radio-over-fiber (DRoF) downlink and uplink systems are shown in Fig. 6. In a DRoF link, the high frequency signal is digitized and the digital signal is modulated onto the optical carrier. Digitizing high frequency signals requires the use of band-pass sampling ADC s and DAC s [15]. 1) DRoF Estimating Power Amplifier Gain: On the uplink, the RF signal is affected by aliasing noise due to band-pass sampling, jitter noise of the ADC, quantization noise, noise on the optical link and the thermal noise of the receiver components. These noise sources are weakly correlated for a digitized signal and can be assumed to be independent [15]. The average SNR

6 3550 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 20, OCTOBER 15, 2014 The term sinc 2 (f c /f s ) represents the RF power loss due to up-conversion using zero-order hold DAC. Using (4), we can obtain the corresponding energy consumption E tx. On the downlink, the main noise contributor is the receiver s thermal noise. Thus the average SNR of the received RF signal is given by Fig. 6. Digitized RF over Fiber. (a) Downlink. (b) Uplink. of the received RF signal on the uplink is given by [15]: 1 SNR up = where 1 SNR WL SNR PD SNR A SNR Q SNR J (15) R 2 Popt 2 SNR PD = (FkT +(R 2 R l ) Popt10 2 RIN/10 +2qRP opt )B SNR WL = P RF FkTB 1 SNR A = ((f h /B) 1) FkTB SNR Q =6.02n + 10log ( M +1/ M 1 ) + 10log (f s /2f h ) 1 SNR J = 4π 2 f 2 στ 2. SNR PD SNR A and SNR J are the SNR values due to the photodetector, the aliasing noise and ADC jitter noise, respectively; a detailed description can be found in [15]. SNR Q is the quantization SNR in db, described in [29]. SNR WL is the SNR of the wireless link. The denominator represents the thermal noise of the receiver. The power amplifier has to compensate for the digital-to-analog conversion loss and the path loss. Assuming the peak-to-peak voltage of the ADC/DAC are equal, the required output power of the power amplifier is given by P Tx = G pa P in = P RF sinc 2 (f c /f s ) G lna G pl. (16) ( ) sinc 2 fif f s G lna G pl G pa P in SNR down =. (17) FkTB For a given BER, E Tx can be calculated, using (4). Finally, the total energy consumption per bit, E DRoF, is calculated by summing up the analog circuit energy and digital circuit energy consumptions of the wireless transmitter and receiver, respectively [similar to (10)]. 2) DRoF Estimating Wireless Transmission Distance: In the case of DRoF, the transmitted power depends on the parameters of the DAC. We assume reliable transmission of bits on the optical link and the RF signal spans the full peak-to-peak of the ADC. Since we consider only the downlink with optical links <1 Km, these are valid assumptions. Under these assumptions, the transmitted power is given by ( ) G RAU V 2 PAPR DAC10 10 sinc 2 fif f s P Tx = (18) R l where PAPR is the peak to average power ratio. Similar to the analysis presented for ARoF links, from the estimated transmitter power, the corresponding energy per bit per unit area for a known receiver sensitivity, P Rx, is calculated using a path loss model and (14). IV. GREEN OPTICAL/WIRELESS IN-BUILDING IT NETWORK ARCHITECTURES In current in-building architectures, wired and wireless networks work independently; each with its own protocols. However, the integration of these two networks (i.e., optical/wireless convergence) would realize a much more flexible architecture with higher end-user data rates. The RoF technology allows for a seamless integration by eliminating wired transmission protocol processing [4]. RoF-based networks (not links) can with proper design be more energy-efficient. RoF allows for centralized resources with simple remote antenna units (RAUs). Despite the fact that RoF links are not more energy efficient than baseband-over-fiber (BoF) links, which limits the benefits of centralized architectures, simpler RAUs yield lower circuit energy consumption (i.e., the penalty of using smaller cells and increasing the number of RAUs will be lower [9]). Furthermore, moving from a copper medium to an optical one reduces energy consumption and allows for higher data rates [4]. Therefore, in our proposed green architectures, we design an optical/wireless in-building network based on the RoF technology. We propose and analyze the following three architectures, depicted in Fig. 7. 1) ARCH1, Distributed Processing Using BoF: This is the legacy architecture [Fig. 7(a)]. It employs wireless access

7 GOWDA et al.: TOWARDS GREEN OPTICAL/WIRELESS IN-BUILDING NETWORKS: RADIO-OVER-FIBER 3551 Fig. 8. Experimental Block diagram. The RoF-based architectures (ARCH2, ARCH3) were designed to be compliant with the common architectures employing RoF [30], [31]. Fig. 7. Green architectures (a) ARCH1. (b) ARCH2. (c) ARCH3. points, each with its own processing circuitry, connected to the main distribution frame (MDF) via 1 Gbps optical Ethernet links. 2) ARCH2, Centralized and Dedicated Analog Processing and Centralized and Shared Digital Processing Using ARoF: In this architecture, all RAUs share a DSP unit but have dedicated analog transceiver chains at the MDF [Fig. 7(b)]. ARCH2 employs ARoF links between the RAU and the MDF. 3) ARCH3, Centralized and Dedicated Analog Processing and Centralized and Shared Digital Processing DRoF: As illustrated in Fig. 7(c), this architecture is similar to ARCH2, but it employs DRoF links between the RAU and the MDF. ARCH2 and ARCH3 share the DSP unit unlike ARCH1. In terms of the resource utilization and management, ARCH2 and ARCH3 would be more efficient. However, the architectures differ in the requirements of these components. Using the model presented in Section III, we analyze the effect of these differences on energy consumption of the architectures. V. RESULTS AND ANALYSIS A. Experimental Results In this section, we present experiments conducted to validate our point-to-point link analysis. The block diagram of the experimental setup is shown in Fig. 8. Here, the RF signal generation and demodulation are conducted off-line using MATLAB. An OFDM signal of sub-carrier bandwidth 732 khz and 26 subcarriers (B = 20 MHz) is considered. Each carrier modulated for 16-QAM symbol is generated and upconverted to a carrier frequency, f c = 2.44 GHz, offline in MATLAB. An FFT size of 2048 and cyclic prefix of 64 is used. The resulting bit-rateis 76 Mbps. Training symbols are used for channel estimation. The channel is re-estimated every OFDM symbol by re-encoding the decoded bits and comparing with the received OFDM symbol. For ARoF, the analog signal generated is directly uploaded onto the arbitrary waveform generator (AWG), which is subsequently modulated onto the optical carrier using a single drive MZM and transmitted through a 500 m single-mode fiber (SMF). After photodetection, the signal is amplified and captured using a digital signal oscilloscope (DSO). The captured data is demodulated offline using MATLAB. In the case of DRoF, the generated analog signal is digitized off-line and the digital signal is loaded onto the AWG. The digital signal is also externally modulated onto the optical carrier, transmitted through a 500 m SMF, photo-detected, amplified and captured using the DSO. The RF signal is reconstructed and demodulated offline. The system is optimized to obtain the best performance in both setups. For ARoF, we get an error vector magnitude of 1% and for DRoF, %. To observe the effect of different losses in the system, we vary the input power injected into the photo-detector. We also consider the effect of different Nyquist zones using an intermediate frequency (IF) and by up-converting to higher carrier frequency using a local oscillator (LO) at RAU. In the experiments, we keep the RF input power constant and measure the output RF power. In the case of DRoF, the RF power measurement is conducted in MATLAB, since the DAC was implemented in MATLAB. For

8 3552 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 20, OCTOBER 15, 2014 TABLE III MODEL PARAMETERS [15] Parameter Value kt (N 0 ) -204 dbw/hz R l 50Ω F 3.3 db ρ 20% η 10 RIN -174 db/hz G lna [33] 19 db G a 1 V DD 2.8 V R PD [34] 0.65 A/W n 10 σ τ 0.3 ps f s 300 MHz Fig. 9. Energy per bit versus PD input power. DRoF, we use a sampling frequency, f s = 300 MHz, and a 10 bit resolution; this results in a data rate of 3 Gbps. To calculate the receiver sensitivity, the noise floor, P N = FN 0 B, is assumed. A required SNR of 19 db is also assumed, since in our experiments we use 16QAM modulation. We also assume a receiver noise figure of 5dBand margin of 5dB. Subsequently, the receiver sensitivity, P Rx, is calculated as P Rx = P N (dbm). (19) The noise floor is calculated as P N = log(B) (dbm). (20) The IEEE n in-building path loss model [32] is used to calculate the maximum distance; it is given by PL = P Tx P Rx = log(d)+20log(f c ) (db). (21) From Fig. 9, we also observe that the sinusoidal approximations in the analytical model closely follow the experimental results for an OFDM signal in both ARoF with an external modulator and DRoF. However, there are some deviations at very low PD input powers. In the case of ARoF, at low input PD powers the system is noise dominated. Since transmitter noise is not considered when estimating the wireless transmission distance, the model does not match the experimental results. In the case of DRoF, at low PD input powers the optical transmission is not reliable; thus the RF signal is distorted and the model predictions are off. Excluding the noise-limited regions, the average error is within 5%. These results show that the model can provide adequate insights into the behavior of energy consumption with different parameters. This flexibility allows us to estimate the energy consumption behavior at the network level by scaling the model, thereby measuring the energy consumption of the proposed green in-building IT architectures. B. Point-to-Point Link Results Theoretical Using the model, we can estimate the gain of the power amplifier required to maintain a certain BER in BoF, ARoF and DRoF links. Assuming certain typical circuit energy consumption (listed in Table II) and link parameters (listed in Table III), we evaluate the total (transmission+circuit) energy consumption per bit of individual ARoF, DRoF and BoF links for 64QAM (spectrally efficient with low robustness) by varying salient parameters that significantly affect the energy consumption. The parameters considered are: 1) Cell Size: changes the path loss. 2) Fiber Length: changes the optical link loss. 3) Nyquist zone (for DRoF): changes the digital-to-analog conversion loss. The investigated energy efficiency results depend on the parameter values; however, the relative performance trends remain the same. 1) Cell Size: In a wireless link, the noise is generally dominated by the receiver thermal noise since the transmitter s thermal noise is degraded due to path loss. On the uplink, the RIN and shot noise in the optical link cause a significant degradation in the SNR. Fig. 10(a) and (b) compare the energy consumption per wireless PHY bit of ARoF, DRoF and BoF on the downlink and uplink, respectively, for a 100 m fiber link at different wireless transmission distances. We assume a RF carrier frequency of 2.44 GHz with bandwidth of 20 MHz in accordance with the IEEE WLAN standard. We observe that the energy per bit for ARoF and DRoF increases rapidly as the wireless transmission distance increases. This is due to the high power amplifiers needed to account for the loss in RF power on the optical link. In the case of ARoF, the loss in RF power is due to modulation depth, optical link loss and coupling loss. By reducing these losses, ARoF can be more energy efficient. In the case of DRoF, modulation depth and any optical link loss does not affect the RF power. However, there is a loss due to up-conversion in the DAC; when recovering the RF signal from the digital signal at the required carrier frequency, there is loss due to the sinc behavior of the DAC. With f c = 2.44 GHz and sampling frequency, f s = 300 MHz, this loss is approximately 36 db. However, in both cases, BoF is the most energy efficient, since there is no loss in RF power due to optical transmission. 2) Fiber Length: Fig. 11 shows the energy per bit of ARoF and DRoF for different fiber lengths for a fixed cell size (d = 10 m). In the figure we notice that the energy per bit for DRoF

9 GOWDA et al.: TOWARDS GREEN OPTICAL/WIRELESS IN-BUILDING NETWORKS: RADIO-OVER-FIBER 3553 Fig. 11. Energy per bit for different fiber lengths (Theoretical). Fig. 10. Comparison of the three dedicated links; Energy per bit versus wireless transmission distance (theoretical). (a) Downlink. (b) Uplink. Fig. 12. Energy consumption per bit versus Nyquist zone (Theoretical). does not change significantly when increasing the fiber length. However, in ARoF, the energy per bit increases when increasing the fiber length. In DRoF, the RF power is not affected by any optical link loss, whereas as the fiber length increases, the RF power loss in ARoF increases. This implies that for a lossy optical transmission or long-reach fiber scenarios (e.g., backhaul networks), it is more energy efficient to use DRoF. The cross over-point beyond which ARoF is more energy efficient depends on the required system components and their power consumption. 3) Nyquist Zone: The Nyquist zone is defined by the carrier frequency and sampling frequency. In the case of DRoF, there is a large loss in the RF power due to the sinc behavior of the DAC. The amount of loss depends on the Nyquist zone in which the carrier frequency exists. Fig. 12 shows the change in energy per bit for different Nyquist zones at a fixed distance of 5 m on the downlink. In this study, we assume that the carrier frequency and sampling frequency are kept the same. At the RAU, a LO is used to upconvert to the desired carrier frequency (essentially DRoF at an IF). The Nyquist zone can also be changed by changing the sampling frequency; increasing the sampling frequency lowers the Nyquist zone of a particular carrier frequency. Lower Nyquist zones have lower energy per bit due to lower loss. However, increasing the sampling frequency increases the data rate on the optical link. Another point to note is that the loss between different Nyquist zones reduces as the Nyquist zone increases, i.e., the difference between the first and second Nyquist zones is larger than the one between the seventh and eighth Nyquist zones. This implies that choosing a sampling frequency that results in a Nyquist zone of 7 rather than 8 does not give as much improvement in energy per bit as does when moving from the third to the second Nyquist zone. C. Energy Consumption of Proposed Green Architectures Theoretical Transmission throughput is considered a salient requirement in the design of a network. By moving to smaller cell sizes, we not only decrease the transmission energy consumption, but also increase the network throughput. To evaluate the proposed architectures, we assume a certain in-building area that requires wireless coverage, and we cover this area with N cells each of size r and bandwidth B. Assuming frequency re-use with low

10 3554 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 20, OCTOBER 15, 2014 Fig. 13. Comparison of the E b net of the three architectures; DRoF: Upconversion at RAU using DAC (Theoretical). Fig. 14. Comparison of the E b net of the three architectures; DRoF: Upconversion at RAU using Local Oscillator (Theoretical). co-channel interference, the maximum total network throughput, i.e., the peak data rate, is given by N B. Now, let s assume the cell size is reduced by half, r/2; hence, the number of cells required to cover the same area will be 4N. These smaller cells will still have the same bandwidth B. Thus, the maximum total network throughput will be 4NB. When considering network architectures that employ different technologies such as 60 GHz versus 2.4/5 GHz networks or architectures which have different point-to-point (p2p) link throughput, as in this paper, scaling bandwidth and comparing power consumption would not capture the added benefit of higher throughput offered. To compare p2p technologies, we use energy per bit as the energy efficiency metric. This metric is independent of the individual p2p throughput and captures their energy consumption. At the network level, we use a similar metric; however, here, we compare the required energy to transmit one bit through the whole network, E b net, which is given as follows: Peak Total Power Consumption E b net =. (22) Total Network Throughput With this metric, we can now measure the energy efficiency of the proposed network architectures, under practical deployment scenarios. The metric also has the advantage of only depending on the technology employed at the physical data layer. Thus, the analysis can be easily extended to include or compare higher level protocols by estimating on average how many effective bits are transmitted over the transmission layer per second and then multiplying the obtained number by E b net to get the overall cross-layer power consumption. In addition, daily user traffic profiles [35] can be used to capture the daily energy consumption in the network. In this work, we focus on the physical plane only. Using E b net, Fig. 13 compares the proposed green inbuilding architectures by estimating the required power amplifier gain for a 100 m 100 m floor plan (i.e., large inbuilding network). In the comparison, the architectures support only WLAN service. Here, we consider an in-building WLAN scenario; thus, the bandwidth per cell is B percell = 20 MHz. As observed in the figure, with ARCH1, E b net remains relatively constant at lower distances (5 15 m). At these distances the network throughput and the number of cells required for achieving complete wireless coverage scale by the same amount. Since at lower distances the power consumption is dominated by circuit energy, the total network power consumption scales linearly with the number of cells. Thus, the network throughput and the total power consumption also scale by the same amount, thereby having a constant E b net. This implies that by decreasing the cell size, the network throughput can be increased; however, the energy required to transmit a bit through the whole network will remain the same. This phenomenon can be seen in the other architectures as well (more explicitly if we consider cell size <5 m). With that, we can conclude that we should use the smallest possible cell size to get the highest throughput and the maximum energy saving. However, this may not be a cost-effective solution since for a given coverage area, we would require a large number of cells, and thus a large number of equipment. This presents a trade-off between cost and energy efficiency. The most energy and cost effective solution would be to design the network to operate at the point at which the curve of E b net starts flattening out. The RoF architectures have higher energy consumption due to the high gain power amplifiers needed to compensate for the optical transmission losses. For the values considered, the curves for ARCH2 and ARCH3 overlap; however, they may not overlap under different assumptions for parameters such as optical noise figure, EOE conversion loss, sampling frequency, etc., as demonstrated in Fig. 14. Fig. 14 compares the three architectures considering an IF (1st Nyquist zone and upconversion using an LO at the RAU) for ARCH3. We observe that the ARCH3 curve does not increase as rapidly as when using a higher Nyquist zone. For the parameters considered, ARCH3 with cell sizes <17 m is more energy efficient than ARCH1. In the above comparisons, we do not consider the power consumption required for cooling or the static power consumption of the digital processing units. In the case of the centralized architectures (e.g., ARCH2, ARCH3), the same processing unit can be shared among all the cells. However in decentralized

11 GOWDA et al.: TOWARDS GREEN OPTICAL/WIRELESS IN-BUILDING NETWORKS: RADIO-OVER-FIBER 3555 architectures (ARCH1), each cell will have its own unit. In cases where the remote access points consume more power, such as mobile base stations, the energy per bit floor for ARCH1 will be higher; thus, for remote access points/base stations that have high static power consumption, centralized architectures are more energy efficient. Also sleep mode can be applied to all the architectures to further reduce the energy per bit floor. VI. CONCLUSION Energy efficiency is now regarded a salient requirement in the operation of next-generation networks. In this paper, we investigated radio-over-fiber (RoF) technologies for achieving energyefficient optical/wireless converged in-building networks. We proposed and analyzed three different fiber-based in-building architectures; ARCH1 is the legacy architecture (i.e., using baseband-over-fiber); ARCH2 and ARCH3 centralize the digital signal processing but have dedicated RoF links for each cell. We presented a model to estimate and compare the power consumption of different RoF technologies. We validated the model using experimentally measured data. Our results show that point-topoint RoF links are not as energy efficient as BoF due to the optical link losses and conversion losses. However, centralized architectures using RoF technology are energy efficient when designed keenly with small cell sizes and when the remote access points/ base stations have high static power consumption. We show that careful design aimed at minimizing the loss due to the optical transmission will improve the energy efficiency of RoF links. For instance, our results show that DRoF-based architectures at lower Nyquist zones can be more energy efficient when the remote access points/base stations have high static power consumption. Under the assumptions considered in this paper, ARCH3 designed at the 1st Nyquist zone with upconversion using a local oscillator at the remote antenna unit is more energy efficient for cell sizes <17 m. APPENDIX DERIVATION OF RF OUTPUT POWER THROUGH AN EXTERNALLY MODULATED OPTICAL LINK The electric field equation at the output of the MZM is given [36] as follows: A c (t) = A 0 2 e ( jπ 2 e jφ 1 (t) + e jφ 2 (t) ) (23) where A 0 is the electric field at the output of the laser. For a sinusoid input and balanced drives: φ 1 (t) = πv RFcos(2πf RF t) (24) φ 2 (t) =V b πv RFcos(2πf RF t). (25) The Fourier transform of (23) is given by A c (f) = A 0 2 n= i n a n δ(f nf RF ) (26) where a n = J n ( πvrf ) ( ) πvrf + J n e jnπ e jπv b. (27) The optical power spectrum can be found as A c (f) A c ( f) which is given by where k= and P c (f) =A c (f) A c ( f) (28) = A2 0 i n 8π a k a k+nδ(f nf RF ) (29) a k a k+n = n= k= + e jθ 2 + e jnπ e jθ 2 k= J k (φ RF ) J k+n (φ RF ) k= + e jnπ k= J k (φ RF ) J k+n (φ RF ) e jkπ k= J k (φ RF ) J k+n (φ RF ) e jkπ J k (φ RF ) J k+n (φ RF ) φ RF = πv RF. The power of the first harmonic is the optical power of the RF signal. Thus, the received electrical power after photodetection is given by ( Popt e αl ( πvb P optrf = sin 4π ) ( ) 2 2πVin J 1 R) R l. (30) A more general derivation in the presence of dispersion can be found in [37]. REFERENCES [1] D. Kilper, Energy efficient networks, in Proc. Opt. Fiber Commun. Conf., OSA Tech. Digest, 2011, Paper OWI5. [2] G. Rittenhouse. (2011). GreenTouch: Putting Energy Into Making Networks More Efficient [Online]. Available: index.php?page=ict-industry-combats-climate-change [3] R. Tucker et al., Charting a Path to Sustainable and Scalable ICT Networks, GreenTouch June Open Forum, [4] L. Kazovsky, T. Ayhan, A. Gowda, K. Albeyoglu, H. Yang, and A. Ng oma, How to Design an Energy-efficient Fiber-Wireless Network, in Proc. Opt. Fiber Commun. Conf./Nat. Fiber Opt. Eng. Conf. OSA Tech. Digest, 2013, Paper OM3D.5. [5] Cisco Visual Networking Index: Forecast and Methodology, (2013). [Online]. Available: collateral/ns341 /ns525 /ns537 /ns705 /ns827 /white_paper_c _ns827_networking_solutions_white_paper.html, accessed Apr [6] D&R International Ltd. (2011) Buildings Energy Data Book, U.S. Department of Energy. [Online]. Available: eren.doe.gov [7] L. A. Barroso and U. Holzle, The case for energy-proportional computing, Computer, vol. 40, no. 12, pp , Dec [8] Q. Zhu, F. M. David, C. F. Devaraj, Z. Li, Y. Zhou, and P. Cao, Reducing energy consumption of disk storage using power-aware cache management, in Proc. IEE Softw., Feb , 2004, p. 118.

12 3556 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 32, NO. 20, OCTOBER 15, 2014 [9] Y. Yang, C. Lim, and A. Nirmalathas, Radio-over-fiber as the energy efficient backhaul option for mobile base stations, in Proc. Microw. Photon., 2011, pp [10] M. Crisp, R. V. Penty, I. H. White, and A. Bell, Wideband Radio over Fiber Distributed Antenna Systems for Energy Efficient In-Building Wireless Communications, in Proc. IEEE 71st Veh. Technol. Conf., May 16 19, 2010, pp [11] Y. Josse, B. Fracasso, and P. Pajusco, Model for energy efficiency in radio over fiber distributed indoor antenna Wi-Fi network, in Proc. 14th Int. Symp. Wireless Personal Multimedia Commun., Oct. 3 7, 2011, pp [12] Y. Josse, B. Fracasso, G. Castignani, and N. Montavont, Energy-efficient deployment of distributed antenna systems with radio-over-fiber links, in Proc. IEEE Online Conf. Green Commun., Sep , 2012, pp [13] Y. Yang, A. Nirmalathas, and C. Lim, Digitized RF-over-fiber as a costeffective and energy-efficient backhaul option for wireless communications, Annales des Telecommun. vol. 68, no. 1/2, pp , 2013 [14] J. C. Fan and L. G. Kazovsky, A fiber infrastructure for microcellular personal communication systems: Infrastructure designoptical dynamic range requirements, Ph.D. dissertation, Stanford Univ., Stanford, CA, USA, [15] Y. Yang, Investigation on digitized RF transport over fiber, Ph.D. dissertation, Dept. of Electrical, and Electronic Eng., The Univ. of Melbourne, Parkville, VIC, Australia, [16] Z. Gong, K. Xu, X. Meng, Y. Pei, X. Sun, Y. Dai, Y. Ji, and J. Lin, Effect of optical losses on the transmission performance of a radio-over-fiber distributed antenna system, Chin. Opt. Lett., vol. 11, pp , [17] L. G. Kazovsky, T. Ayhan, A. S. Gowda, A. R. Dhaini, A. Ng oma, and P. Vetter, Green in-building networks: The future convergence of green, optical wireless technologies, in Proc. 15th Int. Conf. Transparent Opt. Netw., Jun , 2013, pp [18] N. Haralabidis et al., A cost-efficient 0.18 um CMOS RF transceiver using a fractional-n synthesizer for b/g wireless LAN applications, in Proc. IEEE Custom Integr. Circuits Conf., Oct. 3 6, 2004, pp [19] Z. Xu et al., A compact dual-band direct-conversion CMOS transceiver for a/b/g WLAN, in Proc. IEEE Int. Solid-State Circuits Conf., Feb. 10, 2005, vol. 1, pp [20] C. Chang et al., A CMOS transceiver with internal Padigital pre-distortion for WLAN a/b/g/n Applications, in Proc. IEEE Radio Frequency Integr. Circuits Symp., May 23 25, 2010, pp [21] R. S. Tucker, Green Optical Communications Part I: Energy Limitations in Transport, IEEE J. Sel. Topics Quantum Electron., vol.17,no.2, pp , Mar./Apr [22] ArchCom Tech., AC6538 Series 10 Gbps 850nm Multimode ROSA for EDC Applications p-i-n+tia, [23] J. Ryckaert, J. Borremans, B. Verbruggen, L. Bos, C. Armiento, J. Craninckx, and G. Van der Plas, A 2.4 GHz low-power sixth-order Rf bandpass converter in CMOS, IEEE J. Solid-State Circuits, vol. 44, no. 11, pp , Nov [24] C. H. Cox, III, E. I. Ackerman, G. Betts, and J. L. Prince, Limits on the performance of RF-over-fiber linkstheir impact on device design, IEEE Trans. Microw. Theory Techn., vol. 54, no. 2, pp , Feb [25] A. Goldsmith, Wireless Communications. New York, NY, USA: Cambridge Univ. Press, [26] C. H. Cox, Analog Optical Links: Theory Practice. Cambridge, U.K.: Cambridge Univ. Press, [27] B. Nikolic, EE 241. Class LectureTopic: Low power-low energy circuit design, Dept. of Electrical Computer Eng., Univ. of California, Berkeley, CA, USA Spring, [28] R. S. Tucker, Green Optical Communications Part II: Energy Limitations in Networks, IEEE J. Sel. Topics Quantum Electron.,vol.17,no.2, pp , Mar./Apr [29] Q. Gu, RF Systems Design of Transceivers for Wireless Communications. New York, NY, USA: Springer, [30] N. J. Gomes, A. Nkansah, and D. Wake, Radio-Over-MMF Techniques Part I: RF to Microwave Frequency Systems, J. Lightw. Technol., vol.26, no. 15, pp , Aug. 1, [31] M. Sauer, A. Kobyakov, and J. George, Radio Over Fiber for Picocellular Network Architectures, J. Lightw. Technol., vol. 25, no. 11, pp , Nov [32] V. Erceg et al. IEEE P Wireless LANs: TGn Channel Models, Standard: IEEE n-03, May [33] Freescale Semiconductor, MBC13720: SiGe:C Low Noise Amplifier With Bypass Switch, [34] U2T photonics, 50 GHz Photodetector: XPDV2150R, [35] G. Auer, V. Giannini, I. Godor, P. Skillermark, M. Olsson, M. A. Imran, D. Sabella, M. J. Gonzalez, C. Desset, and O. Blume, Cellular Energy Efficiency Evaluation Framework, in Proc. IEEE 73rd Vehicular Technol. Conf., May 15 18, 2011, pp [36] G. P. Agrawal, Introductionin Lightwave Technology: Telecommunication Systems. Hoboken, NJ, USA: Wiley, [37] L. Cheng, S. Aditya, and A. Nirmalathas, An exact analytical model for dispersive transmission in microwave fiber-optic links using Mach- Zehnder external modulator, IEEE Photon. Technol. Lett., vol. 17, no. 7, pp , Jul Authors photographs biographies not available at the time of publication.