Energy Efficiency Analysis of the Watchful Sleep Mode with Delayed Wakeup in PONs

Energy Efficiency Analysis of the ful Sleep Mode with Delayed Wakeup in PONs Raisa O. C. Hirafuji, Ahmad R. Dhaini, Divanilson R. Campelo, Pin-Han Ho, and Limei Peng Abstract Over the last decade, passive optical network (PON) has been taken as the most promising next-generation access network. To improve the energy efficiency in PON systems, the ITU-T standard for next-generation PON system introduced a new power management mode, so-called the ful Sleep mode. This new mode is expected to be operated on any TDM- PON or GPON variant. Although the sleep mechanism under the ful Sleep mode is well defined, its performance can vary greatly depending on the criteria used to start and/or terminate the power saving phase. In this article, we present a comprehensive analytical model for evaluating the energy efficiency of the watchful sleep mode using the Delayed Wakeup (DWU) mechanism. We then compare its performance with the Immediate Wakeup (IWU) scheme. Our analytical and simulation results highlight the accuracy of the presented model and prove its merits. I. INTRODUCTION The information and communication technologies (ICTs) have tremendously expanded in the last two and a half decades. Over 7% of the world youth population are online, and in developed countries, this percentage is around 94% [1]. This expansion has led to an increased concern of the enviroental impact of ICT systems [2]. In the access segment, passive optical networks (PONs) have been considered the most energyefficient access technology [3]. It is possible to reduce the power consumption in PON systems via the Optical Line Terminal (OLT), or via the Optical Network Units (ONUs). As the ONUs are responsible for most of the power consumption in PONs [4], reducing its power consumption by putting it in a low-power state is considered the best method to improve the energy efficiency of PON systems [5] [8]. The Gigabit PON (G-PON) [9] and 1-Gigabit PON (XG- PON) [1] standards specify three protocol-based power management techniques, namely the Cyclic Sleep, Doze and ful Sleep modes. These modes require the ONU to alternate from an active phase to a power saving phase, and vice versa [11], such that the transition between these two phases is managed by both the OLT and ONU. In order to start a power saving phase, it is necessary for both the ONU and OLT to agree on entering the power saving phase, whereas Raisa O. C. Hirafuji and Pin-Han Ho are with the Dept. of Electrical and Computer Engineering, University of Waterloo, Canada. Email: {rodacost, p4ho}@uwaterloo.ca. Ahmad R. Dhaini is with the Dept. of Computer Science, American University of Beirut, Lebanon. Email: ahmad.dhaini@aub.edu.lb. Divanilson R. Campelo is with Centro de Informática, Universidade Federal de Pernambuco, Brazil. Email: dcampelo@cin.ufpe.br. Limei Peng is with the Dept. of Industrial Engineering, Ajou University, Korea. to terminate the power saving phase, it is only necessary for either the ONU or OLT to request switching into the active phase. The difference between the three modes is in the operation of the power saving phase. For instance, with the Cyclic Sleep and Doze modes, in the power saving phase, the ONU switches between an active state, namely the Sleep/Doze Aware state, and a low-power state. The low-power state for the Cyclic Sleep mode is the Asleep state, and for the Doze mode, it is the Listen state [12]. In the Asleep state, the ONU turns off both the receiver and transmitter, and thus would be unable to transmit or receive traffic. Hence, if the OLT desires the ONU to switch into the active phase, it will have to wait until the ONU returns to an active state. In the Listen state, the ONU s transmitter is turned off, yet the receiver is kept on. Hence, the OLT can request the ONU to return to the active phase without having to wait for it to transition into the Aware state. The ful Sleep mode combines both the Cyclic Sleep and Doze modes by alternating between three power levels during the power saving phase [11]: 1) the WSleep Aware state, with both the receiver and transmitter on; 2) the Rx ON state with only the receiver on; and 3) the Rx OFF state, with both the receiver and transmitter off. It has been shown [11] that the ful Sleep mode outperforms the other two modes in energy efficiency, and a PON system supporting this mode can emulate either mode as a special case. Moreover, ever since the NG-PON2 ITU-T specification [13], the ful Sleep mode has been the only allowed power management mode, whereas the Cyclic Sleep and Doze modes are no longer mentioned. In this article, we present an analytical modeling of the ful Sleep mode that, differently from previous analytical models such as [14], considers a more aggressive implementation of the wakeup process initiated by the ONU and OLT: the Delayed Wakeup (DWU). With the DWU implementation, upon a packet arrival during the power saving phase, the ONU and OLT will only initiate the wakeup process after a pre-determined period of time [15]. Note that with the implementation of [14], the Immediate Wakeup (IWU) scheme is employed where the power saving phase is terminated by any packet arrival (i.e., to either the ONU or the OLT), which achieves high performance at the expense of very poor energy saving. The remainder of this article is organized as follows. In Section II, the ful Sleep mode is briefly overviewed. Section III presents the proposed analytical model. Section IV presents our simulation results, which verify the proposed

model and compare the performance of the ful Sleep mode under the DWU and IWU schemes. Finally, we conclude our work in Section V. From Low Power Stimulus T watch Expires Rx OFF WSleep Aware No wake up stimulus T sleep Expires T aware Expires Rx ON To Fig. 1. Power saving phase for the ful Sleep mode. Wake up Stimulus II. OVERVIEW OF THE WATCHFUL SLEEP MODE The ful Sleep mode operates based on finite state machines at the ONU and OLT, respectively. The one at the ONU has four states, namely Held, Free, WSleep Aware, and. The active phase is composed by the two states, the Held and Free states. In the Held state, the ONU is not allowed to leave the active phase. The Free state indicates that the ONU has the permission to leave the active phase and start a power saving phase. Before transitioning to the WSleep Aware state, the ONU must inform the OLT. The WSleep Aware state is transitional to the state. The main difference between this state and the Free state is that the OLT is aware that the ONU may enter into a low power state. The state is divided in two parts, namely the Rx OFF where the ONU turns off most of its components including the receiver and transmitter; and the Rx ON, which is an intermediate state with the receiver on and the transmitter off. The ONU returns to Rx OFF if there is no wakeup message received. If there is a wakeup stimulus, the ONU will transition to the Held state. Upon transitioning to the Held state, the ONU informs the OLT that it is returning to the active phase. For more details, we refer the reader to [15]. The OLT s state machine also comprises four states, namely Awake Forced, Awake Free, Low Power, and Alerted. During the Awake Forced state, the OLT has not given permission to the ONU to enter a low power state, and the ONU is expected to be fully active. In the Awake Free state, the OLT has already granted the ONU permission for switching into a power saving phase. The OLT still expects the ONU to be fully active and forwards downstream traffic. In the Low Power state, the OLT only expects intermittent messages from the ONU and may buffer the downstream traffic. The OLT enters the Alerted state after a local wakeup stimulus. It sends wakeup stimuli to the ONU in every allocation grant so that the ONU terminates the power saving phase. When the OLT receives confirmation that the ONU returned to the active phase, the OLT returns to the Awake Forced state. For more details, we refer the reader to [15]. For the reader s convenience, the most relevant notations used in our model are summarized in TABLE I. During the power saving phase, illustrated in Figure 1, firstly, the ONU will stay in the WSleep Aware state until either T aware expires or until any wakeup stimulus is present. After T aware expires, the ONU enters the state until either T watch expires or the trigger of some wakeup stimulus is present. Within the state, the ONU remains in the Rx OFF state until either T sleep expires, or some local wakeup stimulus is triggered. Note that during the Rx OFF state, the ONU cannot receive any external wakeup stimulus. Thus, after T sleep expires, the ONU enters the Rx ON state. On the other hand, during the Rx ON state, the ONU checks for external wakeup stimuli, and once it is certain that the OLT does not want to wakeup the ONU, the ONU returns to the Rx OFF state. Besides the sojourn time in these three states, we consider the transition time from the Rx OFF state to the Sleep Aware state, referred to as T Transinit ; the transition time from the Rx OFF state to the Rx ON, is referred to as T Rxinit ; and the transition time from the Rx ON state to the, is referred to as T Txinit. III. PROPOSED ANALYTICAL MODEL For the analytical modeling of the power saving phase of the ful Sleep mode, we assume Poisson traffic. The upstream and downstream traffic arrives with rates and λ ds, respectively. To calculate the average power consumption of an ONU, we must first know how long the ONU sojourns in each state of the power saving phase. Ten possible cases are defined in which there could be a packet arrival during the power saving phase [14]. A detailed description of these cases can be found in TABLE II. We define a random variable t us,, for k = 1, 2, 3, 4, 5, as the upstream inter-arrival time for cases 1, 2, 3, 4 and 5, whose average from A to B, t us,, is computed as follows: t us, = = A A e λusa xe λusx dx e λusx dx A + 1 e λusb e λusa e B B + 1 We also define a random variable t ds,, for k = 6, 7, 8, 9, 1, as the downstream inter-arrival time for cases 6, 7, 8, 9, and 1, whose average from A to B, t ds,, is computed as follows: (1)

Notation T hold T aware T sleep T listen T watch T Transinit T Rxinit T Txinit t us t ds ρ ɛ g φ P A P L P S P W Description TABLE I SUMMARY OF NOTATIONS Minimum duration of the Held state Maximum duration of the WSleep Aware state Maximum duration of the Rx OFF state Maximum duration of the Rx ON state Maximum duration of the state Transition overhead from Rx OFF to an active state Transition overhead from Rx OFF to Rx ON Transition overhead from Rx ON to an active state Upstream inter-arrival time Downstream inter-arrival time Traffic intensity Energy consumption during case k Probability of case k Power consumption when the ONU is fully active Power consumption of the Rx ON state Power consumption of the Rx OFF state T sleep P S +T Trxinit P L +T listen P L T sleep +T listen +T Trxinit W n n (T aware + T watch + T Transinit ) W n + W n + T aware Wn W n (T watch + T Transinit ) S m m T sleep + T Rxinit + T listen S + m S m S m + T sleep S m T listen E C Energy consumption during the power saving phase E CA E C considering P A = P S = P L = % E S Energy saving in percentage TABLE II CASES OF POSSIBLE FIRST ARRIVALS DURING THE POWER SAVING PHASE Case Case 1 Case 2 Case 3 Case 4 Case 5 Case 6 Case 7 Case 8 Case 9 Case 1 t ds, = = Description Upstream arrival during the WSleep Aware state. Upstream arrival during the transition overhead from the state to the next WSleep Aware state. Upstream arrival during the Rx OFF state. Upstream arrival during the transition overhead from the Rx OFF to the Rx ON. Upstream arrival during the Rx ON state. Downstream arrival during the WSleep Aware state. Downstream arrival during the transition overhead from the state to the next WSleep Aware state. Downstream arrival during the Rx OFF state. Downstream arrival during the transition overhead from the Rx OFF to the Rx ON. Downstream arrival during the Rx ON state. A A e λusa λ ds xe λ dsx dx λ ds e λdsx dx A + 1 λ ds e λusb e λ dsa e λ ds B B + 1, (2) where A is defined as: W n 1, k = 1, 6 W n T Transinit, k = 2, 7 A = Wn + S m 1, k = 3, 8 Wn + Sm T Rxinit, k = 4, 9 Wn + Sm, k = 5, 1 B is defined as: W n 1 +, k = 1, 6 W n, k = 2, 7 B = Wn + S m 1 +, k = 3, 8 Wn + Sm, k = 4, 9 Wn + S m, k = 5, 1 With the DWU scheme, the ONU and OLT delay the wakeup process for periods H ONU and H OLT, respectively. For the ONU initiated wakeup, the ONU can locally schedule to start the wakeup process right after H ONU expires (see Fig. 2(a)). At the OLT side, however, this process is much slower due to the fact that after the OLT waits for H OLT, the ONU could be in the Rx OFF state, and thus the OLT will have to wait until the ONU returns either to the Rx ON state or to the WSleep Aware state (see Fig. 2(b)). The time needed for the OLT to message the ONU is usually much smaller than H ONU and H OLT, making very little difference in the final results. Thus, to keep the model tractable, we do not consider it in our analysis. We then define a variable T to represent the ONU ONU OLT Awake Free A t us A B SR(WSleep) B Power Saving (a) Low Power t ds packet arrival packet arrival Power Saving (b) H OLT SA(OFF) wake up x x FWI FWI FWI H ONU x Alerted wake up SR(Awake) Awake Forced Fig. 2. ONU phases transitions for: a) ONU-initiated wakeup; (b) OLTinitiated wakeup. (3) (4)

ONU s average sojourn in the power saving phase, computed as follows: t us,+h ONU W + T n T = sleep +T listen +T Rxinit, 1 k 5 (5) t ds, +H ONU W + n T sleep +T listen +T Rxinit, 6 k 1 Another important consideration for the DWU scheme is that, depending on the values of H ONU and H OLT, it is possible that while the ONU is waiting H ONU to start the wakeup process, the OLT may receive a packet, wait H OLT, and wake up the ONU before H ONU expires. The opposite may also happen, where the ONU can locally wake up before H OLT expires. For the first five cases (i.e., k = 1, 2, 3, 4, 5), the upstream arrival occurs before the downstream arrival. Due to the case where the OLT may wake up the ONU before H ONU expires, and since the OLT can only wake up the ONU by the end of a watch cycle, we need to define the variable G k that represents how many watch cycles there are in H ONU + T Transinit : G k H ONU + T Transinit = (6) T sleep + T Rxinit + T listen Thus, E for the first five cases (k = 1, 2, 3, 4, 5), is given as follows: E = M n=1 m=1 G k φ g= e λ dss g ɛ g, where ɛ g for the DWU model is defined in equation 11 and φ for k = 1, 2, 3, 4, 5 is given by: e λ(wn 1) 1 e λustaware, k = 1 ( e λwn e T Transinit 1 ), k = 2 φ = e λ(w +Sm 1) ( n 1 e sleep) λust, k = 3 e λ(w n +S m) ( e λustrxinit 1 ) (8), k = 4 e λ(w +Sm) ( n e λust listen 1 ), k = 5 By observing that for H ONU +T Transinit > S g+t +T Txinit, ɛ g, the average energy consumption of the ONU for an OLT-initiated wakeup can be calculated if there is an downstream arrival during the S g watch cycle. If there is not enough time for the OLT-initiated wakeup, or if there are no downstream arrivals, then ɛ g shall reflect the energy consumption for an ONU-initiated wakeup. For the sake of presentation, we denote P [W + n ] as: P [ W + n and P [S m ] as: (7) ] = (n + 1) Taware P A + T Transinit P A + nt watch P W, (9) P [S m ] = m (T sleep P S + T Rxinit P L + T listen P L ) (1) For the last five cases (i.e., k = 6, 7, 8, 9, 1), where there is a downstream arrival before an upstream arrival, the ONUinitiated wakeup case occurs when there is an upstream arrival between B and W + n 1 +S h+t Txinit. Then, t us, is given by: t us, = e λus B e λus(b ) ( B ) + 1 e λus(w + n 1 +S h+t Txinit) ( e λus(w + n 1 +S h+t Txinit) W n 1 + + S h + T Txinit + 1 ) B e λus(w + n 1 +S h+t Txinit) e λus ( E is given by: E = M n=1 m=1 (12) ) φ ɛ, (13) where φ for k = 6, 7, 8, 9 and 1 is given by: e λ(wn 1) 1 e λ dst aware e T aware, k = 6 ( e λwn e λ ds T Transinit 1 ), k = 7 φ = e λ(w +Sm 1) n 1 e λ dst sleep e T sleep, k = 8 e λ(w n +S m) ( e λ dst Rxinit 1 ), k = 9 e λ(w +Sm) ( n e λ dst listen 1 ), k = 1 (14) Consequently, the average energy consumption (E C ) during the power saving phase would be obtained by: 1 E C = E. (15) k=1 Since the ONU only enters the power saving phase when it is idle, we can obtain the total energy consumption for the power saving phase by considering that the ONU spends (1 ρ) of the total time idle, where ρ is the traffic intensity on the direction with the most intense traffic. The energy saving (E S ) in percentage is then computed as: EC E S = (1 ρ), (16) E CA where E CA is the energy consumption during the power saving phase for P A = P S = P L = %. IV. PERFORMANCE EVALUATION To validate the proposed analytical model, we compare the numerical results with simulation results. The simulation model considers a time-division multiplexing (TDM) PON system with one OLT and 16 ONUs. The downstream rate is 1 Gbps, and the upstream rate is 2.5 Gbps. We assume a propagation delay of µs. The parameters P A, P L and P S are set as %, 4% and 5%, respectively. The bandwidth is equally allocated to all ONUs. Every 2 ms each one of the 16 ONUs receives a 125µs time slot to transmit the upstream traffic, regardless if they are active or have packets to transmit. The upstream and downstream traffic is set such that the traffic intensity (ρ) is the same in both

ɛ g = ( 1 e λ ds S 1 P [ [ ] ) W n 1] + + P S g+t + P A T Txinit P [ W n 1 + ] ) + PW (t us, + H ONU W n 1 + + P A T Transinit, otherwise, H ONU + T Transinit > S g+t + T Txinit (11) directions. This results on the following equation to calculate ρ for the analytical model: ρ = N λ i us i=1 µ us or ρ = N λ i ds i=1 µ ds (17) where N is the number of ONUs, λ i us/λ i ds is /λ ds for the i th ONU and µ us /µ ds is the service rate in the upstream/downstream direction. The incoming traffic is modeled with a Poisson distribution, such that the packet size varies according to the following distribution: 64 bytes (47%), 3 bytes (5%), 594 bytes (15%), 13 bytes (5%), and 1518 bytes (28%) [16]. A confidence interval of 95% is used [11]. The T hold parameter is set to 2 ms, as a bigger value would needless reduce the energy saving. T listen is set to 125µs, which is the maximum amount of time that the ONU has to wait for a frame from the OLT. T watch is set as 1 s, because it must be smaller than the maximum permitted interval between handshakes, but at the same time it is desirable for T watch to be several times bigger than T sleep. T sleep is set as 5 ms, because smaller values have very little energy saving, whereas bigger values have little increase on the energy saving at the cost of an increase in the end-to-end delay. T aware is set as 5 ms, because it must the bigger than the minimum time required for a handshake. The transition overheads T Txinit, T Rxinit and T Transinit are set to 3 ms, 2 ms and 3 ms, respectively [11]. For the DWU implementation, the parameters with the largest impact on the energy efficiency are H OLT and H ONU. Fig. 3 shows the energy efficiency for H ONU = H OLT = 2 ms, 4 ms, 6 ms, 8 ms and ms. It can be also observed that even though there is a big difference in the energy saving between 2 ms, 4 ms, 6 ms, the difference starts becoming insignificant for bigger values, such as 8 ms and ms. This is expected, as the increase of 2 ms in the power saving phase is much more significant for a small value, such as 2 ms, than it is for a bigger value, such as ms. Fig. 4 highlights the impact that H ONU has on the energy efficiency when H OLT is fixed as ms. As noticed, the curves from Fig. 4 are very similar to the ones in Fig. 3. Since the ONU can wake up locally, whereas the OLT can only wake up the ONU when the receiver is on, when both the time inter-arrivals and the difference between H ONU and H OLT are very small, the OLT will not have enough time to wake up the ONU before the ONU wakes up by itself. Thus, unless H OLT is T sleep smaller than H ONU, this parameter will have little effect on the energy saving. This can also be observed in Fig. 5, where H ONU is fixed as ms, and H OLT varies between 2 ms to ms. In contrast, the curves for H OLT = 2 ms and 4 ms are very close. The OLT delays the wakeup process for 2-4 ms after a packet arrival; however, since T sleep = 5 ms, the OLT will have to wait until T sleep expires to be able to wake up the ONU. For H OLT values bigger than T sleep, the OLT will have to wait until the next T sleep expires or for an ONU-initiated wakeup. Moreover, the curves for H OLT = 6 ms, 8 ms and ms are basically the same as the one for H ONU = ms. 8 6 4 2 H ONU = H OLT = 2 ms H ONU = H OLT = 4 ms H ONU = H OLT = 6 ms H ONU = H OLT = 8 ms H ONU = H OLT = ms Fig. 3. Energy efficiency for the ful Sleep mode with the DWU where H ONU = H OLT = 2, 4, 6, 8, ms. 8 6 4 2 H ONU = 2 ms H ONU = 4 ms H ONU = 6 ms H ONU = 8 ms H ONU = ms Fig. 4. Energy efficiency for the ful Sleep mode with the DWU where H OLT = ms for of H ONU = 2, 4, 6, 8, ms. 8 6 4 2 H OLT = 2 ms H OLT = 4 ms H OLT = 6 ms H OLT = 8 ms H OLT = ms Fig. 5. Energy efficiency for the ful Sleep mode with the DWU where H ONU = ms for of H OLT = 2, 4, 6, 8, ms.

Delay (s) In order to validate our analysis, we compare in Fig. 6 the analytical model with simulation results. The model s results are very close to the one obtained via simulations, what validates the correctness of our model. 5 5 Analytical Model Simulation Result IWU DWU.1.2.3.4.5 (a) Comparison of the energy saving of IWU and DWU. 1-1 Mean delay for the IWU Max delay for the IWU Mean delay for the DWU Max delay for the DWU Fig. 6. Energy Saving for H ONU = 5 ms and H OLT = 5 ms under different network loads. To illustrate the difference between the DWU model and the IWU model from [15], we compare in Fig. 7(a) the two implementations of the ful Sleep mode. As noticed, although both implementations start with very good energy efficiency, the energy efficiency with IWU is negligible around ρ =.1. In contrast, with DWU, the energy efficiency reaches around 8% in this same region, at the expense of increased upstream packet delay due to the delayed wakeup in H ONU /H OLT, as shown in Fig. 7(b). However, it can be noticed that the mean packet delay with DWU is very close to H ONU /2, while the maximum packet delay is a bit over H ONU. This is significantly higher than the IWU delay. V. CONCLUSIONS In this paper, we presented a new analytical model to evaluate the energy efficiency of the ful Sleep mode under the DWU implementation. The proposed model enables studying the impact of delayed wakeup at the ONU and OLT, what is essential to gain a better understanding of the system behavior. Due to the distance between the ONU and OLT, the ONU parameters are observed to be much more impactful on the network performance. Extensive simulations were conducted; results confirmed the accuracy of the proposed analytical model. We have also compared the performance of DWU with the IWU; results show that 8% more power saving can be attained using DWU under a traffic load of ρ =.1, while with IWU, no energy is saved. REFERENCES [1] B. Sanou, ICT Facts and Figures 217, ITU-T, 217. [2] W. Vereecken, W. Van Heddeghem, M. Deruyck, B. Puype, B. Lannoo, W. Joseph, D. Colle, L. Martens, and P. Demeester, Power consumption in telecommunication networks: overview and reduction strategies, Communications Magazine, IEEE, vol. 49, no. 6, pp. 62 69, June 211. [3] J. Baliga, R. Ayre, K. Hinton, and R. Tucker, Energy consumption in wired and wireless access networks, Communications Magazine, IEEE, vol. 49, no. 6, pp. 7 77, June 211. [4] J. Baliga, R. Ayre, K. Hinton, W. V. Sorin, and R. S. Tucker, Energy Consumption in Optical IP Networks, Lightwave Technology, Journal of, vol. 27, no. 13, pp. 2391 243, Jul. 29. [5] S.-W. Wong, L. Valcarenghi, S.-H. Yen, D. Campelo, S. Yamashita, and L. Kazovsky, Sleep Mode for Energy Saving PONs: Advantages and Drawbacks, in GLOBECOM Workshops, 29 IEEE, Nov. 29, pp. 1 6. 1-2 1-3 (b) Comparison of the mean upstream delay of IWU and DWU. Fig. 7. Comparison of the performance of IWU and DWU with H ONU = 5 ms and H OLT = 5 ms. [6] A. Dhaini, P.-H. Ho, and G. Shen, Toward green next-generation passive optical networks, Communications Magazine, IEEE, vol. 49, no. 11, pp. 94 11, November 211. [7] L. Valcarenghi, D. R. Campelo, S.-W. Wong, S.-H. Yen, S. Yamashita, D. P. Van, P. G. Raponi, P. Castoldi, and L. G. Kazovsky, Energy Efficiency in Passive Optical Networks: Where, When, and How? vol. 26, no. 6, pp. 61 68, Nov. 212. [8] A. Dhaini, P.-H. Ho, G. Shen, and B. Shihada, Energy Efficiency in TDMA-Based Next-Generation Passive Optical Access Networks, Networking, IEEE/ACM Transactions on, vol. 22, no. 3, pp. 85 863, June 214. [9] ITU-T, Series G: Transmission Systems and Media, Digital Systems and Networks - Gigabit-capable passive optical networks (G-PON): Transmission Convergence layer specification G.987.3, ITU-T, Jan. 214. [1] ITU-T G.987.3, 1-Gigabit-Capable Passive Optical Networks (XG- PON): Transmission Convergence (TC) Specifications Release 2, ITU- T, January 214. [11] D. A. Khotimsky, D. Zhang, L. Yuan, R. O. C. Hirafuji, and D. R. Campelo, Unifying Sleep and Doze Modes for Energy-Efficient PON Systems, Communications Letters, IEEE, vol. 18, no. 4, pp. 688 691, April 214. [12] B. Skubic and D. Hood, Evaluation of ONU Power Saving Modes for Gigabit-Capable Passive Optical Networks, IEEE Network, March/April 211. [13] ITU-T G.989.3, 4-Gigabit-capable passive optical networks (NG PON2): Transmission Convergence Layer Specification, ITU-T, October 215. [14] R. O. C. Hirafuji, A. R. Dhaini, D. A. Khotimsky, and D. R. Campelo, Energy efficiency analysis of the watchful sleep mode in next-generation passive optical networks, in 216 IEEE Symposium on Computers and Communication (ISCC), June 216, pp. 689 695. [15] R. O. C. Hirafuji, K. B. da Cunha, D. R. Campelo, A. R. Dhaini, and D. A. Khotimsky, The watchful sleep mode: a new standard for energy efficiency in future access networks, Communications Magazine, IEEE, vol. 53, no. 8, pp. 15 157, August 215. [16] ITU-T, Series G: Transmission Systems and Media, Digital Systems and Networks - End-user multimedia QoS categories, ITU-T, Nov. 21.