Dynamic Power Management at the Access Node and ... - IEEE Xplore

1 downloads 0 Views 2MB Size Report
ment in point-to-point (PtP) and time-division-multiplexed passive optical network (TDM-PON) systems to maximize the achievable power-saving performance ...
IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 32, NO. 8, AUGUST 2014

1575

Dynamic Power Management at the Access Node and Customer Premises in Point-to-Point and Time-Division Optical Access Jie Li, Member, IEEE, Ka-Lun Lee, Member, IEEE, Chien Aun Chan, Member, IEEE, N. Prasanth Anthapadmanabhan, Member, IEEE, Nga Dinh, Member, IEEE, and Peter Vetter, Member, IEEE

Abstract—In order to achieve the maximum power-saving performance in optical access networks, all network elements including the access node or optical line terminal (OLT) and the optical network unit should be considered. In this paper, appropriate power-saving mechanisms are determined for each network element in point-to-point (PtP) and time-division-multiplexed passive optical network (TDM-PON) systems to maximize the achievable power-saving performance while satisfying the quality-of-service requirements for different applications. By applying these powersaving mechanisms in PtP and TDM-PON systems, it is found that 1) the power consumption of OLTs in the central office is reduced by up to 51% and 38%, respectively, and 2) the total power consumption per subscriber is reduced by up to 61.5% and 67%, respectively. Using simulations, the most energy-efficient optical access network technology for different Internet applications is also identified. Index Terms—Access network, passive optical network, sleep mode, dozing mode, energy-efficient ethernet, packet coalescing.

I. I NTRODUCTION

F

IBER-TO-THE-HOME (FTTH) using passive optical networks (PONs) is a relatively energy-efficient broadband access technology compared to copper-based wireline access networks and wireless access networks. However, the total energy consumption per subscriber is still quite significant, and so additional improvements to the energy-efficiency of FTTH technology can further lower operating costs and reduce climate impact [1]. Two commonly deployed FTTH networks are Ethernet PON (EPON) and Gigabit PON (GPON), which use Manuscript received September 30, 2013; revised March 1, 2014; accepted April 28, 2014. Date of publication July 8, 2014; date of current version September 19, 2014. This work was supported in part by the University of Melbourne, by Alcatel-Lucent Bell Labs, by the Victoria State Government, and by the Seoul Metropolitan Government R&BD Program WR080951. This paper was presented in part at the IEEE International Conference on Communications (ICC) Green Broadband Access: Energy Efficient Wireless and Wired Network Solutions Workshop, Budapest, Hungary, June 2013. J. Li, K.-L. Lee, and C. A. Chan are with the Centre for EnergyEfficient Telecommunications (CEET), University of Melbourne, Vic. 3010, Australia (e-mail: [email protected]; [email protected]; chienac@ unimelb.edu.au). N. Prasanth Anthapadmanabhan and P. Vetter are with Alcatel-Lucent Bell Labs, Murray Hill, NJ 07974-0636 USA (e-mail: [email protected]; [email protected]). N. Dinh is with Alcatel-Lucent Bell Labs, Seoul 135-798, Korea (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/JSAC.2014.2335333

time-division multiple access (TDMA) for upstream transmissions and time division multiplexing (TDM) over a broadcast channel for downstream transmissions. These two systems, known as the TDM-PON, enable a single optical fiber to be shared among a large number of subscribers using an optical passive splitter, hence minimizing the capital expenditure (CAPEX). As the demand of bandwidth intensive services grows, higher access rate with stringent quality-of-service (QoS) is mandatory. In this case, the point-to-point (PtP) fiber access network is an alternative network architecture that provides the potential for higher sustained user rates and simpler infrastructure upgrade paths despite its higher initial CAPEX. Fig. 1 shows the network architectures of the PtP and TDMPON access networks. The access node or optical line termination (OLT) located at the central office (CO) consists of a number of optical access interfaces (AIs) that connect the optical network units (ONUs) and multiple stages of Ethernet aggregators (EAs) located at both line terminals (LTs) and network terminals (NTs). In conventional optical access networks, ONUs and all network elements in the CO are constantly active and consume a significant amount of power even in idle periods. Since ONUs consume the most power in PONs [2], network energy efficiency can be improved by applying sleep mode mechanisms to ONUs. Several sleep mode mechanisms have been described by the International Telecommunication Union (ITU) for GPON and XGPON ONUs [3], [4]. However, sleep mode mechanisms for TDM-PON OLTs are difficult to implement mainly because an OLT is time shared by many subscribers, preventing the OLT from entering into the sleep state [5]. In PtP systems, power savings at both OLTs and ONUs are even more crucial. Since an AI only connects to a single ONU, it results in higher power consumption per subscriber, higher CAPEX and higher operational expenditure (OPEX) than that of a TDM-PON. In this paper, we investigate different mechanisms that enable power saving at ONUs, AIs and EAs in PtP and TDM-PON systems. For each case, we combine the power saving mechanisms at different network elements to minimize the power consumption of the entire access network. The rest of this paper is organized as follows. In Section II, we first discuss and review the existing power saving mechanisms in optical access networks. A power saving scheme for the PtP access network is proposed in Section III. The proposed power saving scheme

0733-8716 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

1576

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 32, NO. 8, AUGUST 2014

B. Fixed-Cyclic Sleep Mode (FCSM)

Fig. 1. Network elements of (a) PtP and (b) TDM-PON access networks.

is based on sleep-mode mechanisms that can be applied to ONUs as well as the AIs and EAs in the OLT. Factors affecting the power saving and delay performance are addressed and evaluated. Then, a different power saving scheme is proposed in Section IV for TDM-PON. This scheme uses adaptive mechanisms according to the operating conditions, aiming to improve the TDM-PON upstream delay performance. In Section V, we evaluate and compare the achievable power saving and packet delay performance, for both PtP and TDM-PON based on the proposed power saving schemes. Section V presents some guidelines for the most energy efficient optical access network based on the targeted power saving, the bandwidth and QoS.

II. OVERVIEW OF P OWER S AVING M ECHANISMS A. Dozing and Sleep Modes The most intuitive way to save power in a network is to turn off the network element whenever it is idle. Two power saving mechanisms for ONUs are recommended by ITU-T [3], [4]: (i) dozing mode and (ii) sleep mode. In dozing mode, the ONU turns off non-essential components (such as the Gigabit Ethernet (GbE) PHY and optical transmitter) while maintaining the receiver active to receive incoming packets. In contrast, sleep mode allows the ONU to turn off most of its functional blocks (including receiver) to minimize the power consumption. However, after the sleep duration expires, the ONU requires a certain period of time to recover from the sleep state to the active state. This period of time is the wakeup transition time (Twake ) for resynchronization of incoming data. The Twake is mainly determined by the hardware with a typical value between 0.6 ms and 14 ms [7]. This is the main bottleneck that limits the power saving and delay performance. The power saving performance of an ONU is directly affected by the ratio of the time spent in the sleep state to the time spent in the active state. We use Tsleep to denote the sleep duration, which is governed by the traffic load and QoS requirements. To maximize the achievable power saving, Tsleep should be as long as possible. However, with stringent QoS requirements, a short Tsleep is necessary to guarantee a tolerable additional packet delay attributed to sleep mode mechanisms. In practice, depending on the QoS requirements of different services, the sleep duration can be either fixed (fixed-cyclic sleep mode (FCSM)) or dynamic (dynamic sleep-duration adjusted sleep mode (DASM)).

FCSM is one of the most fundamental sleep mode mechanisms due to its simplicity [8], [9]. In FCSM, the downstream packets destined to a specific FCSM-enabled ONU are buffered during the Tsleep and are forwarded when the ONU is fully awakened. The same mechanism applies to the upstream transmission while upstream packets are buffered at the ONU during its Tsleep . Consequently, this mechanism introduces additional packet delay to both downstream and upstream traffic. One approach to reduce the packet delay is to interrupt Tsleep based on the incoming packet [4], [10], which is known as the interrupted FCSM. Interrupted FCSM is able to achieve better delay performance at any traffic condition [10], [11]. However, it can only be initiated by the upstream traffic. Furthermore, since waking up an ONU always incurs an overhead Twake with relatively higher power consumption, the interrupted FCSM has less opportunity to stay in the low power mode and demonstrates a poorer power saving performance compared to the conventional non-interrupted FCSM. C. Dynamic Sleep-Duration Adjusted Sleep Mode (DASM) The main idea of the DASM is to dynamically-adjust the value of Tsleep based on the traffic condition [9]. The length of Tsleep is predicted by the OLT based on the traffic conditions [12]–[14]. This approach can be applied to both downstream and upstream traffic. However, the prediction of the sleep duration based on the DASM at low traffic condition can result in a large packet delay due to large variation of the inter packetarrival time during the low traffic condition. In an attempt to attain better power savings, the following two solutions have been suggested (i) shortening the wake-up time Twake [14]–[17], and (ii) applying constraints to Tsleep to prevent ONUs from either exiting the sleep state too frequently or staying in the sleep state for too long [12], [13]. Shortening Twake , considered as a hardware-based technique, requires either new circuits or architecture, which increases the CAPEX and hence it is not feasible for the already deployed PON systems. In this paper, we mainly focus on the case where the hardware is fixed with a specific Twake and investigate software-based techniques as they are viewed as the most promising power saving strategy for existing PONs [18]. D. Energy-Efficient Ethernet for Ethernet Aggregators The IEEE 802.3az Energy Efficient Ethernet (EEE) has been standardized to reduce power consumption of the Ethernet PHY layer transceiver by introducing the low-power-idle (LPI) mode during idle state [19]–[21]. EEE enables the Ethernet PHY to enter LPI dynamically according to the real-time traffic. However, the wake-up overhead due to the copper-based PHY limits the power saving performance [22]. An analytical model of EEE shows its power saving is limited to 50% of full traffic load even though its wake-up overhead is relatively shorter than that of the FCSM ONUs [23]. The actual power saving measurement reported in [26] is even less than 5% of its full capacity. Although the use of optical front-end with a

LI et al.: DYNAMIC POWER MANAGEMENT AT THE ACCESS NODE AND CUSTOMER PREMISES

Fig. 2.

1577

Proposed power saving scheme for PtP access networks.

shorter wake-up overhead has been proposed to attain better power saving performance for EEE-EAs, this hardware-based technique also increases the CAPEX [24]. In Sections III and IV, two different power saving schemes are proposed for PtP and TDM-PON systems based on a combination of the different power saving mechanisms described above. III. P ROPOSED P OWER S AVING S CHEME FOR N ETWORK E LEMENTS IN PTP N ETWORK In PtP systems, each OLT AI is connected to a single ONU. This dedicated link architecture enables a simple handshake scheme and allows the application of power saving mechanisms at both terminals. As shown in Fig. 2, we propose to use two sleep-based power saving mechanisms in the PtP systems. FCSM is applied to both ONUs and OLT AIs, while EEE is applied to EAs. We then use simulation models of the proposed PtP system to evaluate the power saving and delay performance of each of these power saving mechanisms. A. Evaluation of FCSM in PtP ONU and PtP OLT-AI In our proposal, FCSM is chosen for the existing PtP access network since it provides a good trade-off between power saving and delay performance. The average additional packet delay is the average time delay between the packet arrival time at the ONU during the sleep period and the time of transmit. The simulation models and key parameters are described in Appendices I and II, respectively. The simulation model uses Poisson distribution to generate different traffic loads. Fig. 3(a) shows the impact of applying FCSM to a Gigabit Ethernet (GbE) PtP ONU with different values of Twake . Hence, shortening Twake results in both a better power saving and an improved delay performance. It is noted that since FCSM is applied to both OLT AI and ONU in PtP access networks, the estimated average packet delay for both upstream and downstream are the same. The power saving and delay performance of FCSM is also affected by the packet size (ps). As shown in Fig. 3(b), the power saving performance is dramatically degraded if the size is reduced from the maximum Ethernet packet size (1518 Bytes) to the minimum Ethernet packet size (64 Bytes). This is due to the reason that a smaller packet size will result in a higher packet arrival rate (which degrades the power saving performance) under the same traffic load. Therefore, using larger packet size could potentially reduce the inter-packet arrival rate, and hence improves the power saving performance. For

Fig. 3. Power saving and delay performance of a FCSM-enabled 1G PtP ONU with (a) different Twake values and (b) different packet sizes (ps).

example, the use of Ethernet jumbo frame (9000 Bytes) enables the ONU to benefit further on the power saving at higher traffic loads as shown in Fig. 3(b). At a low traffic load range, the average additional delay is approximately equal to half of the sleep duration as shown in Fig. 3(a) and (b). It is noted that the delay slightly increases to a value larger than half of the sleep duration when the traffic load increases to low-medium traffic load range. This is because the packet arrival interval is close to the sleep trigger time. In order to be able to enter the sleep state, the average packet arrival interval needs to be larger than the sleep trigger time. If the average packet arrival interval is close to the sleep trigger time, the average additional packet delay will be close to the sleep duration time. When the traffic load is high, the ONU has less chance to enter the sleep mode and hence less average additional delay will be incurred by the sleep mode operations. The power saving performance of the PtP OLT AI exhibits a similar behavior as compared to the ONU. The plots for the PtP OLT AI are therefore omitted due to space constraints. B. Performance Evaluation for EAs Under EEE In terms of the maximum achievable power saving, for an EA with small number of ports, the power saving can be as high as 70% by using of EEE [25]. With the increase in the number of ports, power saving of approximately 30–50% of its full power can be achieved for EA [25], [26]. This is because a significant amount of power is consumed by multiple stages of

1578

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 32, NO. 8, AUGUST 2014

Fig. 5.

Fig. 4. Power saving and delay performance of (a) 1 GbE PtP link and (b) 10 Gbps Ethernet link under different power saving mechanisms.

switches and control blocks for traffic management and packet processing in EAs. Similar to the sleep mode mechanism, the power saving performance of EEE is also affected by the wake-up overheads and the packet size. In an EA, the wake-up overhead to switch a GbE-PHY port to active is in the order of micro-seconds [23], which enables DASM to be applied to EAs. EEE with packet coalescing, where multiple packets are grouped and processed in batches, was proposed in [19] to improve the power saving performance. Coalescing multiple packets allows the EA PHY to stay longer in the low power state. However, waiting for packets to accumulate at a low traffic load also results in an unacceptably large packet delay. Therefore, we propose a coalesced FCSM mechanism for EEE which mitigates both the delay and power consumption issues as shown in Fig. 4. The main idea of this mechanism is to reduce the packet arrival rate by coalescing multiple packets during the sleep triggering time (similar to the Ttrigger shown in Appendix I). This increases the opportunity of entering the sleep mode, defined as low-power-idle (LPI) state in EEE. Once the EA PHY exits from the LPI state, it starts sending backlogged packets. The packet transmission occurs even if the targeted number of coalesced packets cannot be reached. This guarantees the delay performance by setting an upper bound, which equals to the LPI duration. This overcomes the large coalescing delay issue as observed in [19]. The conventional EEE synchronizes each pair of EA physical layers during the refresh period and the time spending in sleep period is considerably short with only a few or tens of microseconds. The proposed coalesced FCSM for the EA only requires the resynchronization after a longer sleep period and hence reduces the refresh period and power consumption. However, the drawback of this mechanism is the degraded maximum

Proposed power saving scheme for TDM-PON access networks.

achievable power saving and the increase in packet delay in low traffic condition. In Fig. 4, the power saving performance and packet delay of a 1 GbE link (Fig. 4(a)) and a 10 GbE link (Fig. 4(b)) with standard EEE mechanisms are compared. Since the 10 GbE link has a shorter wake-up overhead than that of the 1 GbE link, it exhibits better power saving and packet delay performance. The number of packet coalesced for EA under coalesced FCSM is assumed to be 5 in our simulation because this number gives a good balance between the power saving and delay performance. A larger packet coalesced number could potentially improve the power saving performance, but due to short wakeup overhead of the EA transceiver, this improvement is not significant. Furthermore, larger packet coalesced number will increase the delay dramatically. In contrast, a smaller packet coalesced number could improve the delay performance with the trade off of worse power saving performance. IV. P ROPOSED P OWER S AVING S CHEME FOR N ETWORK E LEMENTS IN TDM-PON In this section, we propose a power saving scheme for the entire TDM-PON system as shown in Fig. 5. The ONU operates in a hybrid sleep mode, which dynamically switches between FCSM and DASM according to the traffic conditions. As we will explain in this section, it is challenging to implement the sleep-based power saving mode, especially the FCSM, to the OLT AI in TDM-PON systems. Therefore, we propose to operate the AI in dozing mode to provide a better balance between the upstream packet-delay and power savings performance. The details of this proposed power saving scheme is explained later in the section. A. Operational Scheme and Delay Performance of Hybrid Sleep Mode for ONUs In TDM-PONs, it is essential to keep the optical receiver of the OLT’s AI active at all times to minimize the incoming upstream packet delay. Therefore, dozing mode is proposed to be applied to the OLT’s AI. On the other hand, for the ONUs, the proposed hybrid sleep mode is applied to reduce the upstream packet delay. In order to enable this hybrid mode, the existing protocols require modifications. Under the hybrid mode, the ONUs will use FCSM if there is no incoming upstream traffic because FCSM provides a better power saving. When there is incoming upstream traffic, the ONU switches to DASM and adjusts the sleep duration for better upstream delay performance.

LI et al.: DYNAMIC POWER MANAGEMENT AT THE ACCESS NODE AND CUSTOMER PREMISES

1579

Fig. 7. Power saving and delay performance of a GPON ONU with different packet scheduling for downstream traffic.

Fig. 6. (a) Flowchart for a TDM-PON under hybrid mode to switch between FCSM and DASM and (b) the length of sleep duration is computed according to the predicted upstream packet interval time.

A flowchart of the hybrid sleep mode in the ONU and dozing mode in the OLT AI for TDM-PON are illustrated in Fig. 6(a). With the presence of only the downstream traffic, the sleep duration of the ONU is fixed (Tsleep_F CSM ). During Tsleep , the OLT AI stays in dozing mode and buffers all incoming downstream packets for all sleeping ONUs. When the ONU detects upstream traffic from the user network interface, the ONU calculates and predicts the next sleep duration (Tsleep_DASM ) based on the previous packet arrival intervals. The scheduling of sleep duration for the DASM (Tsleep_DASM ) is based on [13]. As shown in Fig. 6(b), the length of Tsleep_DASM is computed according to the predicted upstream packet inter-arrival time. Tsleep_DASM is constrained by two thresholds: Ttrigger as the lower boundary and Tsleep_F CSM as the higher boundary. In other words, if Tsleep_DASM is larger than Tsleep_F CSM , Tsleep_F CSM will be applied during the next transmission cycle. Otherwise, the sleep duration for the next cycle will be dynamically adjusted based on the upstream traffic. Since the downstream and upstream data can only be sent after the expiry of Tsleep , an ONU with DASM has better delay performance compared to (non-interrupted) FCSM if the sleep duration is adjusted according to the traffic condition. B. Performance Evaluation for ONU Under Hybrid Sleep Mode Similar to the PtP ONU, the power saving and estimated average packet delay of TDM-PON ONUs under FCSM is

affected by the wake-up overhead and the packet/frame size. Considering downstream traffic, with the implementation of GPON Encapsulation Method (GEM), GTC frame in GPON (and similarly also in XGPON) allows a very efficient packaging of packets with multiple frame segmentations to form a much larger frame. Compared to an Ethernet-based frame, the GTC frame demonstrates a better power saving performance. However, since a GTC frame contains frame segments for different ONUs, scheduling plays a pivotal role in allocation of GTC frames for different ONUs. In Fig. 7 we demonstrate the impact of packet scheduling. The “best case” for GPON (solid line in Fig. 7) assumes an ideal scheduling is implemented in which the AI only transmits packets destined to a particular active ONU and buffers downstream packets separately for the sleeping ONUs. A GTC frame could be allocated to a particular ONU only if required. Nevertheless, due to the application of FCSM, the maximum additional packet queuing delay at low traffic load is constrained within the tolerant range. The “average case” for GPON (dotted line in Fig. 7) represents the packet transmission without any scheduling for the sleep mode. Hence an ONU wakes up and receives all incoming broadcasted packets, among which only a small number of packets are actually addressed to it. In other words, this can be viewed as an increase in the packet arrival rate for this ONU and thus reduces the power saving that is achieved. For the upstream traffic, the main motivation for introducing DASM in the TDM-PON ONUs is to improve the upstream packet delay while maintaining significant power savings. Ideally, the DASM should significantly improve the overall power saving and delay performances compared to FCSM because the receiver of the OLT is remained ON at all time implicating a reduced wake-up overhead. Furthermore, an ONU can turn its transmitter ON much faster than the receiver since synchronization is done at the downstream direction and therefore the ONU could transmit packets to OLT as soon as the transmitter is turned ON. However, in practice, an upstream transmission requires both the transmitter and receiver of the ONU to be turned ON due to the following reason. Based on existing PON protocols, the upstream transmission is managed by the OLT via an upstream bandwidth allocation map (BWmap) encapsulated in the downstream packet. In other words, before transmitting any upstream bursty packets, the ONU receiver needs to be

1580

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 32, NO. 8, AUGUST 2014

Fig. 8. Power saving and delay performance of an EPON ONU under different power saving mechanisms for packets with different priorities.

turned on to receive BWmap information. Therefore, the wakeup overhead of downstream traffic dominants the overall wakeup overheads regardless of upstream or downstream. With such a large wake-up overhead, the benefit of DASM is suppressed. Despite the above constraint, applying DASM in the TDMPON ONU could still improve the upstream packet delay compared to FCSM at the low to medium traffic range as shown in Fig. 8. The further enhancement of the DASM requires the design of new transmission protocol and a more detailed impact analysis, which is beyond the scope of this paper. The scheduling algorithm sets the minimum and maximum sleep duration based on an estimated packet interval time. The minimum and maximum sleep duration is set to Ttrigger and Tsleep_F CSM as shown in Fig. 6(b). The parameter TSD_max avoids a large packet delay penalty due to the DASM as shown in Fig. 8. However, a DASM-enabled ONU exhibits lesser energy-efficiency compared to FCSM as the sleep duration reduces with the increase of the upstream traffic load. C. Performance Evaluation for TDM-PON OLT AI Under Dozing Mode In a conventional TDM-PON, it is challenging to implement a sleep-based power saving mode in OLTs (e.g. the FCSM) because OLTs are constantly monitoring the status of all ONUs. In order to better understand this issue, we first apply FCSM to an EPON OLT AI, which connects to 32 ONUs with a maximum packet size of 1518 Bytes. With the OLT-AI connected to multiple ONUs with an average upstream packet arrival rate of λus at each ONU, the average packet arrival rate at the AI equals to 32 × λus . A similar scenario occurs during the downstream transmission. In other words, the OLT AI receives and transmits more packets compared to any individual ONU, and therefore it has less opportunity to enter into the sleep state. In practice, the actual power saving performance of an OLT AI with FCSM-enabled could be worse due to a large number of small size handshaking packets. For instance, more than 40 percent of the total number of packets in a server are small packets with 64 Bytes [27]. Hence the power saving is significantly deteriorated as shown in Fig. 9 (dashed line). Therefore, we propose to apply dozing mode to the OLT AI where its optical receiver is always active. Fig. 9 also shows the power saving performance for EPON OLT AI with dozing

Fig. 9. Power saving and upstream packet delay performance of an EPON AI under different power saving mechanisms. TABLE I P OWER SAVING S CHEMES FOR PTP AND TDM-PON ACCESS N ETWORKS

mode enabled. Although the maximum achievable power saving of dozing mode enabled OLT AIs is poorer than FCSM at the low traffic load, they are able to provide power saving over a wider range of traffic load. Furthermore, since applying dozing mode at OLT AIs minimizes the upstream overhead from the AI side, it is assumed that there is no additional packet delay from the AI. Note that the power saving performance of the OLT AI under dozing mode can be further optimized if all ONUs are able to sleep at the same time with same Tsleep . In summary, the proposed power saving schemes for PtP and TDM-PON access networks are shown in Table I. V. P OWER S AVING P ERFORMANCE In the previous section, we investigated the factors that affect the power saving and the additional packet delay via a model for each element in PtP and TDM-PON access networks. In this section, we consider the complete access link and the impact of different power saving mechanisms are evaluated separately for downstream and upstream traffic. A. Power Saving Model for Complete Access Networks The total power consumption consists of the power of ONUs, OLT AIs and EAs. The power consumption per subscriber for PtP and TDM-PON system is given as [1]:   PEA_P tP ,(1) PP tP = PON U _P tP + 1.5 · PAI _P tP + NEA_P tP PP ON = PON U _T DM   PAI _T DM PEA_T DM + 1.5 · + ,(2) NP ON NP ON · NEA_T DM where PON U _P tP , PON U _T DM , PAI _P tP , PAI _T DM , PEA_P tP and PEA_T DM are the power consumption of PtP-ONU, TDM-PON ONU, PtP AI, TDM-PON AI, and PtP

LI et al.: DYNAMIC POWER MANAGEMENT AT THE ACCESS NODE AND CUSTOMER PREMISES

1581

TABLE II T HE M AXIMUM ACHIEVABLE P OWER S AVING PER ACTIVE S UBSCRIBER FOR D IFFERENT ACCESS N ETWORKS

TABLE III T HE M AXIMUM ACHIEVABLE P OWER S AVING AT CO FOR D IFFERENT ACCESS N ETWORKS

EA and TDM-PON EA, respectively. NP ON is the number of subscribers that share an AI port in a TDM-PON. Typically, 32 ONUs can be served by an AI port where the bandwidth of each TDM-PON port is shared among active ONUs. NEA_P tP and NEA_T DM are the number of AI port sharing an EA. A factor of 1.5 represents an estimate of the overhead due to cooling and power inefficiency at the CO [1]. Table II shows the resultant achievable power saving for PtP and TDM-PON networks. It is calculated based on the power consumption distribution of each element across the networks [28] and the corresponding maximum power saving of each network element [2], [26]. The power consumption of each component shown in Table II includes the power attributed to optics front-ends and digital processing electronics. Details of power consumption parameters are listed in Appendix II. The total resultant power saving achieved by a PtP system and a TDM-PON system can be greater than 60% of its full power consumption. It is noted that the power of the OLT in the PtP access network can be saved by up to 23% of its original power for each subscriber, which is much higher than that of a TDMPON OLT with a power saving of approximately 4%. From the service provider point of view, as shown in Table III, up to 51% of power saving could be achieved for a PtP system and 38% for a TDM-PON by applying these power saving mechanisms to the network elements in the CO. In the other words, the benefit on OPEX saving for PtP system will be more than that of the TDM-PON. B. Downstream Performance Fig. 10(a) and (b) show the power consumption and total additional packet delay as a function of downstream traffic load for 1Gb PtP, EPON, 10G-EPON, GPON and XGPON. 1 Gbps Ethernet links (1000BASE-T) are used in the 1Gb PtP OLT EAs, EPON and GPON, while 10 Gbps Ethernet

Fig. 10. (a) Total power consumption and (b) additional packet delay per active subscriber for PtP and TDM-PON systems with power saving mechanisms at downstream traffic.

links (10GBASE-T) are used for 10G-EPON and XGPON. In general, using the proposed power saving schemes, TDMPONs demonstrate a lower power consumption compared to PtP networks, due to the fact that the OLT AIs and EAs are shared among TDM-PON subscribers. EPON has the lowest power consumption at low traffic since both the ONUs and OLT of EPON consume less power than that of GPON. The use of GTC frame structure and the optimization of packet scheduling in GTC-based TDM-PONs can result in a greater benefit from sleep-based mechanisms compared to Ethernetbased TDM-PON systems as discussed in Section IV-B. Hence, GPON/XGPON systems show lower power consumption in the traffic range between few Megabytes to hundreds of Megabytes. At high traffic load, a PtP network provides the lowest power consumption per subscriber compared to TDM-PONs as shown in Fig. 10(a). The stepwise nature of the power consumption of TDM-PON shown in Fig. 10(a) is due to the assumption that the number of connected subscribers in a TDM-PON link (NP ON ) is rescaled (reduced) in order to achieve a higher average bandwidth per subscriber. This is not observed in PtP because of the dedicated link between the OLT AI and the ONU. In other words, if the ratio between the average traffic load per subscriber and the achievable peak traffic load approaches unity, PtP network architecture is more suitable to be used to fulfill the network traffic requirements. However, if the ratio approaches zero, the TDM-PON is preferred compared to PtP system. Fig. 10(b) shows the total additional downstream packet delay of different optical access networks. The total additional packet delays are mainly contributed by FCSM for ONUs and the coalesced FCSM for EAs. An explicit trade-off is observed

1582

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 32, NO. 8, AUGUST 2014

TABLE IV E NERGY-E FFICIENT O PTICAL ACCESS N ETWORKS FOR D IFFERENT A PPLICATIONS U NDER D IFFERENT BANDWDITH R EQUIREMENTS

between the power saving and delay performance over the entire traffic load. Based on our observations for power consumption for different optical access networks, Table IV lists the recommended access networks for different Internet applications with respect to energy-efficiency. Applications like Voice over IP (VoIP), Internet gaming, and video streaming with standard resolution, which do not require high bandwidth, are better to use Ethernet-based PON system since it provides the lowest power consumption with the use of power saving mechanisms. In contrast, for high bandwidth demand applications such as standard definition TV (SDTV), high definition TV (HDTV), super HDTV, standard 3DTV or even HD 3DTV, GPON and XGPON performs better in terms of energy efficiency. However, for applications that require even higher bandwidth, i.e., Ultra HDTV or Ultra 3D TV, a PtP system is preferred. For the bandwidth requirement of more than 1Gbps (e.g. 3D Ultra HDTV), both XGPON and 10G-EPON provide similar energy efficiency performance. C. Upstream Performance Fig. 11 illustrates the power consumption and delay performance of upstream traffic for PtP and TDM-PON systems. For TDM-PON, the traffic load range over which power saving can be achieved is poorer for upstream traffic than for downstream traffic due to the use of DASM. The average upstream packet delays for different optical access networks attributed to the sleep-based power saving mechanisms are shown in Fig. 11(b). The total delay is added separately from ONUs, AIs and EAs. Since the upstream packets can be transmitted only in the active state, ONU has fully woken up and suffers the same wake-up overhead as the case for downstream packets. With such a wake-up overhead, TDMPON ONUs with DASM show an improved packet delay for upstream than that of downstream. The OLT AI with dozingmode does not introduce additional upstream delay since the receiver always remains active. For the EA, the delay due to the application of coalesced FCSM should also be taken into account.

Fig. 11. (a) Total power consumption and (b) delay performance for PtP and TDM-PON networks with power saving mechanisms at upstream traffic.

Figs. 10 and 11 can be used by network operators to compute the potential power savings of the entire optical access networks after applying different power saving schemes. However, it should be noted that in practice, network operators design their access networks based on average and peak network traffic loads, which vary over the time depending on the number of users and customer behaviors. Once the network architecture has been decided based on the network traffic profile, Figs. 10 and 11 could be used to estimate the overall network power saving and delay performances. VI. C ONCLUSION We have evaluated the performance of several power saving mechanisms for all network elements located at both the customer premises and the access node for different optical access network systems. Our results show that the best power saving performance can be achieved by using the fixed cyclic sleep mode (FCSM), with the trade-off of upstream packet delay. In contrast, the dynamic sleep-duration adjusted sleep mode (DASM) adjust the sleep duration of the ONU according to the traffic condition to improve the upstream packet delay with the trade-off of lower power saving performance. By combining the main features of both FCSM and DASM, we proposed a hybrid sleep mode for TDM-PON ONUs, which balances the power saving and packet delay performance. For the optical line terminals (OLT) for the TDM-PON, due to the shared nature and the availability of traffic scheduling, dozing mode is the preferred option in the optical access interfaces (AIs) in an OLT to achieve good packet delay and power saving performance. While for the PtP system, its dedicated link architecture allows FCSM to be applied to both the ONU and the OLT AI to achieve higher power saving at each terminal. By implementing

LI et al.: DYNAMIC POWER MANAGEMENT AT THE ACCESS NODE AND CUSTOMER PREMISES

1583

A PPENDIX II K EY PARAMETERS FOR P ERFORMANCE A NALYSIS ONUs and OLT AIs in FCSM/DASM [2], [26], [28]:

Fig. 12. (a) State transition and (b) timing diagram of FCSM in ONU.

a modified FCSM, which allows packet coalescing, the Ethernet aggregators (EAs) in an OLT attain a better power saving performance compared to conventional Energy-Efficient Ethernet (EEE). In summary, our findings suggest that the analysis of power saving performance in optical access networks should consider all components (i.e., ONU, OLT, and aggregation switch). When considering all components in the optical access network, we found that a hybrid sleep mode is the ideal candidate in maximizing network power savings while reducing packet delay. This important finding suggests that future research directions need to be focusing on improving the hybrid sleep mode mechanism such as (i) identifying the ideal packet/frame sizes, (ii) developing packet coalescing and scheduling techniques, and (iii) developing transceivers with reduced wake-up overheads. Furthermore, our analysis on the trade-off between power consumption and additional packet delay under different power saving mechanisms and different optical access networks can be used as a guideline for achieving energy-sustainable optical access networks.

OLT EAs in modified FCSM [2], [26], [28], [29]:

A PPENDIX I S IMULATION M ODELS FOR S LEEP M ODE The ONU state transition and timing diagram for cyclicsleep mode in XGPON (G.987.3) [3], [4] are shown in Fig. 12. In the Asleep state, both the receiver and transmitter of the ONU are turned off, and in all the remaining states, they are both turned on. Thus, sleep mode mechanisms generally require buffering of downstream packets at the OLT and upstream packets at the ONU while the ONU is in the Asleep state. In the SleepAware state, the ONU wakes up to check the status of traffic. It goes back to the Asleep state if there is no downstream or upstream packet. Otherwise, it returns to the ActiveHeld state when it receives a forced wakeup indication (FWI) from the OLT in case of downstream packet arrival, or by triggering a local wakeup indication (LWI) in case there is upstream traffic. Since the time used to send/receive the wake-up indication is almost negligible, the time spent in the Taware is mainly constrained by the wake-up overhead (Twake ), i.e., synchronizing the downstream data. In our model, we assume Taware = Twake and the first SleepAware period Taware (0) together with Thold is the triggering time (Ttrigger = Thold + Taware (0)).

R EFERENCES [1] J. Baliga, R. W. A. Ayre, K. Hinton, and R. S. Tucker, “Energy consumption in wired and wireless access networks,” IEEE Commun. Mag., vol. 49, no. 6, pp. 70–77, Jun. 2011. [2] A. Dixit, B. Lannoo, D. Colle, M. Pickavet, and P. Demeester, “ONU power saving modes in next generation optical access networks: Progress, efficiency and challenges,” Opt. Exp., vol. 20, no. 26, pp. B52–B63, Dec. 2012. [3] ITU-T, “GPON power conservation,” presented at the Series G Supplement 45: Transmission Systems Media, Digital Systems Networks, Geneva, Switzerland, 2009. [4] ITU-T, “Transmission convergence (TC) specifications,” presented at the G.987.3: 10-Gigabit-Capable Passive Optical Networks (XG-PON), Geneva, Switzerland, 2010.

1584

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 32, NO. 8, AUGUST 2014

[5] J. I. Kani, S. Shimazu, N. Yoshimoto, and H. Hadama, “Energy efficient optical access network technologies,” in Proc. OFC/NFOEC, 2011, pp. S-22–S-26. [6] P. Vetter et al., “Energy-efficiency improvement for optical access,” IEEE Comm. Magazine, pp. 136–144, April 2014. [7] L. Valcarenghi et al., “Energy efficiency in passive optical networks: Where, when, how?” IEEE Netw., vol. 26, no. 6, pp. 61–68, Nov./Dec. 2012. [8] N. Dinh and A. I. Walid, “Power saving protocol for 10G- EPON systems: A proposal and performance evaluations,” in Proc. IEEE GLOBECOM, Anaheim, CA, USA, 2012, pp. 3135–3140. [9] J. Mandin, “EPON powersaving via sleep mode,” presented at the IEEE 802.3 Interim Meeting, Seoul, Korea, 2008. [10] Y. Ying et al., “Energy management mechanism for Ethernet Passive Optical Networks (EPONs),” in Proc. IEEE ICC, 2010, pp. 1–5. [11] J. Li et al., “Sleep mode mechanism with improved upstream performance for passive optical networks,” in Proc. IEEE ICC, 2009, pp. 1–6. [12] S. Lei, B. Mukherjee, and S.-S. Lee, “Energy-efficient PON with sleepmode ONU: Progress, challenges, solutions,” IEEE Netw., vol. 26, no. 2, pp. 36–41, Mar./Apr. 2012. [13] R. Kubo, J.-I. Kani, Y. Fujimoto, N. Yoshimoto, and K. Kumozaki, “Adaptive power saving mechanism for 10 gigabit class PON systems,” IEICE Trans. Commun., vol. E93-B, no. 2, pp. 280–288, Feb. 2010. [14] N. Suzuki, K. Kobiki, E. Igawa, and J. Nakagawa, “Dynamic sleep-mode ONU with self-sustained fast-lock CDR IC for burst-mode power saving in 10 G-EPON systems,” IEEE Photon. Technol. Lett., vol. 23, no. 23, pp. 1796–1798, Dec. 2011. [15] S.-W. Wong et al., “Sleep mode for energy saving PONs: Advantages and drawbacks,” in Proc. IEEE GLOBECOM Workshops, 2009, pp. 1–6. [16] J. I. Kani, “Power saving techniques and mechanisms for optical access networks systems,” J. Lightw. Technol., vol. 31, no. 4, pp. 563–570, Feb. 2013. [17] N. Suzuki et al., “Ultra fast-lock burst-mode CDR technology for 10 Gb/sbased PON systems,” in Proc. 17th OECC, 2012, pp. 423–424. [18] B. Skubic and D. Hood, “Evaluation of ONU power saving modes for gigabit-capable passive optical networks,” IEEE Netw., vol. 25, no. 2, pp. 20–24, Mar./Apr. 2011. [19] K. Christensen et al., “IEEE 802.3az: The road to energy efficient ethernet,” IEEE Commun. Mag., vol. 48, no. 11, pp. 50–56, Nov. 2010. [20] P. Reviriego, K. Christensen, J. Rabanillo, and J. A. Maestro, “An initial evaluation of energy efficient ethernet,” IEEE Commun. Lett., vol. 15, no. 5, pp. 578–580, May 2011. [21] I. Seoane, J. A. Hernandez, P. Reviriego, and D. Larrabeiti, “Energy-aware flow allocation algorithm for energy efficient ethernet networks,” in Proc. 19th Int. Conf. SoftCOM Telecommun. Netw., 2011, pp. 1–5. [22] P. Reviriego et al., “Energy efficiency in 10 Gbps ethernet transceivers: Copper versus fiber,” in Proc. OFC/NFOEC, 2010, pp. 1–3. [23] M. A. Marsan et al., “A simple analytical model for energy efficient ethernet,” IEEE Commun. Lett., vol. 15, no. 7, pp. 773–775, Jul. 2011. [24] D. Larrabeiti, P. Reviriego, J. A. Hernández, J. A. Maestro, and M. Urueña, “Towards an energy efficient 10 Gb/s optical ethernet: Performance analysis and viability,” Opt. Switching Netw., vol. 8, no. 3, pp. 131–138, Jul. 2011. [25] P. Riviriego et al., “An energy consumption model for energy efficient ethernet switches,” in Proc. Int. Conf. High Perform. Comput. Simul., Madrid, Spain, 2012, pp. 98–104. [26] Cisco, IEEE 802.3az Energy Efficient Ethernet: Build Greener Networks. [Online]. Available: http://www.cisco.com/ [27] AMS-IX, AMS-IX Amsterdam Internet Exchange. [Online]. Available: https://www.ams-ix.net/ [28] E. Commission, Code of Conduct on Energy Consumption of Broadband Equipment, Ispra, Italy 2011. [29] P. Reviriego, J. A. Hernandez, D. Larrabeiti, and J. A. Maestro, “Performance evaluation of energy efficient ethernet,” IEEE Commun. Lett., vol. 13, no. 9, pp. 697–699, Sep. 2009.

Jie Li (M’12) received the B.Eng. (Hons) and Ph.D. degrees in electronic and integrated circuit design from Massey University, Auckland, New Zealand, in 2007 and 2012, respectively. In April 2012, he joined the Centre for EnergyEfficient Telecommunications, University of Melbourne, Melbourne, Australia, as a Research Fellow, where he is developing circuits and systems for lowenergy fiber access networks and wireless networks. He has published several papers in the circuit and system design for wireless transceiver and optical networks. His research interests include mixed signal IC designs and system integrations.

Ka-Lun Lee (M’00) received the Ph.D. degree in electronic engineering from the Chinese University of Hong Kong, Shatin, Hong Kong, in 2003. In May 2004, he joined the University of Melbourne (UoM), Melbourne, Australia, as a Research Fellow, where he has been actively working in the area of highspeed multiwavelength optical pulse generation for the application of ADC for optical label processing. During 2007–2010, he was with the Centre for Ultra-Broadband Information Networks, UoM, participating in an industry linkage project with NEC-Australia on extended-reach fiber access networks. Since April 2011, he has been a Senior Research Fellow with the Centre for Energy-Efficient Telecommunications, UoM. He is currently developing new approaches for low-energy fiber access networks. His research interests include optical wireless integrated access, microwave photonics, optical signal processing, and electrooptic sampling.

Chien Aun Chan (M’10) received the B.Eng. degree (Hons) in electrical and electronic engineering from the University of Adelaide, Adelaide, Australia, in 2003 and the M.Eng. and Ph.D. degrees in electrical engineering from the University of Melbourne (UoM), Melbourne, Australia, in 2005 and 2010, respectively. In 2011, he joined the Centre for Energy-Efficient Telecommunications, UoM, where he is undertaking research into the energy efficiency of the Internet, optical communications, and content distribution networks.

N. Prasanth Anthapadmanabhan received the B.Eng. degree in electronics and instrumentation engineering from Birla Institute of Technology and Science, Pilani, India, and the M.S. and Ph.D. degrees in electrical engineering from the University of Maryland, College Park, MD, USA. He was a Postdoctoral Fellow with the Wireless Networking and Communications Group, University of Texas at Austin, Austin, TX, USA. Since 2010, he has been a Researcher with Alcatel-Lucent Bell Labs, where his research focuses on developing architectures and technologies for next-generation wireline access networks. His research interests include information theory and applications, error-correcting codes, access networks, and green communications.

Nga Dinh (M’11) received the Ph.D. degree in electrical engineering from the Korea Institute of Science and Technology, Seoul, Korea, in 2009. From September 2009 to June 2010, she was a Postdoctoral Researcher with Gwangju Institute of Science and Technology, Gwangju, Korea. Since then, she has been with Alcatel-Lucent Bell Labs, Seoul, as a member of the technical staff. Her current research interests include power-saving mechanisms to improve power efficiency of telecommunication networks and measuring quality of user experience/quality of service with Internet services via laptop, smartphone, and tablet PC applications.

Peter Vetter (M’96) received the Engineering Physics Degree from Gent University, Ghent, Belgium, in 1986 and the Ph.D. degree in 1991. After a postdoctoral fellowship at Tohoku University, Sendai, Japan, he joined the research center of Alcatel (now Alcatel-Lucent) in Antwerp in 1993. He subsequently worked on liquid-crystal displays, optical interconnections, optical access, access platforms, and access network architectures, first as a Researcher and later as a Department Manager. In 2000, he became a founding member and an R&D lead for the PON technology of an Internal Venture that produced the first FTTH product in Alcatel. He also initiated and managed activities in various European research projects. From 2004 to 2008, he was the overall Project Manager of IST MUSE, which is a major European integrated project with 36 partners, covering all aspects of broadband access. Since 2009, he has been with Bell Labs, Murray Hill, NJ, USA, where he leads research on optical access, access platforms, and energy-efficient access networks. He is currently the Department Head for Access Systems at Bell Labs. He has authored or coauthored more than 60 international papers, including several that were invited. He has served in the Technical Program Committee of ECOC, NOC, and BB Europe.