Open Access
Add to BibSonomyAbstract
The energy savings of 10 Gbps vertical-cavity surface-emitting lasers (VCSELs) for use in energy-efficient optical network units (ONUs) is critically examined in this work. We experimentally characterize and analytically show that the fast settling time and low power consumption during active and power-saving modes allow the VCSEL-ONU to achieve significant energy savings over the distributed feedback laser (DFB) based ONU. The power consumption per customer using VCSEL-ONUs and DFB-ONUs, is compared through an illustrative example of 10G-EPON for Video-on-Demand delivery. Using energy consumption models and numerical analyses in sleep and doze mode operations, we present an impact study of network and protocol parameters, e.g. polling cycle time, network load, and upstream access scheme used, on the achievable energy savings of VCSEL-ONUs over DFB-ONUs. Guidance on the specific power-saving mode to maximum energy savings throughout the day, is also presented.
©2012 Optical Society of America
1. Introduction
The exponential growth in global Internet traffic is currently being fuelled by the combination of increasing customer demand for bandwidth-intensive services, and the penetration of fiber networks into the access network segment. As the Internet expands, its energy consumption will not only exacerbate the global carbon footprint but will also significantly contribute to the operating expenditure of the network operator. Energy modeling studies of the Internet have highlighted that in the short-to-medium time frame, energy consumption is overwhelmingly dominated by the access network, particularly by the optical network unit (ONU) at each subscriber premises [1]. To reduce energy consumption, efforts in developing optical transceivers and electronic circuits of low power consumption are underway. The IEEE 802.3az [2] and ITU-T G.sup 45 [3] standards have specified idle mode operation for power savings. In an ONU, both the User Network Interface through power shedding operation, and the PON interface through doze and sleep mode operations, can be powered down.
Under doze mode operation, the ONU transmitter (TX) block which comprises the laser and driver, is powered down when no upstream traffic is to be transmitted, and under sleep mode operation, both TX and receiver (RX) blocks are powered down when no upstream or downstream traffic is observed. The RX block comprises the photodetector, front-end circuit, and back-end circuit. Out of the two modes, sleep mode is considered to be the more energy-efficient mode due to the powering down of both TX and RX blocks. However, sleep mode incurs an overhead Trec of 2 ms for clock recovery and synchronization [4]. This overhead is attributed to the time taken for the ONU to resynchronize to the central office (CO) clock after transitioning from sleep mode to active mode. Under heavy traffic conditions, a long Trec prevents an ONU from transitioning into sleep mode. In contrast, doze mode operation eliminates the overhead Trec through keeping the receiver and front-end circuitry continuously powered up. Instead, a small overhead of typically 760 ns [5] corresponding to the settling time Tsett of the TX block exists. Doze mode operation yields smaller energy savings as compared to sleep mode since only the TX block is powered down during dozing intervals.
In order to maximize energy-efficiency in an ONU, (a) the power consumed in active and power-saving modes should be minimized, and (b) the duration in which the ONU is in power-saving mode should be maximized. In this work, we address these critical points whilst maintaining the cost-efficiency of access networks through the use of 10 Gbps vertical-cavity surface-emitting lasers (VCSELs) as laser transmitters of next-generation energy-efficient ONUs. We critically investigate the energy savings arising from using 10 Gbps VCSEL-ONUs under active, doze, and sleep mode operations. To clarify, an ONU that operates in doze mode can transition between dozing (i.e. only RX is powered-up) and active intervals (i.e. both TX and RX are powered up). Likewise, an ONU that operates in sleep mode can transition between sleep (i.e. both TX and RX are powered down) and active intervals. In contrast, an active and always-on ONU has both its TX and RX continuously powered up.”
We show that in active mode, the power consumption of the VCSEL TX block is smaller than that of a DFB laser by an order of magnitude. Further, we demonstrate that the VCSEL-ONU considered in our work is able to transition between active and dozing with a fast Tsett of 330 ns, thereby maximizing the dozing duration. We also study the interdependency between network and protocol parameters, e.g. polling cycle time TCYC, network load, and upstream access scheme, and the maximum achievable energy savings of VCSEL-ONUs. We show that the polling cycle time is not only an important protocol parameter that influences delay, jitter, and upstream utilization, but also the energy-efficiency of an ONU. When using an upstream scheme based on static time division multiple access (TDMA), we show through numerical analysis that energy savings is maximized under light network loading levels and long TCYC. When using a dynamic bandwidth allocation (DBA) upstream access scheme, we show that energy savings is maximized under heavy network loading levels and long TCYC. Through the observations of our work that is presented in this paper, we provide guidance on the specific power-saving mode to use in order to maximize energy savings. We highlight that the ability to switch between doze and sleep modes in a VCSEL-ONU provides the flexibility to maximize energy savings throughout the day irrespective of the network loading level, polling cycle and upstream access scheme.
2. Characterization of VCSEL transmitter block
Figure 1(a) shows the TX block of our proposed VCSEL-ONU, comprising a high-speed 1340 nm buried-tunnel-junction (BTJ) VCSEL and a 11.3 Gbps VCSEL driver. The design and characteristics of the 1340 nm BTJ VCSEL used in this work was previously reported in [6,7]. The threshold current of the BTJ-VCSEL used in our experiments was measured to be 2.5 mA. For a bias condition of 11.5 mA, the overall power consumption was measured to be 20 mW. Since typical DFB-laser thresholds at room-temperature are ~10 mA which coincide with typical bias-currents of VCSELs for which output-powers are in the range of 3 to 4 mW, the VCSEL is the laser of choice for upstream PON transmission when considering the power efficiency at these optical power levels.
Fig. 1 (a) 10 Gbps VCSEL transmitter (TX) block. (b) Measurement of Tsett of the VCSEL TX block: Oscilloscope traces of doze mode control from ON to OFF, and the resulting VCSEL output. (c) Oscilloscope traces of doze mode control from OFF to ON, and the resulting VCSEL output.
Another important point to note is that temperature stabilization is not required for the operation of our VCSEL. This is due to the use of an integrated heat sink which provides excellent heat management. A summary of the power consumption of the VCSEL-ONU under active, doze and sleep mode operations is listed in Table 1 . Note that the values listed include only the power consumption of the PON interface but not that of the User Network Interface [8]. In active mode, the 10 Gbps VCSEL-ONU consumes a total of 3.984 W [9,10]. A comparable 10 Gbps DFB-ONU consumes 5.052 W [10–12]. The power consumed by the 10 Gbps VCSEL TX block was calculated to be 0.134 W [9,10]. In comparison, a 10 Gbps DFB TX consumes 1.202 W of power [10,11]. The RX block for both types of ONUs was intentionally chosen to be identical to highlight the energy saving benefits of the VCSEL TX. The difference in power consumption between both types of ONU in active mode therefore lies in the power consumed by their respective TX block.
Table 1. Summary of Power Consumption and Transition Times of 10 Gbps VCSEL-ONU
Also listed in Table 1 are the setting time, Tsett and recovery time, Trec of the VCSEL-ONU considered in our work. The total Tsett of 330 ns of the VCSEL-ONU was measured experimentally by summing up the individual transition times of the TX block between active and doze modes and vice versa. This is shown in Figs. 1(b) and 1(c) which compares the oscilloscope traces of the doze mode control and the resulting VCSEL output. A ‘0’ to ‘1’ transition of the doze mode control powers down the TX block where as a ‘1’ to ‘0’ transition powers up the TX block. The VCSEL is inherently well-matched to 50 Ohm circuits as its differential series resistance is typically in that range [13]. Therefore, its connection to well-designed, commercially available 50 Ohm driving circuits (such as that used in this work) results with low settling times.
Plots of the experimental bit-error-ratio (BER) measurements of a free-running and uncooled VCSEL-ONU are shown in Fig. 2 . The VCSEL-ONU is directly-modulated with 10 Gbps non-return-to-zero data (PRBS 223-1 pattern length) for back-to-back and 20 km single mode fiber (SMF) transmissions. The optical spectrum with a full width half maximum of 0.06 nm is shown inset. Considering that forward error correction (FEC) is mandatory as specified by the 10G-EPON standard [14] and in combination with the use of an avalanche photodiode at the OLT, results in Fig. 2 indicate that the VCSEL-ONU is able to support channel insertion losses specified in the standard, namely 20 dB (PR10) and 24 dB (PR20). FEC is necessary in high bit-rate links to compensate for the decrease in optical sensitivity of 10 Gbps receivers and to meet the standardized link budgets.
Fig. 2 Bit-error-ratio (BER) measurements of free-running and uncooled 10 Gbps VCSEL for back-to-back (B2B) and 20 km single mode fiber (SMF) transmission. Inset: Optical spectrum of laser output.
3. Power consumption for Video-on-Demand (VoD) delivery over 10G-EPON
The power consumption per customer for VoD downstream delivery over a 10G-EPON using VCSEL-ONUs and DFB-ONUs is compared. In our work, we consider distributed storage solutions for VoD delivery whereby video storage arrays are located at the CO, as shown in Fig. 3 . Studies of energy-efficient VoD and IPTV architectures have highlighted that distributed storage which are located close to customers minimizes transport energy requirements, while implementing centralized storage away from the customers minimizes storage energy requirements [15,16]. Since high downloads of popular content such as new release movies increase transport energy consumption, such content should be widely replicated throughout the network. Further, storing content closer to customers is now more than feasible than before given the ever-decreasing cost of data storage [17]. The storage arrays at the CO would only store frequently requested popular content, thus leverage storage energy requirements with its transport energy requirements. Equation (1) describes the power consumption per customer for downstream VoD delivery over the 10G-EPON using data from major equipment vendors as listed in Table 2 .
Fig. 3 Video-on-Demand (VoD) delivery architecture over a 10G-EPON where distributed video storage is located at the central office (CO).
Table 2. Equipment Specifications
In Eq. (1), parameter V = 28.8 Gbit is the HD video size (60 mins in length), M = 1000 is number of movies stored, N is the number of ONUs supported, K is the number of 10 Gbps OLT line cards required to support N ONUs, and B = 0.008 Gbps is the bit rate for a HD video stream, respectively. The terms on the right hand side represent the power contributions arising from the CO storage arrays, CO video server, OLT chassis, and ONU, respectively. Each line card serves a maximum of 32 ONUs and each ONU serves a single customer. The power consumption model described in (1) takes into account redundancy and a utilization of 60% to allow for some reserve capacity for storage and services. In (1), the value of PONU is dependent on whether the ONU is in DFB active (Pactive = 5.052 W), VCSEL active (Pactive = 3.984 W) or VCSEL doze (Pdoze = 3.85 W) mode. Sleep mode operation is not considered in this section of the work since downstream delivery of VoD traffic requires that the RX block to be powered up.
The power-saving effectiveness of using VCSEL rather than DFB transmitters in ONUs can be observed in Fig. 4 . Plots of the power consumption per customer for downstream VoD delivery over a 10G-EPON evaluated using (1) for VCSEL and DFB ONUs are compared. The general trend indicates that as the network scales, the power contributions from CO storage arrays, CO video server, and OLT chassis become less significant since these are shared between an increasing number of supported ONUs, N. As validated by the results in Fig. 4, the ONU becomes the dominant contributor of power consumption as the network scales up. At N = 2048, the power consumption per customer in a network that uses active DFB-ONUs is 8.6 W. By comparison, that of network that uses active VCSEL-ONUs is 7.52 W. In doze mode, the power consumed per customer is 7.38 W. The percentage of power savings arising from implementing VCSEL-ONUs over active DFB-ONUs is superimposed in Fig. 4. For N = 2048, a 12.8% of power savings can be achieved through using active VCSEL-ONUs. A further improvement of up to 14% can be achieved when these ONUs transition into doze mode.
Fig. 4 Power consumption per customer and percentage of power savings as a function of supported ONUs.
4. Energy savings of VCSEL-ONUs in doze and sleep mode operations
An impact study of the interdependency between network and protocol parameters such as polling cycle time TCYC, network load, and upstream access scheme, and the energy savings of VCSEL-ONUs, is presented in this section. In our work, we consider a 10G-EPON that supports up to 32 VCSEL-ONUs. The arrival rate of the Ethernet packets for upstream transmission at each VCSEL-ONU follows the Poisson distribution with exponential arrival times. The Ethernet packet length is chosen to be uniformly distributed between 64 and 1518 bytes, and with an average packet length of 791 bytes. The guard time between consecutive VCSEL-ONU transmissions is 1 µs [18]. In order for 32 VCSEL-ONUs to efficiently share the upstream bandwidth, we consider two distinct upstream access schemes, namely the static TDMA and DBA upstream access schemes. For the static TDMA case, each VCSEL-ONU is allocated a designated transmission slot within each polling cycle. The polling cycle, TCYC, defined as the time between consecutive transmissions from the same VCSEL-ONU, has a fixed duration and is a crucial design parameter that impacts the packet delay, jitter, and bandwidth utilization [18]. The optimum choice of TCYC is a tradeoff between satisfying the quality-of-service requirement of these different performance parameters.
For the DBA scheme, standardized REPORT and GATE messages are used to dynamically request and allocate bandwidth between the VCSEL-ONUs and the OLT. The REPORT control message is a fixed 64 byte overhead that is transmitted by the ONU once per polling cycle. Upon receiving the REPORT messages from all ONUs, the OLT sends a GATE message to each ONU allocating its bandwidth. In order to avoid upstream collisions between multiple transmitting VCSEL-ONUs, the transmission start-time of each ONU is also provided in its GATE message. Hence for DBA upstream access, depending on the requested bandwidth which in turn reflects the upstream network load, the actual value of TCYC differs from cycle to cycle. Typically, the length of a polling cycle is proportional to the total amount of demanded upstream bandwidth from all ONUs or equivalently total network load. Some DBA access schemes specify a maximum polling cycle time TCYCMAX in order to limit the allocated bandwidth per polling cycle to a maximum value. Specifying TCYCMAX also bounds packet delay to a maximum value. A complete list of network and protocol parameters used in our numerical analyses presented in this section can be found in Table 3 .
Table 3. Network and Protocol Parameters
As discussed in Section 1, the energy-efficiency of a VCSEL-ONU increases with the time spent dozing or sleeping, and the amount of power saved in these modes. The general formula for evaluating the percentage of energy savings η is given by:
In (2), Tactive is the time interval within a polling cycle when the VCSEL-ONU is on (i.e. both TX and RX blocks are powered up), Tdoze/sleep is the time interval within a polling cycle when the VCSEL-ONU is dozing or sleeping, and Pdoze/sleep is the power consumed during the dozing or sleeping. The parameter Tdoze/sleep is reduced by the overheads incurred during the polling cycle. These include Tsett in doze mode, Trec in sleep mode, and TREPORT in doze and sleep modes.For static TDMA upstream access, the percentage of energy savings η arising from implementing a VCSEL-ONU in doze or sleep mode is given by:
where L is the aggregate network load and N is the number of supported ONUs in the 10G-EPON. In our work, a normalized network load of 1 represents 10 Gbps upstream traffic that is aggregated from N = 32 ONUs. In (3), note that the overheads Tsett and Trec result in additional time spent within Tactive, and thus reducing Tdoze/sleep. For DBA upstream access, the percentage of energy savings η from implementing doze/sleep mode is given by:The energy savings from implementing a VCSEL-ONU in doze and sleep modes is summarized in Fig. 5 . Using (3), the energy savings for the case of static TDMA upstream access is evaluated. Results in Fig. 5(a) indicate that a VCSEL-ONU in doze mode is minimally dependent on the network load and value of TCYC. The variation in energy savings is small ranging from (a) a minimum of 22.9% at maximum network load and minimum TCYC, i.e. Tdoze is minimize, to (b) a maximum of 23.7% at minimum network load and maximum TCYC, i.e. Tdoze is maximized. The small Tsett incurred in doze mode ensures that tangible energy savings can still be achieved at very low values of TCYC. As for sleep mode operation, results in Fig. 5(b) show zero energy savings for TCYC < Trec (~2 ms). For TCYC > Trec, energy savings of up to 84.7% at minimum network load and maximum fixed polling cycle can be achieved. Such high energy savings is attributed to a maximized Tsleep and to the large reduction in VCSEL-ONU power consumption when in sleep mode.
Fig. 5 Percentage of energy savings η vs. network load and polling cycle for doze and sleep modes with static TDMA and DBA upstream access.
Using (4), the energy savings for the case of DBA upstream access is evaluated. Results for doze mode operation shown in Fig. 5(c) show energy savings ranging from (a) a minimum of 22.7% when network load and TCYCMAX are minimum, to (b) a maximum of 23.05% when network load and TCYCMAX are maximum. These results can be explained as follows. For the case of DBA upstream access, the length of the polling cycle is proportional to the network load. Therefore, a lightly-loaded network would result in a shorter polling cycle. If this polling cycle is less than the specified maximum of TCYCMAX, then Tdoze and hence the percentage of energy savings are minimized. As the network load is increased, the polling cycle is extended, resulting in an increase in Tdoze and the percentage of energy savings. Results obtained for sleep mode operation shown in Fig. 5(d) show a similar behavior to that of the doze mode. No energy savings is observed at least until TCYCMAX > 2 ms. Beyond that, the energy savings of up to 82.3% can be achieved when network load and TCYCMAX are maximum. Such high energy savings is attributed to a maximized Tsleep and to the large reduction in VCSEL-ONU power consumption when in sleep mode.
Comparing doze mode and sleep mode operations, the latter has always been considered as the more energy-efficient mode due to the powering down of both TX and RX blocks. However, results in Fig. 5 clearly show that depending on the network load, chosen polling cycle time, and implemented upstream access scheme, the energy savings from sleep mode can be lower than that of doze mode. We illustrate this fact by evaluating the energy savings of a 10G-EPON which network load varies throughout the day. The results in Fig. 6 highlight that rather than implementing just sleep mode in VCSEL-ONUs, the flexibility to switch between doze and sleep modes in the device will maximize energy savings throughout the day. Figure 6(a) shows a normalized upstream network load of the PON as a function of time of day (TOD). This daily profile was simulated based on upstream video requests as outlined in [19]. The upstream network load is low during early morning but peaks around 1400 hrs and 2100 hrs. To provide a fair comparison with identical maximum bounded transmission delay, TCYC of the static TDMA upstream access scheme and TCYCMAX of the DBA upstream access scheme were intentionally fixed to 2.8 ms. For the TDMA upstream access, the polling cycle is a constant 2.8 ms. For DBA upstream access, the polling cycle varies according to the network load, reaching a maximum of 2.8 ms at 2100 hrs as shown in Fig. 6(b).The energy savings from implementing doze, sleep, and active modes in a VCSEL-ONU over an active DFB ONU for the case of static TDMA and DBA upstream access scheme is illustrated in Fig. 6(c) and 6(d) respectively. The blue line in both figures represents the percentage of energy savings from implementing an active VCSEL-ONU over an active DFB, i.e. η = 100% (1- (Pactive,VCSEL/Pactive,DFB)). Results in Fig. 6(c) show that in order to maximize energy savings throughout the day, sleep mode is favored during the first half of the day whereas doze mode is preferred during the second half. Both sleep and doze modes provide improved energy-efficiency as compared to an always on, active VCSEL-ONU. The energy savings from using a VCSEL-ONU in doze, sleep, and active modes for the DBA upstream access scheme, is illustrated in Fig. 6(d). From an energy conservation viewpoint, doze mode operation provides the highest energy savings with the least variance throughout the day. This is due to the fact that for DBA upstream access, the polling cycle is constantly minimized for all network loading levels. Consequently, at low network loading levels, the polling cycle cycle may be less than Trec of 2 ms. In this case, the ONU cannot transition into sleep mode to conserve energy and therefore, energy savings is zero. Regardless of the choice of upstream access, our results indicate that using a VCSEL-ONU and more importantly one that can switch between doze and sleep modes will ensure that energy-savings of the network is always maximized for any given network load.
Fig. 6 (a) Normalized network load vs time of day (TOD), (b) Polling cycle vs TOD, (c) Percentage of energy savings from sleep and doze mode operations with static upstream access, and (d) Percentage of energy savings from sleep and doze mode operations with dynamic upstream access.
5. Summary
In this work, we addressed the efforts in reducing the power consumption of next-generation optical transceivers by critically investigating energy-savings arising from using 10 Gbps VCSEL-ONU transmitters. In terms of energy-efficiency, we showed that there are two main benefits in using VCSEL-ONUs, namely a fast setting time and low power consumption during active and power-saving modes. These benefits allow the VCSEL-ONU to maximize its duration in doze and sleep modes and to minimize its power consumption during active, doze, and sleep modes. The percentage of energy savings from using VCSEL-ONUs in doze and sleep modes over active DFB-ONUs was numerically investigated. Our calculations show that in the case of static TDMA upstream access, the energy savings from doze mode is minimally affected by network load and the fixed polling cycle. As for sleep mode, tangible energy savings is observed for polling cycles only beyond the recovery and synchronization overhead time. For DBA upstream access, energy-efficiency is maximized for heavy loads and long maximum polling cycle times. Since the polling cycle for DBA upstream access is dynamic and proportional to the network load, a heavily-loaded network and one with a long maximum polling cycle time will maximize the duration in power-saving mode.
Regardless of static or dynamic upstream access, for the VCSEL-ONU considered in our work, results show that the energy savings from doze mode operation is least affected by both network load and polling cycle. Results also show that sleep mode is not necessarily the most energy-efficient mode. The ability to switch between doze and sleep modes in a VCSEL-ONU thus provides the flexibility to maximize energy savings throughout the day irrespective of the network loading level, polling cycle and upstream access scheme. Finally, it is imperative note that the implementation of power saving modes along with the bandwidth allocation scheme in a network must not compromise its quality-of-service (QoS) performance. Our preliminary results obtained so far via packet level simulations and which we endeavor to report in a future journal publication, show that implementing sleep and doze mode VCSEL-ONUs does not affect the QoS performance of the network.
Acknowledgments
The authors would like to thank VERTILAS GmbH for providing the VCSELs used in our work, and also the Australian Research Council and Go8-DAAD Germany Joint Research Co-Operation Scheme for funding the research.
References and links
1. J. Baliga, R. Ayre, K. Hinton, W. V. Sorin, and R. Tucker, “Energy Consumption in Optical IP Networks,” J. Lightwave Technol. 27(13), 2391–2403 (2009). [CrossRef]
2. IEEE. Std, 802.3az–2010, [Online]. Available: http://standards.ieee.org/getieee802/download/802.3az-2010.pdf
3. “GPON power conservation,” ITU-T G-series Recommendations –Supplement 45 (G.sup-45), 05/2009.
4. J. Mandin, “EPON Powersaving via Sleep Mode,” IEEE P802.3av 10GEPON Task Force Meeting (2008).
5. E. Igawa, M. Nogami, and J. Nakagawa, “Symmetric 10G-EPON ONU BMT Employing Dynamic Power Save Control Circuit,” Proc. of IEEE/OSA Opt. Fiber Commun. Conf., Los Angeles, USA, NTuD5, (2011).
6. W. Hofmann, M. Mueller, G. Bohm, M. Ortsiefer, and M. C. Amann, “1.55um VCSEL with enhanced modulation bandwidth and temperature range,” IEEE Photon. Technol. Lett. 21(13), 923–925 (2009). [CrossRef]
7. A. Gatto, A. Boletti, P. Boffi, C. Neumeyr, M. Ortsiefer, E. Rönneberg, and M. Martinelli, “1.3 µm VCSEL transmission performance up to 12.5 Gbps for metro access networks,” IEEE Photon. Technol. Lett. 21(12), 778–780 (2009). [CrossRef]
8. E. Wong, M. Mueller, P. I. Dias, C. A. Chan, and M. C. Amann, “Energy Saving Strategies for VCSEL ONUs”, Proc. of IEEE/OSA Opt. Fiber Commun. Conf., Los Angeles, USA, OTu1H5 (2012).
9. Analog Devices ADN2530 data sheet, [Online]. Available: http://www.analog.com/static/imported-files/data_sheets/ADN2530.pdf
10. S. W. Wong, L. Valcarenghi, S.-H. Yen, D. R. Campelo, S. Yamashit, and L. Kazovsky, “Sleep mode for energy saving PONs: Advantages and Drawbacks,” Proc. of IEEE GLOBECOM Workshops, (2009).
11. NEL Laser Diodes, [Online]. Available: http://www.nttelectronics.com/en/products/photonics/pdf/NLK5B5EBKA.pdf
12. Analog Devices ADN2531 data sheet, [Online]. Available: http://www.analog.com/static/imported-files/data_sheets/ADN2531.pdf
13. M. Mueller, C. Grasse, K. Saller, T. Gruendl, G. Boehm, and M. C. Amann, “1.3 μm High-Power Short-Cavity VCSELs for High-Speed Applications”, CLEO/QELS, San Jose, USA, CW3N.2, (2012).
14. IEEE. Std, 802.3av – 2009, [Online]. Available: “http://standards.ieee.org/getieee802/download/802.3av-2009.pdf
15. J. Baliga, R. Ayres, K. Hinton, and R. Tucker, “Architectures for Energy-Efficient IPTV Networks,” Proc. of IEEE/OSA Opt. Fiber Commun. Conf, San Diego, USA, ThQ5 (2008).
16. C. Jayasundara, A. Nirmalathas, E. Wong, and N. Nadarajah, “Energy Efficient Content Distribution for VoD Services,” Proc. of the IEEE/OSA Opt. Fiber Commun. Conf, Los Angeles, USA, OWR3 (2011).
17. M. Rabinovich and O. Spatscheck, Web Caching and Replication: Addison Wesley (2001).
18. B. Skubic and D. Hood, “A comparison of DBA for EPON, GPON, and NG TDM PON,” IEEE Commun. Mag. 47, 540–548 (2009).
19. Y. Hongliang, Z. Dongdong, Y. Z. Ben, and Z. Weimin, “Understanding user behavior in large-scale video-on-demand systems,” Proc. 1st ACM SIGOPS/EuroSys European Conference on Computer Systems 2006, Leuven, Belgium, 1–12 (2006).
