At present, distributed storage systems have been widely studied to alleviate Internet traffic build-up caused by high-bandwidth, on-demand applications. Distributed storage arrays located locally within the passive optical network were previously proposed to deliver Video-on-Demand services. As an added feature, a popularity-aware caching algorithm was also proposed to dynamically maintain the most popular videos in the storage arrays of such local storages. In this paper, we present a new dynamic bandwidth allocation algorithm to improve Video-on-Demand services over passive optical networks using local storages. The algorithm exploits the use of standard control packets to reduce the time taken for the initial request communication between the customer and the central office, and to maintain the set of popular movies in the local storage. We conduct packet level simulations to perform a comparative analysis of the Quality-of-Service attributes between two passive optical networks, namely the conventional passive optical network and one that is equipped with a local storage. Results from our analysis highlight that strategic placement of a local storage inside the network enables the services to be delivered with improved Quality-of-Service to the customer. We further formulate power consumption models of both architectures to examine the trade-off between enhanced Quality-of-Service performance versus the increased power requirement from implementing a local storage within the network.
©2013 Optical Society of America
In recent years, the deployment of passive optical networks (PONs) has resulted in significant growth of on-demand services over the access segment. Studies have forecasted that global Video-on-Demand (VoD) traffic will triple by 2016 . The amount of VoD traffic in 2016 will be equivalent to 4 billion DVDs per month . Providing VoD streams whilst maintaining the required Quality-of-Service (QoS) levels will be especially challenging during peak hours with potentially hundreds of customers of a single PON, watching the same video. Furthermore, research is currently being carried out to enhance the Quality-of-Experience (QoE) for the customer with the aid of higher resolution screens, 3D effects, and higher definition audio features. Forecasts in  further predict that by 2016 High Definition (HD) internet video will comprise 79% of the global VoD. Therefore, it is to be expected that video file sizes will become larger, thus raising the bandwidth requirements to suit their functions. QoS requirements concerning network parameters for different classes of services provided over IP networks including VoD have been specified in the ITU-T and ETSI recommendations [2–4]. Further, it has been shown that customers are demanding even higher QoS attributes in studies conducted using techniques such as user opinion scoring systems . Therefore, the minimum QoS requirement levels can be expected to be more stringent in years to come.
As a solution to address the increasing VoD-related Internet traffic, distributed storage systems have been proposed to achieve higher capacities andlower network transmission costs in VoD systems . Additionally, studies on VoD and IPTV architectures highlight that delivering video content from strategic locations in the network in a distributed manner [7–9] can help improve the energy-efficiency of the network. With these benefits, the placement of a local storage (LS) server within the PON was previously proposed as a viable distributed storage solution for VoD delivery over PONs . In , the LS arrays carry a set of highly popular videos which is dynamically updated depending on customer demand using a novel last-k algorithm to measure the popularity of the considered videos. The proposed LS architecture was shown to achieve a 40% bandwidth saving in the downstream direction from the central office (CO) of the PON .
In this paper, we study and compare the QoS attributes and power requirement of PON with, and without the use of the above-mentioned LS. We carry out packet level simulations for two architectures whereby VoD services are delivered with and without the presence of LS within the PON. Further, we introduce a dynamic bandwidth allocation (DBA) algorithm to optimize packet delivery in the LS PON with the aim of enhancing QoS levels. We formulate power consumption models for the two architectures to analyze the additional power requirement that may have been introduced to the PON by the LS equipment. The power consumption of a LS is attributed to the processing and video storage of LS server and storage arrays respectively. The QoS and power consumption values are then critically analyzed to study the trade-off between the QoS performance and the network power consumption.
2. Video-on-demand over passive optical networks
Video content is generally distributed using unicast video streams from central IPTV servers or intermediate cache servers upon customer request. Due to the downstream broadcast nature of the PON, copies of the content are transmitted to every customer connected to the network, thus resulting in very inefficient downstream bandwidth utilization. Multiple customers requesting a popular video at different times will cause the same content to be distributed repeatedly, maximizing downstream bandwidth wastage in an exponential manner. Each customer connects to the PON using an interface called an optical network unit (ONU) which is located at the customer premises. In the illustrative example used in our analysis, we consider the PON to service 32 ONUs or equivalently 32 customers. This number can be increased to 64 depending on the power budget of the network. Figure 1 shows the PON architecture that was previously proposed in . The LS is placed at the remote node location of the optical distribution network. At the remote node, the LS is strategically connected to a 2 × 32 splitter enabling the LS to receive a copy of all content distributed from the Optical Line Terminal (OLT) which is located at the central office (CO). The LS also receives a copy of all video requests and content uploaded by the ONUs towards the OLT. As such, the LS has the advantage of knowing the exact downstream and upstream statuses of the PON at all times. Most importantly, the LS can broadcast content within the PON in a localized manner. It is equipped with a transmitter allowing it to broadcast data on a separate wavelength, denoted λS. Correspondingly, the ONUs are equipped with an additional photo-detector to receive content on λS. Downstream communication is a broadcast amongst all connected ONUs on λD. Time division multiplexing is used to efficiently and equally share downstream bandwidth. Upstream communication is governed by the IEEE 802.3ah standard Multi-Point Control Protocol that uses REPORT and GATE control frames for bandwidth request and allocation . During operation, the ONU sends a request to the OLT on λU requesting for a specific video, Video h. The LS receives a copy of this request on λU. If Video h is available in its library, the LS will deliver Video h on λS. During operation, the OLT will keep track of the videos that are stored in the LS and will ignore a request on the knowledge that the LS will service that particular request. If the requested video, e.g. Video j, is not available in the LS, the LS will ignore the request and will allow the OLT to deliver Video j on λD. In this case, if Video j is recognized as a new popular movie by the last-k caching algorithm at the LS, the LS will store the contents of Video j it receives on λD in its storage array. Subsequently, when another ONU requests for the same video, the LS can service this request locally without utilizing additional bandwidth on λD. As a result of this continuous process, bandwidth usage for video delivery of popular moves is limited to λS, thus allowing bandwidth on λD to be made available for other downstream services.
3. Dynamic bandwidth allocation algorithm for local storage PONs
In this section, the proposed dynamic bandwidth allocation (DBA) algorithm that optimizes packet delivery in a PON with LS, is presented. The time division multiplexing access (TDMA) scheme is used in downstream direction on λD to effectively share the bandwidth amongst the ONUs. Upstream communication is managed by the centralized OLT granting time windows for each ONU to upload its content based on the reported queue sizes. The proposed architecture in Fig. 1 uses REPORT and GATE control frames to manage upstream communication within the PON as suggested by IEEE 802.3ah standard . REPORT control frames are sent upstream on λU by each ONU to communicate its bandwidth requirement to the OLT. The OLT collects all REPORT frames and allocates bandwidth as required for the next cycle according to the DBA. These allocations are then communicated downstream to the ONUs using GATE control frames on λD.
According to the IEEE 802.3ah standard, the DBA is left open for vendor implementation. The formats of control frames are adjustable to suit the services delivered by the PON. Here, we modify the frame format of both GATE and REPORT Multi-Point Control Protocol (MPCP) control frames. In the REPORT control frame message, one octet is allocated to carry the requested video identity number from the ONU to the OLT. The format of the REPORT control frame is adjusted to accommodate the VoD service parameter, the Requested Video ID, as highlighted in Fig. 2 . The opcode specific fields, which can carry a multiple number of report bitmaps and queue sizes, depending on the number of queue sets, are generally allocated 40 octets by the standard. One octet is taken out from this set to be used as the Requested Video ID carrier to the OLT. This adjustment will eliminate one cycle from the initial request communication between the ONU and OLT, which in return will directly help improve the delay performance of the network.
Similarly, in the GATE control frame, one octet is allocated to carry VoD content information from the OLT to the LS, as highlighted in Fig. 3 . This information is used to communicate the popularity information derived by the Last-k algorithm at the OLT, to the LS. If the transmitted video is recognized by the caching algorithm to be a popular video, LS will store a copy of the video in its storage arrays during the first downstream delivery on λD. Maintenance of the most popular set of videos in the LS is thus optimized since it does not require additional bandwidth on λD other than what will be already utilized for the first downstream delivery of that video. This improvement is only possible due to the above mentioned format adjustment.
The algorithm by which the VoD service is delivered is illustrated further using the timing diagrams shown in Fig. 4 . Figure 4(a) illustrates how a video request is handled in the PON when the requested video is not available in the LS video library. Figure 4(b) shows how the same request is handled when the LS library contains the requested video. When a video is requested, the ONU will add the Requested Video ID to the REPORT control frame to be sent next. This frame is then uploaded to the OLT in the next transmission window. Therefore, the initial waiting time for a request at the ONU is limited to at most one cycle. This enhancement is only possible because of the proposed format for the REPORT frame which consists of a field dedicated to accommodate the video ID. Without the format adjustment highlighted in Fig. 2, sending the Requested Video ID to the OLT will extend over two cycles, i.e. one cycle for requesting for a timeslot to upload the data portion carrying the Requested Video ID and another cycle for the actual transmission of the Video ID. Once the OLT receives the REPORT frame with the Requested Video ID, it will search for the requested video in the LS content index which the OLT manages for reference. If the requested video is unavailable in the LS library, the OLT will begin to broadcast the video packets in the next cycle as shown in Fig. 4(a). In the case where the LS library contains the requested video, the OLT ignores the request allowing the LS to service the request on λS. The OLT continues its transmissions from the next cycle as shown in Fig. 4(b). The LS transmission differs from that of the OLT mainly because the time division multiplexed bandwidth on λS is shared amongst only those ONUs which are receiving video from the LS. In comparison, the OLT will allocate downstream timeslots for each and every ONU irrespective of their active or idle status. The generic steps of the above mentioned DBA algorithm is represented in the flow chart shown in Fig. 5 .
4. QoS analysis and discussion
The impact on QoS attributes from placing a LS within a PON is presented in this section. In this work, we consider delay, jitter, and downstream bandwidth availability to be the main QoS attributes for VoD services. The QoS attributes of a VoD service over a PON mainly relies on the speed of video contents reaching the client ONU as quickly as possible and with minimal jitter. From a vendor’s point of view, decreasing the response time is considered very important for such video delivery schemes. The response time is highly dependent on packet delays. As explained in Section 3, our proposed architecture reduces the initial packet delay by one cycle time. Further, packet delays are attributed to transmission delays, queuing delays and processing delays. However, we consider in this work that the processing delay is negligible when compared to transmission and queuing delays. Hence, we ignore processing delays in our delay measurements. In accordance to the proposed architecture, when a LS that is located within the PON is transmitting video packets on λS to a ONU, the transmission delay will be significantly reduced as compared to the case when video is transmitted from the OLT. This is mainly due to the difference in transmission distances in the two cases. In our simulations, the OLT is considered to be 20 kilometers away from the remote node yielding a round trip transmission delay of 200 μs. The LS which is located at the optical splitter is placed 1 kilometer away from the ONU, giving rise to a round trip time of only 10 μs. This reduction is more significant in long-reach deployments whereby the round trip transmission delay between the OLT and ONU could be up to 1 ms (i.e. 80-100km between OLT and ONU). The queuing delay is the duration a packet spends at buffer queues of transmitting equipment. This value depends heavily on the buffer queue sizes of the OLT and LS. We consider exact buffer sizes in both OLT and LS server. Due to this usage of exact buffer sizes, and the fact that a majority of traffic is downstream from either the OLT on λD, or from LS on λS, queuing delay should converge to similar final average values in both cases.
As it is important to comprehend the practical aspects of how individual packets behave during the transmission, we carry out packet level simulations of the PON with and without LS to study its impact on the QoS attributes of VoD services. The network and protocol parameters used in our simulations are listed in Table 1 . The simulation programs were coded in C# and executed in a computer with .NET framework 3.5 as the run-time environment. The simulations were carried out to measure the resulting delay, jitter, and available downstream bandwidth values for the two above mentioned architectures. In order to improve the comparability of results, two extremes were simulated where in the first case all requests are serviced by the OLT at the CO and in the second case all requests were serviced by the LS.
Further, to improve the granularity of the results, the number of active ONUs using the VoD service at a given time is varied from 1 to 32. The results thus represent how the QoS attributes behave under different conditions of network load. The simulation time for each run was chosen to be 5 seconds which is equivalent to 2500 cycles (cycle time = 2ms).
The packet delay values obtained by simulating the two cases are plotted in Fig. 6 . The flat line represents the average delay values measured at the ONU when the VoD requests are solely handled by the OLT. These values do not vary with the number of active ONUs due to the time division multiplexed nature of the downstream transmission. That is, the time window for an ONU transmission is constant irrespective of the number of active ONUs. The dotted line represents the average delay values from the packets received by the ONUs on λS from the LS. Since the LS assigns timeslots in which duration is dependent on the number of active ONUs, the delay varies with different numbers of active ONUs along the horizontal axis. Nonetheless, results show that irrespective of how many ONUs are active at any given time during the day, the delay attribute observed by the ONUs is lower when the LS is present in the PON. As discussed, this improvement in the delay performance when LS is present, will be increasingly significant as the distance between the OLT and ONUs increases.
Figure 7 plots the jitter values whereby the continuous flat line and the dotted line represent values measured from the VoD packets delivered from the OLT on λD and from LS on λS, respectively. We define jitter as the variation of packet delays from their mean valueover a period of time. This variation is minimal when VoD is serviced from the OLT since the downstream transmission is a continuous process. The queuing delay of video packets will converge from the initial packet, thus making the total delay a steady value. However, in the case of LS, the LS will have to adjust its downstream timeslots on λS every time the number of serviced ONUs changes. This dynamic adjustment will interrupt the smooth packet flow on λS creating a variance from the mean delay. This will cause the LS transmissions to have higher jitter levels compared to the downstream delivery of VoD from the OLT, as indicated in Fig. 7. Jitter levels from using the LS (dotted line) is always greater than that from using the CO (flat line), irrespective of number of active ONUs. However the use of a jitter buffer at each ONU will enable such delay variations to be offset even in the worst case in which the recorded jitter is approximately 0.25 ms. In literature, jitter levels of less than 50 ms is considered to be acceptable for high-definition VoD delivery over the Internet . Therefore, the observed jitter from our simulation results satisfies this requirement.
The available bandwidth on λD for both architectures were also measured and plotted in the graph shown in Fig. 8 . The continuous line represents the percentage of available bandwidth in the case where the OLT services the video requests. This trend is due to the time division multiplexed scheme used for downstream transmission. When the number of active ONUs is increased, more timeslots will be utilized for distribution of VoD contents, thus decreasing the bandwidth available for other services. However, in a case where only one ONU is downloading a video and the rest is idle, the OLT will still use the dedicated timeslot to transmit video content, thus wasting a significant portion of downstream bandwidth. With the use of LS, bandwidth utilization of λD and λS is more efficient. The top and bottom dotted lines represent the bandwidth availability on λD and λS respectively in the case where the LS is servicing the video requests. Bandwidth on λD is almost fully available (~99.9%) for other downstream services since the LS is servicing all VoD requests on λS. Note that a very small level of bandwidth on λD is used for the transmission of GATE control frames even when no downstream video is transmitted to the ONUs. This continuous transmission of GATE control frames is required to maintain communication between the OLT and ONUs. The dotted blue line in Fig. 8 represents the available bandwidth on λS when the LS is transmitting the video content. The availability is small (~0.8% to 1.5%) due to the optimal utilization of bandwidth on λS for VoD delivery. The observed fluctuation of available bandwidth on λS as a function of increasing number of active ONUs is due to an unused remainder of bandwidth per timeslot. The effect of such unused bandwidth on delay and jitter is explained in detail using the results in Fig. 9(a) and 9(b).
Figures 9(a) and 9(b) highlight the correlation between bandwidth utilization on λS for VoD delivery, and the delay and jitter attributes respectively. At the LS, each polling cycle is divided into a set of timeslots, this number being dependent on the number of active ONUs whose video streams are delivered by the LS. It is important to note that since we model the video packets as fixed-length Ethernet packets of 1518 bytes, the queues are filled only with these fixed-sized Ethernet packets and a small portion of unutilized bandwidth per timeslot ensues. This unused allocated bandwidth per active ONU will cause the delay and jitter attributes to fluctuate depending on the number of active ONUs, as plotted in dotted red lines in Fig. 9(a) and Fig. 9(b) respectively. The trends in Fig. 9(a) and Fig. 9(b) clearly suggest that whenever utilization decreases, the delay and jitter values increase and vice versa.
5. Power consumption for VoD delivery over PONs using local storage
In this section, we study the impact on the power requirement of the PON due to the introduction of a LS in the architecture. We present models of power consumption per customer for VoD delivery over the two architectures, with and without LS. As discussed, the LS equipment connected to the splitter comprises a video server and the storage arrays that contain video data. Power consumption of the LS is attributed to the power requirements of the equipment used. We have previously published power consumption models for VoD delivery over 10 GE-PONs  to study the impact of energy-efficient ONUs. To maintain consistency and comparability, we consider the same network equipment used previously in the models presented in this work. Table 2 lists the equipment specifications used for formulating our power consumption models of this work. The power consumption per customer values for the proposed PON architecture is given by:
In (1), the parameter V is the average size of a HD video, which we consider to be 28.8 Gbit assuming an average duration of 60 minutes. The parameter N is the number of ONUs supported (N = 1024 = 32 × 10Gbps OLT line cards × 32 ONUs) where parameters Ncentral and Nlocal are the number of active ONUs using video downstream from the OLT at the CO and from the LS respectively. The parameter K is the number of 10 Gbps OLT line cards required to support N number of ONUs (maximum 32 line cards) and B = 0.008 Gbps is the bit-rate for the HD video stream. The terms on the right hand side of the equation accounts for different portions of power consumption per customer arising from CO storage arrays, CO video server, LS storage arrays, LS video server, OLT chassis and VCSEL-ONU  respectively. Here, PONU represents the combined power consumption of two receiver modules and single transmitter module at each ONU. The parameters Mcentral and Mlocal are the number of videos stored in storage arrays at CO and LS. Since the LS is dynamically updated with the most popular video content, we can consider it to store less than 10% of the videos available at the CO . In our analysis, we consider a range of conditions where the Mlocal value is varied from 1% to 10% of the Mcentral value. That is, we consider 1000 videos are to be stored at CO (Mcentral = 1000) and a specific percentage (a range of 1% - 10%) of most popular videos are stored at LS.
In the architecture without LS, the video server and storage arrays are located at the CO and all video requests are served by the CO video server. The power consumption per customer values for the PON architecture is given by:
Equation (2) describes the power consumption per customer for downstream VoD delivery from the CO. In (2), parameters V, N, K, and B are the same as in (1). Parameter M is the number of videos in CO storage and the parameter Nactive is the number of active ONUs connected to the server.
6. Power consumption analysis & discussion
The impact of implementing the LS within the PON on the power consumed per customer can be observed from the results shown in Fig. 10 . Curves of power consumed per customer from architectures with LS and without LS both show similar trends. Results from a range of percentage of videos stored in the LS, were considered. As the network scales, the power contribution from the video server, storage arrays and OLT chassis is shared amongst many ONUs thus reducing the power requirement per ONU towards saturation. Further, results indicate that the power consumption in the case of the LS architecture is always higher without LS, irrespective of the percentage of the number of videos stored in the LS. This is expected because the equipment in the LS will consume additional power. Figure 10 shows that as the percentage is increased, the curves are shifted upwards attributing to the additional power consumed by the LS storage arrays to store additional videos. Additionally, curves in Fig. 10 indicate periodic sharp edges along the horizontal axis as the number of ONUs increases. These edges are due to the additional power requirement of periodically added OLT line cards to support the increase of ONUs. As a single OLT line card can support only 32 ONUs, every 33rd ONU added to the network necessitates a new OLT line card. This periodic addition causes the above mentioned sharp increments in the power values, thus resulting in periodic edges along the curves. As highlighted in Fig. 10, we consider 5% (average of 1% - 10%) as the percentage of videos stored in the LS. We compare the power requirement of the two architectures further in Fig. 11 to study the power tradeoff in detail.
Figure 11 compares the power consumption per customer of both considered architectures along with the percentage of power consumption increment. At N = 1024, results indicate a power consumption level of about 11.6 W for the architecture with LS and 7.5 W for the architecture without LS. The percentage of increment of power consumption between the two scenarios, i.e. percentage difference between the two power values, is also plotted. At N = 1024, the plot indicates a percentage of power consumption increment of 53.8% due to the addition of the LS equipment in the network. The percentage of increment of power consumption can be viewed as a trade-off for achieving higher bandwidth availability in the LS architecture. The percentage of power increment of 53.8% can be reduced to 32.6% with a compromise of QoS through increasing the number of OLT line cards to 64, each supporting 64 ONUs, and thus increasing the maximum network size to 4096 ONUs. With this adjustment, the average delay and jitter values will rise to 2.208 ms and 0.33 ms respectively. These values are increased mainly due to downstream bandwidth being shared over moreONUs. Even though the increment of jitter seems significant, the final average jitter value of 0.33ms is still marginal enough to be alleviated by use of a jitter buffer at each ONU.
Alternatively, the power consumption in the LS architecture can be reduced by replacing the two receiver modules in each ONU by one tunable receiver module. This tunable receiver is capable of tuning between λD and λS. Such a receiver configuration potentially provides energy-savings through the elimination of one receiver module per ONU but at the expense of increased delay and jitter, and the capability to simultaneously receive packets on λD and λS. We are currently performing a comparative study of the LS architecture with two receiver modules per ONU and that with one tunable receiver per ONU, and we endeavor to report the results from this study in a future journal publication.
In this paper, we addressed the efforts in enhancing the QoS attributes of VoD customers by using local storage within the access network. We conducted packet level simulations to study the delay, jitter behavior, and the available downstream bandwidth for the customer ONUs. Results were subjected to thorough analysis in an attempt to study the implications of local storage on the QoS of VoD delivery. Results indicate that delay is reduced by the reduction in transmission delay, since the local storage is placed within the PON. The improvement in delay performance is expected to increase with increasing distance between the OLT and ONUs, especially in long-reach PON implementations However, with local storage, jitter values are increased. Nonetheless, this increment could be considered marginal enough for the jitter buffers to alleviate and to provide a smooth playback. Most importantly, each customer ONU receives content on an additional bandwidth channel on λS, doubling its available downstream bandwidth.
Further, we analyzed the power requirement introduced to the network through the use of the local storage equipment. The increment of power consumption per customer of 10.4% can be considered as the trade-off for significantly improved bandwidth availability and lower delays. The percentage increment of power can be reduced by scaling up the size of the network, resulting in minor compromises in the QoS values that are still within the QoS levels specified in the ITU-T and ETSI recommendations.
The authors would like to acknowledge Chien Aun Chan and Chamil Jayasundara for their helpful discussions on the formulation of the power consumption models and the video popularity tracking algorithm, respectively.
References and links
1. “Cisco visual networking index: Forecast and methodology, 2011 – 2016,” http://www.cisco.com/en/US/solutions/collateral/ns341/ns525/ns537/ns705/ns827/white_paper_c11-481360.pdf.
2. “End user multimedia QoS categories,” ITU-T Recommendation G.1010 (2001).
3. “Speech and Multimedia Transmission Quality (STQ); Audiovisual QoS for communication over IP networks,” ETSI ES 202, 667 (2009).
4. “Quality of Service ranking and measurement methods for digital video services delivered over broadband IP networks,” ITU-T recommendation J.241 (2005).
5. P. Dymarski, S. Kula, and T. N. Huy, “QoS conditions for VoIP and VoD,” J. Telecommun. Inf. Technol. 3, 29–37 (2011).
6. D. De Vleeschauwer and K. Laevens, “Performance of caching algorithms for IPTV On-Demand services,” IEEE Trans. Broadcast. 55(2), 491–501 (2009). [CrossRef]
7. J. Baliga, R. Ayres, K. Hinton, and R. Tucker, “Architectures for energy-efficient IPTV networks,” in Proceedings of OFC 2008, paper ThQ5 (2008).
8. C. Jayasundara, A. Nirmalathas, E. Wong, and N. Nadarajah, “Energy-efficient content distribution for VoD services,” in Proceedings of OFC/NFOEC 2011, paper OWR3 (2011).
9. C. Jayasundara, A. Nirmalathas, E. Wong, and C. A. Chan, “Improving energy efficiency of Video on Demand Services,” J. Opt. Commun. Netw. 3(11), 870–880 (2011). [CrossRef]
10. C. Jayasundara, A. Nirmalathas, E. Wong, and N. Nadarajah, “Popularity-aware caching algorithm for Video-on-Demand delivery over broadband access networks,” in Proceedings of IEEE GLOBECOM, 1 – 5 (2010).
11. “IEEE Standard for Information Technology – Telecommunications and information exchange between systems – Local and metropolitan area networks – Specific Requirements – Part 3: Carrier sense Multiple Access with Collision Detection (CSMA/CD) access method and physical layer specification,” ANSI/IEEE Standard 802.3 – 2002, http://standards.ieee.org/getieee802/download/802.3-2002.pdf.
12. Y. Chen, T. Farley, and N. Ye, “QoS requirements of network applications on the Internet,” IOS Press, Systems Management 4, 55–76 (2004).
13. E. Wong, M. Mueller, M. P. Dias, C. A. Chan, and M. C. Amann, “Energy-efficiency of optical network units with vertical-cavity surface-emitting lasers,” Opt. Express 20(14), 14960–14970 (2012). [CrossRef] [PubMed]