Bandwidth Allocation in DiffServ Enabled Ethernet Passive Optical Networks

Transcription

1 Bandwidth Allocation in DiffServ Enabled Ethernet Passive Optical Networks Dawid Nowak John Murphy Philip Perry Nortel Mervue Business Park Galway Ireland Performance Engineering Laboratory, Department of Computer Science, University College Dublin, Dublin, Ireland Performance Engineering Laboratory, School of Electronic Engineering, Dublin City University, Dublin, Ireland Abstract Ethernet Passive Optical Networks (EPONs) have attracted considerable attention from industry as they offer a simple, highly flexible and cost effective solution to the problem of providing broadband access to a customer. In this paper, a new approach to bandwidth allocation in EPONs is presented where the Optical Line Terminator (OLT) has full control over the access mechanism. This results in much simpler Optical Network Unit (ONU) architecture. It is shown in this paper that such an OLT-centric architecture offers full support for Differentiated Services and makes enforcement of Service Level Agreements possible. Extensive simulation experiments show that bandwidth allocation algorithms deployed in such a centralized environment can deliver good performance in terms of average and maximum packet delay. In this paper two new algorithms are introduced that target SLAaware EPONs and provide a good protection of offered quality of service against interference from other sources.

2 Bandwidth Allocation in DiffServ Enabled Ethernet Passive Optical Networks Dawid Nowak, Philip Perry, John Murphy May 28, 2008

3 Introduction With the increasing popularity of the Internet, the traffic generated by domestic and small business users has been growing constantly over the last few years. Various technologies have been deployed to provide broadband access to the network in the area known as the last mile and most network providers nowadays offer a broadband access based on Digital Subscriber Loop technology. A Passive Optical Network (PON) is a point-to-multipoint all optical network with no active elements in the path between the signal source and the destination. On the network side there is an Optical Line Terminator (OLT) unit which is usually placed in the local exchange and it acts as a point of access to the Wide or Metropolitan Area Network. On the customer s side there is an Optical Network Unit (ONU). An ONU can be placed either in the curb, building or home and its primary task is to convert data between the optical and electrical domains. Ethernet has been proposed as the data transmission protocol in PONs as it has been successfully used in many local area networks. Ethernet PONs (EPONs) provide no mechanisms that would enable support of traffic with different quality of service (QoS). As a result EPONs architecture is very simple, networks are easy to setup and equipment is available at a lower price. EPONs gained much attention from industry as interoperability between old and new networks can easily be maintained and legacy solutions can be used, thereby driving down the total cost of the deployment of EPON infrastructure. The standard describing EPONs was released in September 2004 [] Typically EPONs are connected in a tree topology [2] with a number of ONUs attached to a single OLT by means of an optical splitter. In the downstream transmission, the OLT can use all the available bandwidth to broadcast packets through the splitter/coupler to every ONU. Each ONU extracts packets from the stream based on the MAC address. In the upstream direction packets sent by an ONU can only reach the OLT as the optical splitter prevents an ONU from receiving packets from other ONUs. In order to avoid collisions between packets from different ONUs at the splitter, only one ONU is permitted to transmit at a time. The OLT is responsible for assigning a non overlapping time-slot to each ONU. During an off period packets are buffered and when the time arrives, the ONU transmits its data in one burst using the full channel bandwidth. One of the key demands for a modern network is their ability to provide levels of QoS tailored to suit the customer s demands. In the Differentiated Services (DiffServ) [3] architecture generally three classes of traffic can be distinguished: Expedited Forwarding (EF), Assured Forwarding (AF), and Best Effort (BE). EF services (primarily voice and video) have strict requirements and demand a constant, low end-to-end delay and jitter. AF services tend to be less sensitive to packet delay but require a guaranteed amount of bandwidth. BE class includes services that have no strong requirements regarding QoS. In [4] we introduced the topic of OLT allocating the bandwidth, however the algorithms presented were limited to proportional dynamic bandwidth allocation () and and SLA aware DBA. This work has been extended to introduce new algorithms and more extensive results. In this approach both inter- and intra-onu bandwidth allocation is performed by the OLT. Also two new bandwidth allocation algorithms based on this principle and that enable full support of Service Level Agreements (SLAs) are proposed. The centralized approach results in a rudimentary ONU s functionality. This offers two main benefits: As no access control or packet scheduling is done at an ONU, various algorithms can be deployed in the OLT without the need for reconfiguration of the equipment on the customer s side. It also allows for an SLA to be created, modified and deleted during normal network operation. As the result of the centralized approach, ONUs have a generic architecture and are less 2

4 expensive to produce. Hence, the cost of implementing EPON installations is reduced resulting in a more affordable solution for private customers and small businesses. In Section 2 we present the background information about bandwidth allocation methods in EPON. The proposed new schemes are introduced in Section 3. The setup used during simulation experiments is described in Section 4. In Section 5 the model and mathematical analysis of EPONs that was used to validate the results is shown. The results of experiments are presented in Section 6 and they are discussed in detail in Section 7, with concluding remarks in Section 8. 2 Background The bandwidth allocation algorithm is a key component of EPON s architecture and the only one that is in charge of all decisions concerned with bandwidth assignment. Two control packets GATE and REPORT are used to exchange necessary information between OLT and ONUs. 2. Polling Cycle Two main polling methods were proposed in the literature. Kramer et al. introduced in [5 7] an algorithm called Interleaved Polling with Adaptive Cycle Time (IPACT). In this algorithm the Multi-Point Control Protocol (MPCP) GATE messages were calculated and sent individually to different ONUs. Hence, the length of the cycle was variable and dependent on the size of transmission windows allocated to ONUs. Su-il Choi in [8] presented a different approach where the decision about bandwidth allocation was done using cyclic polling with fixed intervals and all GATE messages were sent at the same time. Su-il Choi also presented the comparison of the interleaved and cyclic polling dynamic bandwidth assignment algorithms. The presented evaluation showed that IPACT achieved better performance in terms of average delays. On the contrary, algorithms with cyclic polling needed on average less downstream bandwidth for transmitting control messages. Two enhanced polling schemes for TCP traffic in EPONs were presented by Chang and Liao in [9]. Both algorithms are based of IPACT and they try to improve fairness among TCP streams. 2.2 Inter-ONU bandwidth allocation The inter-onu scheduler is responsible for allocating single or many non-overlapping transmission windows to all ONUs attached to the OLT. During the allocation process the control unit might take into consideration the state the ONUs are in, although scheduling policies could be implemented that do not rely on such information. Based on MPCP REPORT messages, the intra-onu scheduler knows the exact status of the system and is able to make the optimal decision. Kramer et al. presented various types of bandwidth allocation algorithms in [0]. Assi et al. in [,2] proposed some improvements to algorithms introduced by Kramer. Bandwidth allocation scheme based on cyclic polling based was also mentioned by Yang et al. in [3]. They proposed Delta-DBA in which the difference is calculated between the reports received in two consecutive cycles. Later this value is used to calculate the new length of the transmission window. A different approach was presented by Lee et al. in [4]. The authors introduced a two step scheduling algorithm that supported two different bandwidth allocation policies: Static () and Dynamic (DBA) Bandwidth Allocation. A different approach to the problem of optimal bandwidth allocation was proposed by Zhu et al. in [5]. The authors introduced the idea of dividing ONUs into two disjoint groups namely Bandwidth Guaranteed (BG) and 3

5 Bandwidth non-guaranteed (non-bg). ONUs are classified into one of these groups based on requirement s agreed in the SLA such as minimum peak rate and the maximum waiting time an ONU can tolerate. The approach presented by Zhu et al. is specific in the sense that the whole connection admission control process is centralized and executed in OLT and this algorithm gives good performance without any intra-onu scheduling. A similar approach was proposed by An et al. in [6]. In the Hybrid Slot-Size/Rate DBA (HSSR) protocol, the authors suggested dividing the available bandwidth into two parts. An approach based on an analogous principle was published in [7], where Xie et al. proposed a Two-Layer Bandwidth Allocation (TLBA). The difference between these two methods was that instead of fixed size slots, a dynamic allocation was used in TLBA. The main problem with the algorithms introduced in [4 7] lies in the fact that tight cooperation between the OLT and ONUs is assumed. Moreover, an ONU must know precisely which class of traffic to send in a particular time slot. Such an assumption is contradictory to the current EPON standard and it might lead to somewhat restricted compatibility with other bandwidth allocation schemes. On the other hand centralized allocation has benefits and a way should be introduced that allows OLT to inform ONUs about the order in which the queues should be serviced. The different approach was presented by Naser et al. [8]. In their Class-of-service Oriented Packet Scheduling the DBA allocates bandwidth based on the credits assigned to different types of traffic. An outline of adaptive scheduling algorithm for distributed inter and intra-onu algorithms was presented by Zheng and Mouftah in [9]. An extensive analysis of bandwidth allocation algorithms in two stage EPONs was proposed by Shami et al. [20]. Authors introduced penetrated-epon networks that offer a longer reach. 2.3 Intra-ONU bandwidth allocation As opposed to the inter-onu scheduling the scope of this part of the allocation method is only restricted to an ONU. It is postulated in the EPON standard that an ONU is free to use the allocated transmission window to the best of its ability. This means that different policies might be deployed and moreover different ONUs might operate under the control of different schedulers. Achieving an efficient DiffServ support has always been a primary concern in designing bandwidth allocation algorithms for EPONs. Kramer et al. postulated incorporation of strict priority queueing scheme in EPONs [5,0]. It was also pointed out that implementing strict priority scheduling with the IPACT algorithm leads to a light load phenomenon, when the average delay time for lightly loaded networks is larger than for medium loaded networks. The authors provided an explanation of this paradox in [5] and they proposed a two stage buffering scheme and partial packet reordering to improve EPON performance. An improved version of a priority scheduler was proposed by Assi et al. in [,2]. An ONU with a separate queue dedicated to each class of traffic was proposed in [7]. The authors postulate the use of a weighted Random Early Detection (WRED) [2,22] mechanism to drop packets in order to monitor admission control and avoid congestion. Ghani et al. in [23] proposed to adapt algorithms used for Connection Admission Control in ATM and IP networks to EPON needs and introduced a modified version of Start-Time Fair Queueing [24,25] which was less complex than other virtual queueing systems where all packets are time stamped. Inter-ONU bandwidth allocation adds a new level of complexity to the functionality of the ONU. By moving the inter-onu scheduling to the OLT a simpler and more generic structure is achieved. 4

6 3 Proposed Schemes The functional analysis of EPON requires looking at its architecture from two different viewpoints. In the downstream direction EPONs behave like Ethernet networks and packets are broadcast to all receivers. In the upstream direction they have to be considered as point-topoint networks, as only one ONU is permitted to send at a time. This creates the problem of the fair allocation of the resources and it must be ensured that only one ONU is transmitting at a time. The OLT, being the central unit, is responsible for assignment of the non-overlapping time slots to ONUs in a static or dynamic manner. In contrast to the algorithms presented in Section 2, this paper presents the idea where the OLT is responsible for both inter- and intra-onu bandwidth allocation. We believe that such an approach offers many benefits: The OLT has full knowledge about the system; not only the current state but also the history of the network. Simple and generic ONU architecture. Effective bandwidth assignment algorithms, tailored to a specific situation, can be deployed at any time without any modification on a customers side. Centralized approach facilitates the introduction of SLAs and offer better support for DiffServ. As the number of ONUs that can be connected to the OLT is limited to 32 due mostly to the signal insertion loss at the splitter [0, 26] we do not envisage problems with scalability of centralized approach. Implementing the centralized bandwidth allocation requires introducing a new way of communication between the OLT and ONUs. In order to make per queue bandwidth allocation possible and separate the functionality of ONUs from the OLT, we propose a modified version of the GATE message whose format is presented in Fig.. The structure of the message is symmetrical to the format of the REPORT message and it allows assigning time slots for up to 8 queues. In this section, we show two QoS aware dynamic bandwidth allocation algorithms that represent two different approaches towards the problem of optimal bandwidth assignment. Both algorithms make a decision about bandwidth allocation in the next granting cycle based on the latest reports received from ONUs. The length of the granting cycle in which all ONUs can transmit their data is variable and its duration is directly dependent on the total length of buffered packets. The maximum duration of the granting cycle is set at 2 ms as similar values were used in the models published in the literature. Ensuring that the bandwidth allocation decision is based on up to date information the minimum granting cycle time must be longer or equal to the RTT measured for the most distant ONU in the network. In our scenario the minimum cycle length τ min is calculated from () where τ prop is a propagation time and τ R is time needed to transmit the REPORT message. τ min = 2 τ prop + τ R () The size of time slots in the next cycle is estimated based on the last REPORT message arrived from an ONU. The exchange of control messages is shown in Fig. 2. In the first algorithm presented here, the bandwidth is assigned proportionally to the reported queue length, whereas in the second it is allocated in an adaptive manner. 5

7 3. SLA-aware Proportional DBA Algorithm In this section the SLA-DBA algorithm is outlined. In the first phase it allocates bandwidth proportional to the reported queue length. Such an approach gives no protection to traffic parameters as classes requesting more bandwidth will receive proportionally more. This may lead to worse QoS offered to other classes of traffic. In order to secure a fair amount of bandwidth for every queue, in the second stage the constraints agreed in the SLA are taken into account. A pseudo code for this SLA aware scheduler is presented in Fig. 3. Let Q i (j) be the number of bytes reported in queue j by ONU i and βi k (j) is the number of bytes allocated to this queue in step k of the algorithm. The number of bytes that can be sent in a particular granting cycle is given by β [n] which must fulfill (2), where τ min and τ max are the minimum and maximum time of a granting cycle and C L is the link capacity in bits per second. τ min β[n] 8 τ max, (2) C L In Phase I, the amount of assigned bandwidth to a particular queue is expressed as (3): β I i (j) = Q i(j) i,j Q i(j) β[n] (3) Let γi min (j) and γi max (j) be the minimum and maximum number of bytes guaranteed to the particular queue. In the Phase II the constraints given in the SLAs are applied to all high and medium priority classes. Three distinct situations have to be considered at this stage:. βi I (j) γmax i (j) Assigned bandwidth has exceeded the amount promised in the SLA. The bandwidth allocated to a particular queue is thus reduced to βi II (j) = γi max (j). 2. βi I (j) γmin i (j) and βi I (j) < γmax i (j) Requested bandwidth is within the limits of the SLA. No changes are made and βi II (j) = βi I(j) 3. βi I (j) < γmin i (j) In a situation where there is enough of excess bandwidth β ex the allocated bandwidth is equal to γi min (j). Otherwise the amount of the allocated bandwidth does not change. The bandwidth that is not allocated during the second phase is shared among all queues in Phase III. The amount of bandwidth assigned to a queue is thus calculated from (4). βi III (j) = βi II (j) + βex βi I(j) (4) After obtaining the new bandwidth allocations for every queue the new transmission windows are assigned. The size of a new transmission window is calculated from (5) where τ(n) is the length of the cycle calculated from (6). ω [n] i (j) = τ(n) β i (j) i,j β i(j) (5) τ(n) = Qtotal C L (6) 6

8 Table : Simulator parameters Parameter name Value Link speed Gbit/s Number of ONUs 6 Number of queues per ONU 3 Inter-Frame Gap (IFG) 96 bits Guarding time µs 3.2 Adaptive DBA Algorithm The adaptive approach to the bandwidth allocation problem is presented here with the description of the A-DBA algorithm. In A-DBA it was assumed that the maximum amount of bandwidth that could be allocated to a particular class of traffic was agreed in an SLA between the network operator and a customer. As long as a particular source transmits packets at a rate lower than the maximum, they are forwarded without an additional delay. If the source exceeds the agreed maximum rate then its packets will be sent at the maximum allowed rate and surplus data will be buffered until the source decreases its rate below the maximum rate or other sources have no data to send. As in Section 3., let Q [n] i,j be the queue length in bytes reported for queue j by ONU i in granting cycle n. Let β [n] i,j be a number of bytes assigned in granting cycle n and βmax i,j be the maximum number of bytes that could be sent by a particular queue in one cycle. The granting cycle time τ(n) is dependent on the total amount of bandwidth allocated to the queues and is given by (7). Time τ(n) can not be longer than τ max which is calculated from (8). τ(n) = i,j τ max = i,j β [n] i (j) 8 C L (7) β max i (j) 8 C L (8) The block diagram of the A-DBA algorithm is shown in Fig. 4. Based on values obtained from the scheduler the OLT calculates new transmission windows from (9) where τ(n) is the length of the cycle and is calculated from (7). ω [n] i (j) = τ(n) β [n] i (j) i,j β[n] i (j) (9) 4 Simulator model To measure the performance of each bandwidth allocation algorithm an event-driven C++ based EPON simulator was designed. A number of ONUs are connected in a tree topology to a single OLT and every ONU had three queues representing different classes of traffic. Every queue had an independent buffering space. The parameters describing the simulator are shown in Table. Two types of traffic sources were used during the experiments. High priority voice traffic is generated by a source with a Poisson distribution with an arbitrary packet length. Leland [27] showed that most network traffic (i.e., http, ftp and VBR services) was best characterized as self-similar with long-range dependence. In our simulator such data is generated by a number of sources with a Pareto distribution. Different sources send packets with different lengths to ensure that the combined traffic as closely as possible resembles characteristics of the traffic 7

9 in real life networks. Measuring the performance of the introduced algorithms under different conditions was crucial. 5 Queuing Model and Analytical Analysis Here, we show the approach that was used to validate the results of the simulation experiments. The general queueing model in a downstream transmission in EPON can be simplified to a Time Division Multiple Access (TDMA) system with M independent M/D/ queues. The detailed analysis of such a system was presented in [28]. Here, we assumed that the size of the packet was equal to the size of the time slot. Also, let the rate of the link be R and packet size be fixed and equal to X. Thus, the average delay T can be expressed as T = X R + M X 2R + X R M S 2( S) where S is network utilization factor. In order to compare the results, the simulator parameters had to be adjusted so they would match the parameters of the the model: The size of the packet is equal to the size of the time slot assigned to a particular queue; only one packet gets sent per gating cycle. All ONUs operate at the same rate, all queues are equally important. Packets arrival times have Poisson distribution. Guarding times and IFG are set to zero, In Fig. 5 the comparison between the results obtained through (0) and simulation experiments is shown. It can be concluded that the results obtained from the simulator match the figures calculated from the model. Small inaccuracies are present due to the functionality of the simulator and inability to obtain exact load measures. 6 Test Cases Extensive simulation experiments have been performed for different configurations of EPONs to observe the behavior of the algorithms presented in Section 3. The performance of SLA-DBA and A-DBA algorithms was compared with the performance of other well known algorithms namely the and. The algorithm allocates a fixed share of bandwidth to an arbitrary queue. The amount allocated is based on some policy but this assignment does not depend on the current status of the queues. In this sense its functionality is similar to the Fixed approach. In the, a source is assigned a bandwidth proportionally to a queue length reported by an ONU in the last REPORT message and in principle it behaves as the Elastic approach. All algorithms were tested for various mixes of traffic generated by different sources and under different load conditions. The average delays as well as throughput were measured to provide numerical statistics. During the simulations we recorded results given by SLA-DBA and A-DBA algorithms for different classes of traffic. The total load consisted of: 20% of EF, 40% of AF and 40% of BE traffic. To model SLAs, every class was assigned some minimum and maximum number of bytes that could be transferred in one cycle. These threshold values are presented in Table 3. The recorded average delays are shown in Fig. 6 as well as average queue lengths in Fig. 6. The comparison of bandwidth utilization and average cycle length is presented in Fig (0)

10 Table 2: Traffic sources parameters Class of Service Source Source Packet length number type (bytes) EF 0 Poisson 70 AF & BE 20 Pareto 00, 500, 000, 500 Table 3: SLA Parameters SLA SLA 2 SLA 3 SLA-DBA A-DBA min max max EF AF BE 6.0 EF AF BE 6.0 EF AF BE 40.0 Limits as a % of link capacity per class of traffic. To simulate the performance algorithm under congestion the rate of 8 ONUs was increased by 00% and the average delays were compared with the delays recorded for remaining ONUs. The performance of different algorithms is shown in Figs. 9 and 0 7 Results Discussion The graphs in Fig. 6 indicate that although the performance of SLA-DBA and A-DBA algorithms is correlated that differences do exist. It is clearly visible that for EF and AF classes of traffic the performance of SLA-DBA algorithm is better for small and medium loads of up to 60-70% of the total link capacity. Conversely, the A-DBA algorithm is better for larger loads. For BE traffic the A-DBA algorithm is much better and clearly outperforms the SLA-DBA approach. The comparison of the average queue lengths in Fig. 7 reveals that the queues measured for SLA-DBA are shorter than for A-DBA. If the average cycle length is compared it can be seen that the A-DBA approach is better and generally smaller. The throughput achieved by both algorithms is comparable, but in case of SLA-DBA it is more dependent on the SLA used. The performance exhibited by both algorithms can be understood better if the architecture of both algorithms is recalled. The SLA-DBA algorithm is based on proportional allocation. Hence, it performs better for bursty traffic and is its behavior is similar to. The A-DBA algorithm assigns bandwidth in chunks which is suitable for traffic where the variation between burst sizes is small. This makes the bandwidth allocation less efficient for medium priority, data oriented services. In the SLA-DBA algorithm, improvements in QoS received by high priority classes come at the expense of smaller amount of bandwidth dedicated to BE traffic. In the the A-DBA algorithm there is no minimum of assigned bandwidth and hence, better bandwidth utilization. If SLA-DBA and A-DBA algorithms are compared with the and it can be seen that they generally outperform the other two. The algorithm gives the longest average delays and it has the worst bandwidth utilization. The performance recorded for the algorithm is similar to the performance of SLA-DBA and only for the large traffic it delivered worse QoS. 9

11 Table 4: The 95% confidence interval values Total load % Mean Error Mean Error ± ± ± ± ± ± ± ± ± ±3.26 The centralized approach performs very well if compared with other algorithms published in literature. Kramer et al. in [0] shows the average delay of s for high loads, which is much higher that the results achieved by our algorithms. Results presented by Assi et al. [] are comparable with results presented here, although the average delay of 5 ms for 80 % of bandwidth utilization is higher than the values recorded for the centralized approach. The main goal of our research was to develop an algorithm that would have good performance in heavily congested networks, where one or more customers breached their SLA and requests more bandwidth that they originally agreed to. The results of experiments for networks where the volume of traffic significantly exceeds the link capacity are shown in Figs. 9 and 0. It is clearly seen that the algorithm offers no support for SLAs. Classes that exceed their contracts will suffer the same penalty as classes which respect the limits of their SLAs. On the contrary, the algorithm shows good protection of traffic parameters, especially for EF class of traffic where there was enough bandwidth assigned. The performance of the SLA-DBA algorithm is good and the delivered QoS depends on the implemented SLA. For high and medium priority classes the difference in QoS delivered to compliant and non-compliant sources is large. Although as the offered load increases the average delays recorded for compliant sources will start to grow, due to the limited length of the cycle and the proportional bandwidth allocation method employed in SLA-DBA. The A-DBA algorithm shows good behavior and the recorded average delays are small for sources within the limits of their SLAs. On the contrary, the non-compliant sources are heavily penalized. Further analysis of the measured delays reveals that for both SLA-DBA and A-DBA algorithms the results recorded for different SLAs show that different QoS can be delivered to different sources. For clarity of the presentation we have not shown the confidence intervals on previous figures. The values of 95% confidence intervals for various algorithms and different loads were small, mainly due to large number of simulated packets. Sample values are shown in Table 4. 8 Conclusions Currently all bandwidth allocation algorithms consist of inter- and intra-onu parts which are executed by the OLT and ONUs respectively. In this paper we presented the centralized approach to bandwidth allocation in EPONs. Our proposal is based on the assumption that the OLT is the only unit that has precise and complete knowledge about the state of the whole system. Hence, it should be responsible for making decisions about bandwidth allocation amongst ONUs. We showed that in OLT-centric EPONs it is possible to create efficient and flexible algorithms that assign bandwidth per queue rather than per ONU. We demonstrated that different levels 0

12 of service can be easily assigned to an arbitrary queue and optimized schemes can be deployed. The performance of the proposed algorithms in terms of average delays was measured and their behavior compared with other well known techniques. The centralized access control mechanism allows for smooth deployment of different algorithms and SLAs can be created, modified or updated at any time. This makes the system more flexible and efficient. In this paper two new A-DBA and SLA-DBA algorithms for OLT-centric EPONs were introduced and compared. We showed that both algorithms give good results in terms of average delays and provide good protection of assigned SLAs. The results presented here outline that the new algorithms severely penalize sources that exceed their SLAs. This allows good control of the QoS in EPONs. The results of our simulation experiments were validated through the analysis of the model of a TDMA system. The outcomes show a high agreement between the analytic and simulation approaches. References [] IEEE, IEEE 802.3ah, Sept [2] G.Kramer and B.Mukherjee, Ethernet PON: design and analysis of an optical access network, Photonic Network Communication, vol. 3, no. 3, pp , July 200. [3] S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, and W.Weiss, RFC An Architecture for Differentiated Services, Dec. 998, [4] Dawid Nowak, John Murphy, and Philip Perry, Bandwidth allocation for service level agreement aware Ethernet passive optical networks, in Proceedings of IEEE GLOBECOM, 2004, vol. 3, pp [5] G.Kramer, B.Mukherjee, S.Dixit, Y.Ye, and R.Hirth, On supporting differentiated classes of service in Ethernet passive optical network, OSA Journal of Optical Networking, vol., no. 8/9, pp , [6] G.Kramer, B.Mukherjee, and G.Pesavento, Interleaved Polling with Adaptive Cycle Time IPACT: A Dynamic Bandwidth Distribution Scheme in an Optical Access Networks, Photonic Network Communication, vol. 3, no. 3, July 200. [7] G.Kramer, B.Mukherjee, and G.Pesavento, IPACT: A Dynamic Bandwidth Distribution Scheme in an Ethernet PON(EPON), IEEE Communications Magazine, vol. 40, no. 2, pp , [8] Su-il Choi, Cyclic Polling-Based Dynamic Bandwidth Allocation for Differentiated Classes of Service in Ethernet Passive Optical Networks, Photonic Network Communications, vol. 7, no., pp , [9] Kai-Chien Chang and Wanjiun Liao, On the Throughput and Fairness Performance of TCP over Ethernet Passive Optical Networks, IEEE Journal on Selected Areas in Communications, vol. 24, no. 2, pp. 3 2, Dec [0] G.Kramer, B.Mukherjee, and A.Maislos, Ethernet Passsive Optical Networks (EPON), in IP Over WDM: Building the Next Generation Internet, S. Dixit, Ed., chapter 8. John Wiley & Sons, Inc, Mar

13 [] Ch.M.Assi, Y.Ye, S.Dixit, and M.A.Ali, Dynamic Bandwidth Allocation for Quality-of- Service Over Ethernet PONs, IEEE Journal on Selected Areas in Communications, vol. 2, no. 9, pp , Nov [2] Ch.M.Assi, Y.Ye, and S.Dixit, Support of QoS in IP-Based Ethernet-PON, in Proceedings of IEEE GLOBECOM, Dec. 2003, vol. 22, pp [3] Yeon-Mo Yang, Ji-Myong Nho, and Byung-Ha Ahn, An Enhanced Burst-Polling based Delta Dynamic Bandwidth Allocation Scheme for QoS over E-PONs, in ACM Next- Generation Residential Broadband Challenges, Oct [4] Ho-Sook Lee, Tae-Whan Yoo, Ji-Hyun Moon, and Hyeong-Ho Lee, A Two-Step Scheduling Algorithm to Support Dual Bandwidth Allocation Ppolicies in an Ethernet Passive Optical Network, ETRI Journal, vol. 26, no. 2, pp , Apr [5] Yongquing Zhu, Maode Ma, and Tee Hiang Cheng, A novel multiple access scheme for Ethernet Passive Optical Networks, in GLOBECOM IEEE Global Telecommunications Conference, Dec. 2003, vol. 22, pp [6] Fu-Tai An, Yu-Li Hsueh, Kyeong Soo Kim, Ian M. White, and Leonid G. Kazovsky, A New Dynamic Bandwidth Allocation Protocol with Quality of Service in Ethernet-based Passive Optical Networks, in OFC, Mar. 2003, vol., pp [7] Jing Xie, Shengming Jiang, and Yuming Jiang, A Dynamic Bandwidth Allocation Scheme for Differentiated Services in EPONs, IEEE Communications Magazine, vol. 42, no. 8, pp , Aug [8] Hassan Naser and Hussein T. Mouftah, A Fast Class-of-Service Oriented Packet Scheduling Scheme for EPON Access Networks, IEEE Communications Letters, vol. 0, no. 5, pp , May [9] J.Zheng and H.T.Mouftah, Adaptive scheduling algorithms for Ethernet passive optical networks, IEE Proceedings-Communications, vol. 52, no. 5, pp , [20] Abdallah Shami, Xiaofeng Bai, Nasir Ghani, Chadi M. Assi, and Hussein T. Mouftah, QoSControl Schemes for Two-Stage Ethernet Passive Optical Access Networks, IEEE Journal on Selected Areas in Communications, vol. 23, no. 8, pp , Aug [2] Sally Floyd and Van Jacobson, Random Early Detection Gateways for Congestion Avoidance, IEEE/ACM Transactions on Networking, vol. 3, pp , 993. [22] Mark Wurtzler, Analysis and Simulation of Weighted Random Early Detection (WRED) Queues, Master s thesis, University of Kansas, Lawrence, Kansas, USA, [23] N. Ghani, A. Shami, C.Assi, and Y. Raja, Intra-ONU Bandwidth Scheduling in Ethernet Passive Optical Networks, IEEE Journal on Selected Areas in Communications, vol. 8, no., pp , Nov [24] Pawan Goyal, Harrick M. Vin, and Haichen Cheng, Start-time fair queueing: A scheduling algorithm for intefrated services packet switching networks, in IEEE/ACM Transactions on Networking, Oct. 997, vol. 5, pp [25] N. Ghani and J. W. Mark, Hierarchical Scheduling for Integrated ABR/VBR Services in ATM Networks, in Proceedings of IEEE GLOBECOM, Oct [26] Klaus Grobe and Jörg-Peter Elbers, PON in Adolescence: From TDMA to WDM-PON, IEEE Communications Magazine, vol. 46, no., pp , Jan

14 Octets GRANT message REPORT message Octets 6 Destination Address Destination Address 6 6 Source Address Source Address 6 2 Length/Type(88 08) Length/Type(88 08) 2 2 Opcode(00 02) Opcode(00 02) 4 Timestamp Timestamp 4 Flags/Number of grant sets Number of queue sets Grant map Report map 4 Start time Queue Length Length Queue Length Start time Queue Length Length Queue Length 0-2 Grant map Report map 4 Start time Queue Length Length Queue Length Start time Queue Length Length Queue Length 0-2 Figure : MPCP GATE and REPORT packet formats. ONU cycle N- cycle N cycle N R:5 R ONU R ONU2 R ONU3 R ONU4 R ONU5 R ONU R ONU2 R ONU3 R ONU4 R ONU5 R ONU R ONU2 R ONU3 R ONU4 G ONU G ONU2 G ONU3 G ONU4 G ONU5 G - GATER - REPORT G ONU G ONU2 G ONU3 G ONU4 G ONU5 G ONU G ONU2 G ONU3 G ONU4 G ONU5 OLT Figure 2: Control messages exchange [27] W. Leland, M. Taqqu, W. Willinger, and D. Wilson, On the Self-Similar Nature of Ethernet Traffic (extended version), IEEE/ACM Transactions on Networking, vol. 2, no., pp. 5, Feb [28] Joseph L. Hammond and Peter J.P. O Reilly, Performance Analysis of Local Computer Networks, Addison-Wesley Publishing Company, Mar

15 :STEP I Require: Q i(j) > 0 for i = 0 to Onu do for j = 0 to Queue do Req Req + Q i(j) end for end for for i = 0 to Onu do for j = 0 to Queue do βi I (j) /Req Q i(j) β max end for end for :STEP II for i = 0 to Onu do for j = 0 to Queue do if βi I (j) γi max (j) 0 then βi II (j) γi max (j) β ex β ex + (βi I (j) γi max (j)) end if if βi I (j) γi min (j) βi I (j) < γi min (j) then βi II (j) βi I (j) end if if βi I (j) < γi min (j) then if β ex γi min (j) > 0 then βi II (j) γi min (j) β ex β ex γi min (j) else β II i (j) βi I (j) end if end if end for end for :STEP III if β ex > 0 then for i = 0 to Onu do for j = 0 to Queue do βi III (j) βi II (j)+ β ex Q i (j) end for end for end if i,j Q i(j) Figure 3: Pseudocode for SLA-DBA Scheduler. 4

16 Q [n] i (j) = 0 Q [n ] i (j) = 0 T β [n] i (j) = β[n ] i (j) 2 N Q = Q [n] i (j) Q [n ] i (j) β [n] i (j) = β [n ] i (j) + Q β [n] T i (j) > βi N max (j) β [n] i (j) = βi max (j) Figure 4: A-DBA algorithm block diagram 0 Diffrence (ms) M=24, X=500, Simulation M=24, X=500, Analysis M=48, X=500, Simulation M=48, X=500, Analysis Utilization M=24, X=500, Simulation M=24, X=500, Analysis M=48, X=500, Simulation M=48, X=500, Analysis Figure 5: Results validation 5

17 Average delay [ms] A-DBA SLA SLA-DBA SLA SL SL (a) EF Average delay [ms] A-DBA SLA SLA-DBA SLA SL SL (b) AF Average delay [ms] A-DBA SLA SLA-DBA SLA SL SL (c) BE Figure 6: Average delays comparison. Maximum granting cycle of 2 ms. 6

18 Queue length [bytes] 0.0 M.0 M 00.0 k 0.0 k.0 k 00.0 A-DBA SLA SLA-DBA SLA SL SL (a) EF Queue length [bytes] 0.0 M.0 M 00.0 k 0.0 k.0 k 00.0 A-DBA SLA SLA-DBA SLA SL SL (b) AF Queue length [bytes] 0.0 M.0 M 00.0 k 0.0 k.0 k 00.0 A-DBA SLA SLA-DBA SLA SL SL (c) BE Figure 7: Average queue length comparison. Maximum granting cycle of 2 ms. 7

19 Bandwidth utilization [%] A-DBA SLA SLA-DBA SLA SL SL (a) bandwidth utilization Cycle length [ms] 2.5 A-DBA SLA SLA-DBA SLA SL SL (b) average cycle length Figure 8: Bandwidth utilization and average cycle length comparison. 8

20 Average delay [ms] A-DBA SLA SLA-DBA SLA SL SL (a) EF Average delay [ms] A-DBA SLA SLA-DBA SLA SL SL (b) AF Average delay [ms] A-DBA SLA SLA-DBA SLA SL SL (c) BE Figure 9: Average delays comparison in congested networks for SLA compliant sources. Maximum granting cycle of 2 ms. 9

21 Average delay [ms] A-DBA SLA SLA-DBA SLA SL SL (a) EF Average delay [ms] A-DBA SLA SLA-DBA SLA SL SL (b) AF Average delay [ms] A-DBA SLA SLA-DBA SLA SL SL (c) BE Figure 0: Average delays comparison in congested networks for SLA non-compliant sources. Maximum granting cycle of 2 ms. 20