A CLASS-BASED DYNAMIC BANDWIDTH ALLOCATION SCHEME FOR EPONS

Transcription

1 A CLASS-BASED DYAMIC BADWIDTH ALLOCATIO SCHEME FOR EPOS Jing Xie,, Shengming Jiang, and Yuming Jiang 3 Institute For Infocomm Research, Heng Mui Keng Terrace, Singapore 963, Singapore Department Of Electrical And Computer Engineering, ational University Of Singapore 3 Centre For Quantifiable Quality Of Service In Communication Systems orwegian University Of Science and Technology ABSTRACT In an Ethernet Passive Optical etwork (EPO), all optical network units (s) share one uplink channel to transmit multiple traffic. To avoid data collision, a contention-free MAC protocol is needed to schedule the transmission order of different s. An available solution is to assign a variable-sized time slot to each. In this paper, we first review some existing bandwidth allocation schemes for EPO in the literature. Then we propose a novel scheme called dynamic weight priority scheduling (), which not only guarantees all s fairly share the bandwidth of uplink channel on demand, but also supports differentiated services. To prevent one service class from monopolizing the bandwidth under heavy load, a bandwidth threshold determined by a weight ensures a minimum bandwidth for each traffic class.. ITRODUCTIO Ethernet passive optical network (EPO), which reduces the fiber deployment dramatically and preserves the merits of Ethernet networks, is gaining more and more attention due to its low cost and high performance. An EPO consists of one optical line terminal (OLT) and multiple optical network units (s), as shown in Fig.. In the downstream direction (from the OLT to the s), the OLT broadcasts Ethernet frames to all s. In the upstream transmission (from the s to the OLT), all s share one optical fiber channel. To avoid data collision, the OLT needs to schedule the transmission order of all s. An available solution is to assign a time slot to each, and each can only transmit in its assigned time slot. Before the time slot assigned to an arrives, this stores all packets from its end users into corresponding queues according to their types. When the correct time slot arrives, the transmits the stored packets at full channel speed. The OLT employs a bandwidth allocation scheme to distribute an amount of bandwidth to every on its demand. Thus, bandwidth allocation is a key issue in EPO to guarantee the bandwidth to be used efficiently and fairly by s. Since EPO is expected to support diverse applications, it needs to guarantee QoS of multiple services. Therefore, bandwidth allocation should satisfy the demand of the s and support differentiated services. Assigning a fixed time slot to each regardless of its demand is simple but not adaptive to bursty traffic. Relatively, dynamic bandwidth allocation (DBA) scheme, which assigns a variable time slot to each on its demand, is more effective and flexible. A DBA scheme requires that the OLT know the bandwidth demand of the, which is used to conduct allocation. Af- OLT Splitter/ Combiner Fig.. System Architecture of EPO ter allocation, the OLT informs each of the allocation results. To this end, the IEEE 8.3ah task force develops the Multi- Point Control Protocol (MPCP) [] to manage the control messages which communicate the information between the OLT and the s. The MPCP relies on two control messages, GATE and REPORT, to grant and request the bandwidth. The GATE message, which is sent from the OLT to the, is used to inform of the beginning time of and the size of the assigned time slot. The sends the REPORT message to the OLT to convey its local conditions, such as the buffer occupancy. Even if an need not send traffic, a minimum bandwidth is guaranteed to transmit a REPORT message. To improve the bandwidth utilization, [] presented a DBA scheme that reduces the polling cycle with an interleaved polling approach. The OLT allocates the bandwidth to an on its demand but no more than a predefined maximum value. This is called limited service which can avoid some s with high load from monopolizing the bandwidth. However, the OLT does not know how the distributes the granted bandwidth among all traffic classes and cannot support inter- priority scheduling. [3] found that the limited service causes a light-load penalty phenomenon, which exhibits that under light load the delay of low priority traffic and the maximum delay of medium priority traffic increase significantly. To eliminate this unexpected behavior, a two-stage queue at an (shown in Fig. ) is proposed. Here, stage-i consists of multiple-priority queues and stage-ii is a First- Com-First-Served (FCFS) queue. When a time slot arrives, packets in stage-ii are transmitted, and packets in stage-i enter stage-ii simultaneously. At the end of the time slot, the reports the current queue occupancy of stage-ii to the OLT. The total queue size of stage-ii is defined exactly as TxW MAX, thus, an can report at most TxW MAX and get a transmission window no larger

2 R Mbit/s (gated) W MAX T X P P P R U Mbit/s USER ferent, the subframe size is variable, which is determined by the request size of the corresponding service class. To prevent a service class from monopolizing the bandwidth under heavy load, a weight set for each class decides a bandwidth threshold, which can guarantee a minimum bandwidth to each class. After allocating the bandwidth amongst classes, the size of each subframe is determined. Within the same class, the bandwidth is fairly distributed to all s following the max-min fairness principle. Stage II Stage I Fig.. Two-stage queue at an [3] than TxW MAX accordingly. This constraint ensures that the assigned time slot is always fully used. However, this architecture augments the average queuing delay of packets due to the waiting time in the stage-ii queue. The preemptive scheme used in the ensures the serves a higher priority queue to exhaustion before serving a lower priority queue. However, within a cycle, low priority packets of s which are assigned to transmit earlier, are transmitted earlier than high priority packets of s assigned to transmit later. Therefore, the queuing delay of high priority packets may be longer than that of the low priority packets. Since high priority traffic is delay sensitive, it is more desirable to transmit all high priority packets earlier than the others. To solve the problem mentioned above, [4] introduced a novel DBA scheme, which transmits traffic of different classes separately. The frame is divided into two parts for two service classes. The first part with a constant size is reserved for high priority (HP) traffic to minimize the jitter. The remaining part is allocated to best effort (BE) data. The OLT assigns consecutive time slots to different s to transmit packets with the same type. The remaining bandwidth reserved for HP traffic of an can only be shared by its own sessions. Thus, some s may not use the allocated bandwidth fully while requirements of the others cannot be satisfied. In this paper, we propose a dynamic weight DBA scheme to support inter- and intra- priority scheduling. A frame is divided into multiple subframes for multiple service classes. During a cycle, higher priority traffic is always transmitted prior to lower priority traffic. A weight set for each service class ensures that all classes proportionally share the bandwidth under heavy load. Thus, the throughput of one traffic class only depends on its traffic load and bandwidth threshold regardless of other traffic classes. The remaining of the paper is organized as follows. Section introduces the proposed scheme in detail. umerical results are discussed in Section 3 and Section 4 concludes the paper.. SCHEME To support QoS of differentiated services and satisfy instantaneous demands of s, we propose a novel scheme called dynamic weight priority scheduling (). can not only support both intra- and inter- priority scheduling but also dynamically adjust the service class s proportion of sharing the bandwidth. The frame is divided into multiple subframes to transmit traffic of different classes. Since the bandwidth demands are dif-.. Bandwidth Allocation in OLT In our scheme, we only discuss three service classes corresponding three priorities, i.e., high, medium and low. To present simply, we use, P and P to denote these three classes respectively. Thus, the frame is divided into three subframes, which are assigned to the three classes in a decreasing order of their priorities, e.g., high priority class is always assigned the first subframe. The bandwidth demand of a service class from all s, ReBW j, is different in every cycle, therefore, the subframe size is variable to meet the demand. Only the maximum frame size is fixed, and the bandwidth allocation is dynamic within the frame. 3 3 Subframe Frame n- Frame n Subframe Subframe Subframe 3 Subframe Frame n+ Subframe Fig. 3. Frame Format of Subframe Subframe Subframe The OLT maintains a bandwidth demand table which stores the bandwidth demands of all s in an increasing order. After the OLT receives the REPORT messages from s, it updates the bandwidth demand table. Fig.. indicates that the size of the time slot assigned to i (i =,,..., ) is different in each cycle. If the total bandwidth of the uplink cannot satisfy the demands from all s, the OLT will confine some demands. A bandwidth threshold, BW threshold j (j =, and ), is defined for each service class on its weight which satisfies < weight j < (j =, and ), j= weightj =. BW thresholdj that is calculated by BW frame weight j, helps to prevent a class from occupying more bandwidth and ensure a minimum throughput for low priority class under heavy load. Here, BW frame is the available bandwidth of a frame. weight j determines the proportion of sharing the bandwidth of service class j and can be adjusted dynamically by the OLT. If ReBW j does not exceed BW threshold j, the bandwidth allocated to class j (GrantBW j) is exactly ReBW j and the remaining bandwidth (RemainBW j = BW threshold j - ReBW j) will be proportionally distributed to other classes with unsatisfied requests on their weights. Otherwise, the OLT only allocates BW threshold j to class j. Thus, under heavy load, the

3 minimum throughput of class j approximates the product of the overall throughput and weight j... Buffer Structure in The buffer consists of three queues to place packets from three service classes and does not need a second stage queue which is used in [3]. To ensure the packets with the same type from different s to be served equally, the OLT assigns consecutive time slots to different s to transmit the traffic of the same service class. An reports the lengths of its three queues to the OLT which grants at most three time slots to this accordingly. The transmission order is from the high priority traffic (i.e., ) to the low priority (i.e., P). The OLT needs to wait for REPORT messages from all s to compute the total bandwidth demand of every service class. Within a service class, the bandwidth is distributed to all s based on the max-min fairness policy, which maximizes the minimum share of a source whose demand is not fully satisfied, with the principles stated in [5]. Since three queues share a limited-size buffer, some packets have to be dropped when the buffer is full. The preemptive scheme used in [3] can preempt one or more lower priority packets stored in the buffer to place a newly arriving packet with higher priority. In our scheme, only when the length of lower priority queue is larger than a threshold, a newly arriving packet with higher priority can preempt lower priority packets. Otherwise, this new packet will be dropped. The reason of this is that the OLT assigns a time slot to lower priority traffic of an, however, before the assigned time slot arrives, some or even all lower priority packets may be preempted by the newly arriving packets with higher priority. Thus, this assigned time slot cannot be used to transmit lower priority traffic, and the throughput of lower priority traffic will decrease. To avoid this problem, a queue-length threshold is defined to ensure that the assigned time slot would be used to transmit lower priority traffic. 3. UMERICAL RESULTS To verify the efficiency of, we conduct simulation with OP- ET (a professional simulation tool) in comparison with the twostage queue scheme [3]. The simulated network includes one OLT and sixteen s. The distance from the OLT to any is fixed and equal. The link rate between the OLT and the is Gbit/s and that from end users to the is Mbit/s. Traffic with the same QoS requirements is aggregated into one service class. Three service classes are simulated. Class generates constant-bit-rate (CBR) voice traffic. Class P emulates variable-bit-rate (VBR) video streams which exhibits self-similar and long-range dependence (LRD). Class P is used to simulate non-real-time (nrt) VBR data traffic. First, we change the weight of the service class. Table lists six sets of simulation results. In this table, the unit of delay is second and that of throughput is Mbit/s. The results show that the throughput of the entire network is around 87 Mbit/s and does not vary sharply when the weights are changed. However, the delay and the throughput of the service classes are different with different sets of weight. The throughput of each class is proportional to its weight because the offered load is at the ultimate value ( Mbit/s). As can be seen in this table, if weight j= weight j /, the average packet delay of is less than ms. According to this formula, there are three sets of weight satisfy this formula, i.e., (,, ), (,, ) and ( 4,, ). Among them, set (,, ) not only distinguishes the priority of different service classes but also satisfies the QoS requirement of average packet delay. Relative to ( 4,, ), this set allocates less bandwidth to the but still gets the same throughput, and the average packet delay of P and P is lower. Since set (,, ) can get better results than the others, 3 6 the later discussed simulation results are performed based on this weight set. Throughput (Mbit/s) Throughput(Mbit/s) Throughput(Mbit/s) 7 6 Load of (a) Class Load of P (b) Class P Load of P (c) Class P Fig. 4. Throughput of a single class (all s) vs offered load

4 Table. Simulation Results of different sets of weight vs offered load equal to 6 Mbit/s weight total delay of throughput delay of throughput delay of throughput class class P class P throughput class of class class P of class P class P of class P /5 /5 / /9 3/9 / / /3 / / /4 / /7 /7 / /3 /3 / Then, we investigate the throughput of the network illustrated in Figs. 4, where, the x-axis is the normailized load of an individual and the y-axis represents the throughput of a single class (all s). Fig. 4 (a) shows the throughput of. When the offered load is below.6 (or 6 Mbit/s), the demands of all can be satisfied, hence, the throughput of is equal to its traffic load. If the load of an exceeds.6, the overall bandwidth demand is larger than the available bandwidth of the network. The bandwidth cannot meet all demands and some of them are refused by the OLT, thus, some packets are delayed in the buffer. Simultaneously, more and more packets arrive at the buffer until the buffer is full and has to drop packets. Since the predefined BW threshold prevents the traffic of from engrossing too much bandwidth, the throughput of this class is confined. However, the two-stage queue scheme permits the traffic of to be transmitted as much as possible, and the arriving packets of can preempt lower priority packets queued in the buffer. Therefore, the throughput of under the two-stage queue scheme is close to its traffic load. Fig. 4 (b) presents the throughput of P. If the offered load is lower than.6, the bandwidth is sufficient to accommodate demands of all s. Hence, the throughput of P is close to its traffic load. Once the offered load increases beyond.6, some packets of P may be dropped and its throughput decreases. evertheless, BW threshold defined by can guarantee a minimum throughput for P, about 8 Mbit/s. With the two-stage queue scheme, when packets of arrive continuously and all queued packets of P are pushed out to place newly arriving packets with higher priority, packets of P will be discarded and its throughput may continue falling. As shown in Fig. 4 (c), even though the offered load reaches the ultimate value, a minimum throughput (about Mbit/s) of P is ensured by the BW threshold. Compared with, the two-stage queue scheme has to drop packets of P to store arriving packets with higher priority, thus, the throughput of P descends to a low point when the load is very heavy. Table lists the load and the throughput of each when the overall load equals Mbit/s. In this scenario, BW threshold i is Mbit/s, Mbit/s and Mbit/s for three classes, respectively. Following the max-min fairness policy, for, the bandwidth demands of i (i=,,...,8) can be satisfied, therefore, it is found that throughput of i (i =,,..., 8) is close to its load. The demand of i (i = 9,,..., 6) cannot be satisfied and these s share the remaining bandwidth evenly, thus, their respective throughputs are almost equal. Since the throughput of an is determined by the amount of bandwidth allocated to this, the simulation result shows that all s fairly share the bandwidth within, and the same for P and P. ow we discuss the packet drop rate. Since the newly arriving packets with higher priority can preempt lower priority packets Table. Throughput of each vs offered load equal to Mbit/s class (Mbit/s) class P (Mbit/s) class P (Mbit/s) ID load throughput load throughput load throughput only when the length of lower priority queue exceeds a threshold, the packet drop rate of may be non-zero. The reason for setting queue-length threshold is explained in section.. There are three curves in Fig. 5 (a), because the buffer drops some packets of under the offered load exceeding.9 (9Mbit/s). Referring Fig. 4 (b) and Fig. 4 (c), when the offered load exceeds.9, the throughput of P and that of P (with using ) keep around Mbit/s and 8 Mbit/s, respectively. Hence, to guarantee the minimum throughput of P and P, if the length of both queue and queue are under their respective thresholds, some newly arriving packets of are discarded. Comparatively, in Fig. 5 (b), the packet drop rate of is zero and that of P is very small. However, the packet drop rate of P increases sharply as the offered load beyond.5. can moderate drop rate of P at the expense of dropping some packets from and P. Finally, Fig. 5 (c) & (d) shows the average packet delay for three service classes. decreases the average packet delays of and P. The reason is that the packets only wait in one queue and the queuing delay is less than that with the two-stage queue scheme (Fig. ). also can transmit more packets from P instead of discarding them. However, these packets delay longer in the buffer until the correct time slot arrives, hence, their average queuing delay is prolonged. The simulation results demonstrate that even if the offered load is very heavy, can guarantee a minimum bandwidth

5 5 P P (a) Packet Drop Rate of 5 P P (b) Packet Drop Rate of Two Stage Queue Packet Drop Rate (Mbit/s) Packet Drop Rate (Mbit/s) P P (c) Average Packet Delay of P P (d) Average Packet Delay of Two Stage Queue Average Packet Delay (s) Average Packet Delay (s) Fig. 5. Packet Drop Rate and Average Packet Delay for each service class due to the bandwidth threshold, which also influences the throughput of the single class. The proposal can also reduce the average queuing delay of high priority and medium priority classes. 4. COCLUSIO This paper first reviewed some typical DBA schemes available in the literature, then proposed a dynamic weight priority scheduling () scheme to efficiently support all s fairly share the uplink bandwidth on demand. ensures that all service classes proportionally share the bandwidth according to their weights. Within the same class, all s are allocated the bandwidth with equal rights regardless of their demands. Because high priority class is served prior to the others, its average queuing delay is shortened. Since the OLT ensures a minimum bandwidth to each class under heavy load, a minimum throughput of each class can be guaranteed. The weight can be adjusted to change the proportion of bandwidth shared by each service class. Thus, if the OLT wants to increase the throughput of one class, it should adjust the weight of this class properly. Our future work is to shorten the prolonged interval between two adjacent frames because the OLT has to wait for REPORT messages from all s. Another one is to decrease the bandwidth consumed by the gap times. 5. REFERECES [] Media Access Control Parameters, Physical Layers and Management Parameters for Subscriber Access etworks, IEEE Draft P8.3ah/D.TM, Aug.. [] G. Kramer, Interleaved Polling with Adaptive Cycle Time (IPACT): A dynamic bandwidth distribution dcheme in an dptical dccess detwork, Photonic etwork Commun., vol. 4, no. pp. 89-7, January. [3] G. Kramer, B. Mukherjee, S. Dixit and Y.H. Ye, Supporting differentiated classes of service in Ethernet passive optical networks, J. Opt. etworks, pp.8-98,. [4] F.T An, H. Bae, Y.L. Hsueh, K.S. kim, M.S. Rogge, and L.G. Kazovsky. A new media access control protocol guaranteeing fairness among users in Ethernet-based passive optical networks, IEEE OFC, Atlanta, GA, Oct.3. [5] D. Bertsekas and R. Gallager, Data etworks, Prentice Hall, Upper Saddle River, J, nd edition, 99.