Abstract
Optical burst switching (OBS) is the switching technology that enables optical internet with current technology. OBS networks are capable of operating with or without optical buffers and wavelength converters. Hence, medium access control (MAC) protocol plays an important role in resolving contention in such networks. Distributed Queue OBS Ring (DQOBSR) is one MAC protocol proposed for OBS ring networks. In this paper, the quality of service (QoS) provisioning scheme for DQOBSR protocol is presented. For the proposed protocol, service differentiation for different classes of traffic in terms of throughput and access delay is verified through simulation. A mathematical model is also presented for the computation of classwise mean access delay. The results obtained using the model are found to be in good agreement with the simulation results.
1. Introduction
OBS is seen as the prospective solution for optical internet that enables efficient utilization of the fiber, withstanding the current technological limitations [1]. In OBS, contention resolution is one of the most important issues, since the network is provided with limited or no optical buffers. An effective method of addressing the contention is selecting an appropriate MAC protocol. A new MAC protocol, namely, distributed access OBS (DAOBS) is proposed for OBS rings in [2], and later it had been extended for mesh topologies [3]. This protocol is an adaptation of distributed queue dual bus (DQDB) to OBS ring networks. For the case of ring topology, DQOBSR protocol [4] is proposed with simpler bus assignment without compromising the performance. In this paper, the QoS provisioning for DQOBSR-based networks is presented, and the performance is analyzed through simulation and mathematical modeling.
2. Overview of DQOBSR Protocol
The topology for the DQOBSR protocol consists of two unidirectional counter rotating rings similar to fiber distributed data interface (FDDI). The capacity of each ring is split into data channels and two control channels. The schematic diagram of the network implementing the DQOBSR MAC protocol is shown in Figure 1. Each wavelength channel is divided into equal length slots. The control slots are further split into number of minislots, and there is a one-to-one correspondence between the minislots and the number of data channels. The control slot format is shown in Table 1.
The nodes implement a firsti-first-out (FIFO) distributed queue similar to the one used with DQDB protocol using the two counters, namely, the request counter (RC) and the countdown counter (CDC). However there are two major differences: single node acts as headend node for both the rings in DQOBSR protocol and the bus is wavelength multiplexed. The multiplexed bus necessitates an entirely different control scheme and slot format. The function of the headend is to generate empty slots for the control channels, that is, slots with starting and ending delimiters. With this, clock recovery can be done all optically using simple optical processing devices [5]. The nodes are equipped with a tunable transmitter and tunable receiver for sending and receiving the bursts.
Before selecting the slot for transmission, the node ensures that no other burst is destined for the same node, as the destination of the burst to be transmitted is in the same slot. Then the burst is transmitted using the corresponding wavelength based on the position of the minislot after an offset time similar to conventional OBS [1]. However, when there are no requests from the downstream nodes, then any node can make use of the passing free slot without making a reservation.
3. QoS Provisioning of DQOBSR Protocol
Since the protocol is based on distributed queuing, QoS provisioning can be implemented through priority queuing. In this priority scheme, at any point in time, a reservation request is sent for the lower priority bursts only if the higher priority queue is empty at the nodes. Moreover, an upstream node receiving a request for a lower priority burst will consider the request and, allow a free slot only if it does not have any higher priority bursts to transmit at that time. However, within a particular priority class the service is according to FIFO discipline.
When there are priority classes, each node will consist of number of RC and CDC pairs, each and pair associated with the corresponding queue for the th priority level. At a particular priority level, a node can be in IDLE or BUSY state. A node is said to be in IDLE state when there are no bursts waiting for transmission. In this state, the node will be updating the based on the requests from downstream nodes and higher priority self-burst arrivals. The node is in BUSY state when it has a burst to transmit at that priority level. In this state, after placing a request, it will update the based on the requests from downstream nodes and higher priority self-burst arrivals. A higher priority burst is given priority at both local as well as at distributed queues by the following actions.
A node in IDLE state at the th priority level monitors the requests from downstream nodes as well as bursts arriving to the node itself, with a priority level equal to or higher than . For each higher priority request or burst arrival, the node increments the request counter, . Similarly, a node in BUSY state increments the countdown counter, for every high-priority self-burst arrival and every high priority request arrival.
4. Mathematical Model
Let the delay incurred in the local queue be and let the delay incurred in the distributed queue be . Then the total queuing delay, can be expressed as The total number of priority classes is . The per slot arrivals can be characterized by the joint probability generating function as follows:
Similarly, the common probability mass function of identical and independently distributed service times can be represented by , and the mean service time of a class-j burst is given by . Then the load offered by class-j bursts is given by , and the total load is given by . Class-i bursts are assumed to have nonpreemptive priority over class-j, if . Using the approach in [6] for the head of line (HoL) priority scheme, the mean delay for class-j bursts in the distributed queue can be expressed as where The Poisson arrivals and deterministic service times are chosen for calculating the delay for DQOBSR-based network.
The priority-based local queuing delay, , can be written as where = holding time in the local queue for class- bursts, and = offset time, = transmission time (1 slot).
Then the total access delay can be obtained by substituting (3) and (5) in (1).
5. Simulation Results
The performance parameters analyzed are mean access delay, blocking probability, and throughput. The bursts are generated according to Poisson process. The burst size is fixed to be one slot size which is set to 0.1 msec. The capacity of each channel is taken as 10 Gbps. The processing time is assumed to be 10 μs. The length of each link is assumed to be 50 Kms. Three service classes are used for the bursts, class-0, class-1, and class-2. The simulation results presented for a 16 node network with a typical traffic proportion of 0.1 (class-0), 0.2 (class-1), and 0.7 (class-2). The class-0 and class-1 refer to the delay sensitive higher priority classes, and class-2 corresponds to the best-effort traffic class. Class-0 corresponds to traffic with very strict delay constraint like interactive audio/video and class-1 corresponds to relatively tolerant but delay sensitive traffic such as audio/video streaming. The discrete event simulator is written in Java and run using the Eclipse IDE.
5.1. Throughput
The normalized throughput is determined by taking the ratio between the total number of bursts transmitted and the total number of bursts generated. Figure 2 shows the throughput performance of different classes. Throughput decreases with the increase in load irrespective of the service class. At loads less than 0.5 Erlangs, there is no significant difference in throughput between the classes. However, for higher loads of more than 0.5 Erlangs, the service difference between classes is observed. There is a difference of more than 3% and 18% observed in the throughput of class-1 and class-2 bursts, respectively, with respect to class-0 bursts at 1 Erlangs. Because of the prioritized scheduling at the local and distributed queues, the probability of a class-0 burst getting transmitted is more compared to the lower priority bursts.
5.2. Mean Access Delay
Mean access delay is the interval between the instant a burst arrives at a node to the instant it gets a free slot and starts transmission. The mean access delay performance for the three classes obtained through simulation and model is compared in Figure 3. The delay of class-0 and class-1 bursts is less by almost 83% and 80%, respectively, when compared with class-2 bursts.
The delay values obtained from the model is found to be in good agreement with the values obtained using simulation. Mean value of the error is found to be 5.1% of the mean delay value. The deviation may be attributed to the single server model approximation of the wavelength multiplexed ring of DQOBSR-based network.
If the relative proportion of the higher priority traffic is increased, the access delay will also get increased. The mean access delay for the three different classes is plotted for two different traffic combinations in Figure 4. Case a refers to the 0.1/0.2/0.7 (class-0/class-1/class-2) distribution of traffic chosen for all other plots, and case b refers to relative proportion of 0.2/0.3/0.5 traffic. There is an increase in delay observed for all the traffic classes in case b with respect to the corresponding delay in case a. In case of class-1, there is an increase of 23.7% of delay with respect to case a. This is due to the relative increase in the higher priority class-0 and increase in its own class-1 traffic as well. The mean access delay of class-0 traffic is also increased by 17.7% due to the increase in its own class-0 traffic from 0.1 to 0.2%. For the case of lowest priority class-2 traffic, the increase in delay is 12.8%. There is an increase of its own priority level traffic from 50% to 70%, which should result in increased delay for class-2 traffic for case b. However, the reduction of 20% in the higher priority traffic has greater impact in the access delay of the lowest priority class leading to reduced delay. Hence, the relative increase in delay for class-2 is between the class-0 and class-1 traffic classes.
From Figures 2, 3, and 4, it can be observed that effective QoS provisioning can be implemented with the proposed priority-queuing-based scheme for DQOBSR networks.
5.3. Blocking Probability
Blocking probability is determined by taking the ratio between number of bursts dropped due to unavailable buffer space and total number of bursts generated. With sufficient electronic buffer capacity at the nodes, it is possible to achieve zero blocking at the expense of increased access delay. The blocking probability performance for the three classes is shown in Figure 5. For buffer capacities of above 5000 bursts, no blocking is observed. For a buffer size of 1000 bursts, no blocking occurs for the higher priority classes but blocking is observed for class-2 bursts. For a reduced buffer size of 100, blocking occurs irrespective of the class of service.
It can also be observed that blocking is present for the class-2 bursts even at lower loads (<0.5 Erlangs), whereas for class-0 bursts, blocking begins to occur at relatively larger loads (>0.7 Erlangs) even at small buffer sizes. This shows the effectiveness of the QoS provisioning scheme.
6. Conclusion
In this paper the QoS provisioning through priority assignment is presented and analyzed for the distributed-queue-based DQOBSR protocol. From the simulation results, it is observed that the proposed priority assignment in the local and distributed queues is able to provide very good service differentiation among classes. A mathematical model is also presented to compute the classwise access delay and the predicted values are found to be in good agreement with the simulation results. Though the scheme is shown for only three service classes, the concept can be extended for more number of classes, if required.