Efficient Control Plane for Passive Optical Burst Switching Network

Transcription

1 Efficient Control Plane for Passive Optical Burst Switching Network Ahmed Triki 1, Paulette Gavignet 1, Bernard Arzur 1, Esther Le Rouzic 1, Annie Gravey. 2 1 France Telecom Orange Labs, Lannion, FRANCE 2 Institut Telecom, Brest, FRANCE Abstract One of the main drawbacks of classical Optical Burst Switching (OBS) solutions is the loss of bursts when contentions occur. Time-domain Wavelength Interleaved Networking (TWIN) is a lossless solution with contention resolution in the edge nodes providing a simple and passive switching in the core nodes and satisfying both low-energy and efficient bandwidth use criteria. However, this solution requires an efficient control plane. In this paper, we compare three different control planes based on either centralized or distributed schemes. Moreover we use two different slot allocation strategies (contiguous or disjoint). The performances of the proposed solutions are compared in terms of total delay, jitter, queue length and bandwidth utilization. The simulation parameters are carefully chosen to take into account implementation constraints. We find that the centralized solution with contiguous slot allocation is the most efficient and it allows a throughput up to 7 Gb/s. T I. INTRODUCTION HE TRAFFIC volume of broadband telecom networks is increasing rapidly and so is its energy consumption; the main driver of this increase being hungry data traffic coming from the generalization of video and internet usage both in wireline and wireless networks. At the same time, green considerations put pressure on telecom actors to find technological solutions being able to cope with these two antagonistic trends. Optical sub-wavelength switching technologies track these two goals: better use of bandwidth by reducing the transport granularity under wavelength scale combined with reduction of energy consumption by removing O-E-O (Optical-Electrical-Optical) conversions. In this sense, Optical Burst Switching (OBS) can be an attractive candidate. OBS compensates for the lack of flexibility of Optical Circuit Switching (OCS) and the technical complexity of Optical Packet Switching (OPS). Moreover, OBS aims at reducing the O-E-O conversions realizing switching functions in the optical domain thus taking a step forward on the way to all-optical networks. Nevertheless, since intermediate nodes do not perform O-E-O conversion and optical memories do not exist, a robust control plane becomes mandatory to manage efficiently potential optical bursts contentions. Since more than ten years, many OBS solutions have been presented either theoretically or experimentally; one of the main drawbacks of classical OBS ones being the loss of bursts when contentions occur. In this rich state of the art, some lossless solutions have been proposed based on the idea of providing a simple and passive switching at the intermediate OBS nodes. Especially, in [1-2] authors propose an all-optical burst switching solution called Time-domain Wavelength Interleaved Networking (TWIN) in which a given wavelength is dedicated to transport the traffic from source nodes to a single destination node. The potential burst collisions are solved at edge nodes. This interesting feature fulfils the low energy all-optical switching requirements discussed above. Multi-point to point tree in TWIN solution seems to be similar to Passive Optical Networks (PON) used in access networks, but differences exist mainly on the fact that in PON information concerning data plane and control plane are transmitted on the same wavelength. Additionally, each source node in TWIN sends data to many destinations which leads to additional constraints concerning resource management. For this reason, an efficient control plane for TWIN solution is required to manage the sending of bursts between each sourcedestination pair. In this context, two main control schemes might be defined: centralized scheme [1] [3] and distributed scheme [4-5]. In the centralized scheme, the resource allocation is done from a centralized control entity that has access to the complete network state, including propagation delays and requests from sources. In the distributed scheme, the control is shared between several nodes. [1] and [3-5] present various control schemes that will be detailed in section II. In this paper, we propose new mechanisms for the control plane in centralized and distributed schemes. Especially we introduce the concept of separated control and data cycles in the TWIN architecture. We compare, by simulation, the networking performance of the proposed control schemes in terms of bursts end-to-end delay, jitter, queue length and total bandwidth utilization. The target of this comparison is to define the control scheme that better suits with experimental implementation. Therefore, experimental considerations are taken into account to determine the simulation parameters. Moreover, as we aim to study metropolitan networks, distances between nodes in simulation scenarios are taken in the range of several hundreds of kilometres. The next section of the paper presents the TWIN concepts. Section III describes the control schemes and their related functionalities. Section IV presents the simulation model and the obtained results. II. TWIN CONCEPT Time-domain Wavelength Interleaved Networking (TWIN) [1-2] is a cost-effective network architecture that provides flexible connectivity using passive optics in core nodes. Indeed, a particular wavelength is attributed to each egress node to receive its data. When a source needs to send a burst to a given destination, the source tunes, during the emission of the burst, its laser to the wavelength uniquely assigned to that destination. Each intermediate node switches passively optical

2 bursts from its inputs to its outputs according to the color of the burst. Thus, the physical topology of TWIN can be viewed as overlaid optical multipoint-to-point trees. Each of these trees has a unique color and it is associated to a unique destination. To perform resource management and signaling, TWIN adopts an out of band control plane by using a dedicated channel for this purpose. Fig.1 shows an example of TWIN architecture. In this example, nodes S1 and S2 use the wavelength represented by the dotted lines to reach node D1. Nodes S2 and S3 use the wavelength represented by solid lines to get to node D2 and node S3 uses wavelength displayed by dashed lines to reach node D2. trip time. Moreover, in this study we propose to use contiguous slots and we compare the schemes in terms of delay, jitter, service time, queue length and resource utilization. III. CONTROL PLANE DESCRIPTION The main task of the control plane in TWIN architecture is to manage source transmissions in order to avoid bursts collisions in the core and destination nodes. In this paper, we propose three schemes for TWIN control plane: two schemes based on centralized approach and the third scheme is based on a distributed approach. These schemes are designed to accomplish the same set of functionalities: signaling, traffic estimation, resource allocation and slot assignment. As depicted in Fig.2, traffic estimation and slot assignment are implemented in the source side, whereas resource allocation is implemented in the control entity side. In the distributed scheme, control entities are located in each destination node, while, in the centralized schemes, the control entity is a unique and a particular node of the network. The functionalities composing the control plane will be described in more details in the following sections. Fig.1: Example of TWIN architecture According to this architecture, the complex processing functions are pushed away to the network edge (inside the edge nodes) such that the network core has only to deal with optical transmission. Edge source node performs burst assembly process, then, it inserts assembled bursts into queues until they are allowed to be sent. Source nodes utilize fast tunable lasers and destination nodes use burst-mode receivers. Whereas, intermediate nodes consist of a passive wavelength selective cross-connect (WSXC) capable of merging and routing incoming colored bursts to the appropriate outgoing ports. The cross-connect configuration is kept unchanged for long time scales. The fact that all sources share the same wavelength channel to reach a specific destination leads to a risk of collisions at each merging point of the tree. To resolve this problem, TWIN relies on a control plane allowing the coordination of sources transmission. The control schemes presented in [1] and [3] are centralized ones which mean that the resource allocation is done by a centralized control entity that has access to the complete network state, including propagation delays and requests from sources. The control schemes presented in [4] and [5] are distributed ones, so the control entity is shared between several nodes. The proposed scheme in [4] is source-based but it can lead to bursts loss. The proposed scheme in [5] is destination-based. It needs several grants re-sending in case of conflicts in a source for the emission of a burst for various destinations. Our work considers a centralized scheme with two resource allocation solutions and a distributed scheme. Compared to previous works, our schemes introduce an additional cycle (data cycle) for the resource allocation in order to ensure the independence of the slot allocation mechanism from the round Fig.2: Control plane functionalities In our proposition, burst reception is organized by repetitive cycles that we call control cycle in order to ease the resource assignment. The duration of a control cycle is common to all destinations and it exceeds the duration of the round-trip time of the most distant source-destination pair nodes in the network. As depicted in Fig.3, a control cycle consists of a predetermined number of data cycles (we refer to this number by α). Each data cycle is divided on a predetermined number of slots (we refer to this number by β). Each time slot can carry only one burst. Adjacent bursts are inter-spaced by a guard time to take into account implementation factors such as time-of-day synchronization errors.

3 to reach a given destination are arranged side-by-side if possible. We refer to this strategy by contiguous allocation. The second one does not consider this constraint and we call it disjoint allocation. Fig.5: Contiguous and disjoint slot allocation Fig.3: Time repartition in TWIN architecture 1) Signaling The source nodes make an estimation of their traffic and send the number of required slots to the Control Entity (CE) via a request message. By taking into account requests coming from source nodes, the CE calculates the resources that can be allocated to each source in order to allow bursts transmission without collisions. The CE attributes slots to sources and sends its indexes via a grant message. Thus, each source receives a grant notifying the allocated slots for the data cycles of the next control cycle. The slot allocation pattern is the same for all data cycles belonging to the same control cycle. Fig.4 illustrates this mechanism in the case of two sources. Fig.4: Signaling mechanism 2) Traffic estimation As the grant received from the Control Entity concerns resource allocation for the next control cycle, the request should estimate needs of source for resources during each data cycle belonging to this control cycle. The number of needed slots is calculated as a function of the queue size and the received bursts in the previous control cycle. 3) Resource allocation The resource allocation consists in reserving slots of time for a given source to send its burst for a given destination. We choose to study two different resource allocation strategies as shown in Fig.5. In the first one, slots attributed to each source These two resource allocation strategies may be implemented in both centralized and distributed schemes. In the centralized scheme as well as in the distributed scheme, the CE takes benefits from the tree structure of the logical topology. In fact, bursts that are timed not to collide at the destination cannot collide anywhere else in the network. This characteristic of TWIN alleviates the complexity of the resource allocation functionality. a) Distributed scheme In the distributed scheme, a CE is located in each destination node. It manages burst transmission for only sources having traffic to send to this destination. As control entities run their resource allocation algorithm independently, a source may receive multiple grants asking it to transmit simultaneously toward multiple destinations. In the case where the number of transmitters in the node is smaller than the number of simultaneous grants, a conflict called slot blocking happens. Thus, slot blocking in the case of contiguous allocation can lead to loosing a block of slot reservation, then, a considerable under-utilization of resources occurs. Therefore, we exclude this case afterwards. In the distributed scheme, the CE in a destination node j computes the proportion of resources that it will allocate to a source i by the following manner: represents the number of required slots by the source i to transmit traffic to the destination j, s is the number of sources related to the destination node j and β represents the number of slots per data cycle. Besides, the CE generates randomly a slot allocation configuration (e.g. Fig 5). b) Centralized scheme In the centralized scheme, the CE firstly determines the number of slots to allocate to each source-destination pair (i,j) as follows: min, " (2) Where, (1) and represents the number of required slots by the source i to transmit traffic to the destination j, s is the number of sources

4 related to the destination node j, d is the number of destinations related to the source node i and β represents the number of slots per data cycle. In this way, the CE ensures a fair resource repartition between couple source-destination. Then, the CE performs the allocation of slots by treating source-destination pairs successively. To allocate a slot m (1 ) to a given source-destination pair (i,j), the CE verifies firstly that the number of reservations for this slot doesn t exceed the number of receivers and secondly that the source has a free transmitter during the moment of use of the slot ( ). is calculated as follows:. (3) where represents the start time of the data cycle, is the duration of one slot and is the propagation delay between source i and destination j. In the contiguous allocation, CE accomplishes this step by taking neighboring slots, while, in the disjoint allocation, slots are taken randomly. The main concern in the distributed scheme is to avoid the bursts collisions in the destination receiver. Thus, the time complexity of the allocation resource algorithm in this scheme is estimated to be O(s). It depends linearly on s (the number of node sources related to one destination). In the centralized scheme, the control entity deals with one more issue: the avoidance of bursts collisions in the destination nodes and the avoidance of slot blocking in the source nodes. Thus, additional iterations are required. The time complexity of the algorithm is polynomial and it is estimated to be O( ) where s is the number of source nodes and d is the number of destination nodes. 4) Slot assignment This functionality is performed in the source side. It consists in attributing bursts to slots. In the distributed scheme, the source has to manage the various grants (coming from various destinations) and chooses the adequate grant to use in the case of slot blocking. Thus, the slot assignment in the distributed scheme pattern may differ from one data cycle to another in contrast to the centralized scheme where this process is reduced to a repetitive attribution during an entire control cycle. IV. SIMULATION RESULTS AND DISCUSSION We compare the performance of the three proposed schemes using a simulator based on OMNET++ software. In order to guarantee the reliability of results, we verify that the confidence intervals are sufficiently small with regard to the system model of this study. Thus, we perform 50 runs for each simulation with same parameters but different random number seeds. We consider a metropolitan network topology composed of four source nodes and four destination nodes with non-equal propagation times between source-destination pairs (Table 1). Each source node sends traffic to all destination nodes. However, bursts considered in evaluating performance belong to only one source-destination flow. We verify that performances are still similar for the other flows. We assume that arrival of the bursts from the burst assembly module follows a Poisson process. This assumption is well justified in [6-7]. Bursts are not differentiated with respect to class of service and are supposed to be completely filled. The capacity of Tx and Rx is set to 10 Gbit/s. The time slots have a fixed duration equal to the duration of a burst plus a guard time equal to 500 ns, in order to take into account laser tuning time [8] and synchronization accuracy issues. Table 2 summarizes the parameters of the simulator. Table 1: Source-destination distance (km) Dest. 1 Dest. 2 Dest. 3 Dest. 4 Source Source Source Source Table 2: Simulation parameters Parameters Values Capacity of Tx/Rx 10 Gbit/s Number of Tx/Rx per node 1 Time slot 5 µs Guard time 0.5 µs Size of burst 5600 bytes Number of slots per data cycle 100 Data cycle duration 500 µs Control cycle duration 10 ms We focus in this comparative study on four main performance parameters: delay, jitter, queue length and total resource utilization. These parameters are evaluated as a function of the offered load. Hereafter, we mean by offered load, the ratio between the rate of data entering an ingress node and destined to an egress node and the channel capacity between both nodes. Fig. 6 represents the end-to-end delay seen by the bursts as a function of offered load for the three schemes. The end-to-end delay includes the waiting time, the service time, the transmission time and the propagation delay between source and destination. The waiting time corresponds to the time the burst spends from the entering to the queue until it reaches head of the queue; the service time is the time spent by the burst in the head of the queue waiting for an available slot. The transmission time is the time taken by the transmitter to completely release the burst from the node (it is equal to 5µs). The propagation time between the two studied nodes is equal to 1065 µs (corresponding to 213 km).

5 Fig.6: End-to-end delay versus offered load We can observe that below an offered load of 0.6 the centralized scheme with contiguous allocation achieves the lowest delay while the distributed scheme performs a delay slightly longer than other schemes. This behavior is probably due to the presence of slot blocking in the distributed scheme which disturbs the burst assignment process. Beyond 0.6 load, the delay for the centralized scheme with disjoint allocation and the distributed scheme increases abruptly (from 4 ms at 0.6 to more than 20 ms at 0.7). The delay in the centralized scheme with contiguous resource allocation increases slowly until a load equal to 0.7 (6.5 ms), besides, it undergoes a sudden rise (38 ms at 0.8). Fig.8: Service time versus offered load To examine the delay more closely, we depict in Fig. 8 the service time versus the offered load. The service time in the centralized scheme with contiguous allocation is higher at low load than at high load: it decreases from 60 µs at a load equal to 0.1 to 30 µs at a load equal to 0.6. Reversely the service time values in the two other schemes, which are both based on a disjoint allocation, increase from 20 µs at 0.1 up to 30 µs at 0.6 load. This can be explained by the fact that, in the contiguous allocation when a burst arrives to the head of the queue after the end of the block of slots reserved to its transmission, it has to wait for the next data cycle. The higher the load is, the fewer sources loose opportunities to insert bursts. However, in the disjoint case, the arriving burst has more chance to find an available slot in the current data cycle. So, it will have to wait for less time in the head of the queue. In the distributed scheme, the source is still unable to benefit from all its opportunities since it suffers from slot blocking. This becomes more visible at high load (a load superior to 0.5). According to the previous results, the delay is mainly dominated by the waiting time in the queue. Fig.7: Jitter versus offered load Fig. 7 shows the jitter versus offered load. To calculate this parameter, which represents the variability over time of the latency across the network, we take the difference between the 99 th percentile and the 1 st percentile of the delay distribution. The jitter curve presents almost the same behavior as the delay curve. For a load between 0.1 and 0.4, the three schemes present a low jitter (1 ms). Then, between 0.4 and 0.6, the jitter increases up to almost 10 ms. Beyond a load of 0.6, the distributed scheme and the centralized scheme with disjoint allocation become unstable (jitter value > 40 ms). The centralized scheme with contiguous allocation shows a steady value of about 10 ms until a load of 0.7. Fig.9: Queue length versus offered load Fig. 9 shows the length of the queue as a function of load. In the centralized scheme with contiguous allocation, beyond the load of 0.7, the system becomes unstable and queue size would continue to increase infinitely. For the two other schemes, the system stability threshold is reached earlier at a load of 0.6. In the three schemes, the average of queue length is almost the same for the range of traffic load within which the network is stable. The queue length value explains the

6 significant rise of delay at 0.5 load. In fact, by multiplying the average service time at 0.5 (30 µs) by the length of queue at this load (almost 65 bursts), we obtain a waiting time of about 2 ms, which is consistent with the recorded delay value. Fig. 10 shows the resource utilization (average proportion of used slots during a data cycle) in function of the offered load. The centralized scheme with contiguous allocation presents greater efficient resource utilization than the two other schemes. It can reach a resource utilization ratio of about 80% which enable the emission of more than 7 Gbit/s of traffic. The percentage of resource utilization in the distributed case is limited to 66%. Compared to the distributed scheme, the centralized scheme with contiguous allocation allows the utilization of almost 20% of additional resources among the available ones. Referring to the previous results, we conclude that the performance and the stability of the system are related to the ability of control scheme to manage the available resource. Indeed, at high offered load traffic, the average delay and queue length increases significantly when the CE is unable to satisfy requests of its related sources. further improvement of the centralized scheme performances. In fact, using resource reservation based on timestamp avoids the attribution of guard times between consecutive bursts. Thus, it reduces bandwidth waste. However, the centralized approach still needs significant computational capabilities and high protection requirements. REFERENCES [1] Widjaja, I.; Saniee, I.; Giles, R. & Mitra, D., Light core and intelligent edge for a flexible, thin-layered, and cost-effective optical transport network, Communications Magazine, IEEE, 2003, 41, S30-S36 [2] Saniee, I. & Widjaja, I., A new optical network architecture that exploits joint time and wavelength interleaving, Optical Fiber Communication Conference, 2004, TuH4 [3] Ross, K.; Bambos, N.; Kumaran, K.; Saniee, I. & Widjaja, I., Scheduling bursts in time-domain wavelength interleaved networks, Selected Areas in Communications, IEEE Journal on, IEEE, 2003, 21, [4] Brzezinski, A.; Saniee, I.; Widjaja, I. & Modiano, E., Flow control and congestion management for distributed scheduling of burst transmissions in time-domain wavelength interleaved networks, Optical Fiber Communication Conference, Technical Digest. OFC/NFOEC, 2005, 3, 3, OWC4 [5] Widjaja, I. & Saniee, I., Simplified layering and flexible bandwidth with TWIN, Proceedings of the ACM SIGCOMM workshop on Future directions in network architecture, 2004, {13-20} [6] J. Cao, W. Cleveland, D. Lin, D. Sun, Internet traffic tends toward Poisson and independent as the load increases, C. Holmes, D. Denison, M. Hansen, B. Yu, B. Mallick (Eds.), Nonlinear Estimation and Classification, Springer, New York, 2002 [7] X. Yu, Y. Chen, and C. Qiao, Study of traffic statistics of assembled burst traffic in optical burst switched networks, in Proceeding of Opticomm, 2002, pp [8] Intune Networks, Optical Packet Switch and Transport. A Technical Introduction, 2009, chnical_introduction/ Fig.10: Resource utilization versus offered load V. CONCLUSION TWIN concept is interesting in terms of fast lossless switching and avoidance of optical buffers in the intermediate nodes. It ensures transparency in transit nodes and enables self-routing in the core network because it relies on the wavelength rather than label or address. Nevertheless, the performance of this technology is mainly related to an efficient control plane. In this paper, we compared three control schemes in terms of end-to-end delay, jitter, queue length and wavelength utilization. Simulation results prove that a centralized scheme with contiguous resource allocation allows a throughput up to 7 Gbit/s. Thus, it outperforms centralized scheme with disjoint allocation by almost 12% and the distributed scheme by almost 15%. Even if the source enjoys more flexibility in the slot assignment mechanism in the case of distributed scheme comparing to the centralized one, its performances are still the lowest which contradicts our expectations. Based on these results, we estimate that the substitution of the slotted approach by a timestamp approach might lead to