1

Protection and Spatial Reuse in Optical Burst Transport (OBT) Networks Saurav Das, Jaedon Kim, David Gutierrez, Leonid Kazovsky, Richard Rabbat*, Ching-Fong Su* and Takeo Hamada* Photonics and Networking Research Laboratory, Stanford University, 058 Packard Building, Stanford, California, 94305, USA Email:[email protected] Abstract— Optical Burst Transport network is a novel WDM ring architecture for Metropolitan Area Networks (MAN). It leverages the advantages of optical burst switching and WDM rings, while using a token-based medium access scheme. In this paper, we investigate protection together with spatial reuse property. Specifically we show how protection in OBT can be different from conventional protection and propose three algorithms for implementing a fast and low implementation cost, unidirectional 1:1 protection scheme. We analyzed the performance of these schemes via simulation and found an optimum algorithm for minimizing data loss and maximizing data delivery during the fault. Index Terms—MAN architecture, optical burst switching, optical burst transport, spatial reuse, WDM rings.

I. INTRODUCTION

M

ETRO networks interconnecting high speed backbone and low speed access network, are mostly based on SONET / SDH ring networks. Although SONET / SDH combined with WDM technology has increased transmission capacity, the efficiency issue is always raised, especially with the increasing data traffic. More recently, access networks are providing an ever-increasing amount of bandwidth by employing advanced LAN technologies, such as xDSL, cable modems and passive optical networks (PON). With the advent of user-oriented applications such as real-time video, VoIP, online gaming, and wireless access, metro networks are expected to suffer a lack of bandwidth to deal with access network traffic in the near future. In order to resolve the difficulty in MANs, various approaches have been proposed, which can be broadly classified into one of two categories; Optical packet switching (OPS) and optical burst switching (OBS) [1]. Both schemes aim to develop feasible network architecture to fully utilize the optical transmission link. OBS presents several attractive features. First, an edge node aggregates incoming traffic into a burst and establishes a single hop transmission path from the source to the destination using a control message. Consequently, OBS does not require an optical buffer in the intermediate nodes (unlike OPS), and it contributes to the scalability of the network. Secondly, with such traffic grooming, different kinds of local area traffic can be appropriately combined into higher-speed streams without any frame conversions that are necessary for current MANs [2].

*Fujitsu Labs of America 1240 East Argues Ave., M/S345 Sunnyvale, California, 94085, USA

Thirdly, by separating the control header from the payload and assigning dedicated WDM channels, the payload does not need to be processed at any intermediate node. Therefore, the nodes do not have to suffer from dealing with the transit traffic. Finally, as packet synchronization requirement does not increase with the bit rate of the data channel due to the packet aggregation, the implementation complexity and cost can be lower than that of an optical packet switched network. In a previous paper [3], we demonstrated a token-based optical burst-switching scheme for WDM ring networks named Optical Burst Transport (OBT). In the OBT network, using a token-based scheduled medium access mechanism, we can also guarantee fairness among the network nodes. Although OBT follows the conventional Token Ring medium access control scheme, it differs from the latter in that an OBT node can communicate with other nodes using control messages even during the data transmission, since a separate control channel is always available to every node. Thus, using this property, we can improve the network performance while keeping the benefits of the Token Ring medium access control scheme. Compared with the legacy MAN networks, another property required for the new metro network architectures is a robust protection mechanism. SONET / SDH ring networks are famous for their ability to provide survivability through the use of protection switching. In this paper, we introduce protection schemes for OBT networks. Three possible schemes will be suggested and compared using numerical simulation model. Although all the schemes belong to 1:1 protection, there is significant performance variation between them and the optimum solution for OBT with spatial reuse will be shown. The rest of this paper is organized as follows. Section II gives a brief introduction to the Optical Burst Transport protocol and the specific spatial reuse scheme used to maximize bandwidth usage. Section III reviews conventional protection mechanisms and discusses how protection in OBT differs from them. Section IV delves into the details of the proposed OBT protection schemes and identifies performance metrics. Section V presents simulation results for the various proposed schemes under different conditions and Section VI concludes this paper.

2 II. OPTICAL BURST TRANSPORT (OBT) PROTOCOL A. Optical Burst Transport The network architecture is based on WDM rings. As described in Fig. 1, one control channel and multiple data channels occupy different wavelengths. Signals on the control channel are added or dropped through the WDM MUX and DeMUX, while data are add or dropped by the optical switch. Each network node processes incoming control signals on the control channel and reacts accordingly. Control signals include (1) token (2) control header and (3) other network management messages, which can be used for network protection.

Fig 1. Simplified node and queue structure in the OBT node

Multiple tokens propagating on the ring carry access grants for respective data channels (wavelengths) to realize the Media Access Control (MAC) protocol. When a token arrives at a node, that node holds it and utilizes the corresponding data channel to transmit data packets if it has data to send. In the node, packets waiting for transmission are stored in several virtual output queues (VOQs) associated with their destination nodes on the ring. Discrete packets are aggregated for transmission to reduce encapsulation overhead. At the transmitting node, when a token corresponding to one of the data channels arrives, the node holds it and starts the transmission. A traffic scheduler monitors VOQ lengths and allocates appropriate transmission windows for packets in different VOQs. Before sending out data packets toward a destination node, the node sends a control header (CH) to inform the destination node of the incoming data. Then, after a destination-dependent offset time, the node starts transmitting the data burst to the destination node during the allocated transmission window. That is, the sender transmits the burst in a Tell-and-Go manner. When a VOQ is served, it is temporarily locked to avoid multiple transmissions when the next token for a different wavelength arrives. When the destination node

receives the control header, the destination node throws the switch corresponding to the incoming burst’s data channel wavelength for the specific amount of time, which is specified by the control header from the source that preceded the payload. After dropping the burst, the destination node throws the switch back, to allow the bypass of bursts meant for other destination nodes. B. Spatial reuse in OBT networks Scrutinizing the signal path on the OBT network, when the token is held by a source node for transmission to the destination node, the data path from the source to the destination is occupied, while that from the destination to the source is empty. Therefore, during the transmission from source to destination, a destination-stripping regime creates a collision free area, which can be used by the destination. In the OBT protocol, the source node may create multiple sub-bursts targeting different destination nodes when it holds a token. Every time the source transmits its burst to a destination, the data path between the destination node and the source node can be reused.

Source: Node 4 / Destination: Node 0,1,2,3 Fig 2. Timing diagram of burst transmissions in the OBT network

Fig. 2 illustrates a timing diagram for burst transmission on an OBT network with 5 nodes, where node 4 sends sub bursts to every other node. When node 4 holds a token, it determines the optimal bandwidth allocation according to the size of its VOQs dedicated to different destinations. Node 4 then creates a data burst consisting of several sub-bursts. It sends control headers at destination dependant offset time ahead of each sub-burst. Each control header contains the information of incoming sub burst size so that a destination can properly terminate the sub burst destined to it. During the time when a destination node is receiving the data, it is also granted an opportunity to transmit its own data for a certain duration, which is marked in solid gray color in Fig. 2. That is, whenever a node receives a control header destined to it, the destination node knows the time duration for transmitting as well as receiving its data at the same time. Therefore, destination nodes are allowed to initiate a secondary transmission without a token to make full use of the available capacity. In general, the secondary transmission can be made for a destination set different from the transmission initiated by the token. In Fig. 2,

3

250

w/o spatial reuse w spatial reuse Throughput per node (Mbps)

200

0.8

w/o spatial reuse w spatial reuse

0.7 0.6 0.5

Delay(sec)

node 0 has node 1, 2, 3 and 4 as the destination set, while node 3 only has node 4 as the destination set. The size of the secondary transmission should be less than the reception time of the sub burst because the control header process time would be included in the secondary transmission time slot. Since the secondary transmission from the destination node uses the same timeslot as the primary sub-burst to that node, the size of secondary transmission could have a different value each time. Note that during the secondary transmission, the destination of the secondary transmission may also create another (secondary) transmission to enable higher network throughput and in principle, this process could continue as long as the final destination is not beyond the source of the primary transmission. When a node receives a primary control header, its scheduler assigns the time slot for the secondary traffic to the longest VOQ amongst its valid destination nodes. That is, from Fig. 2, when node 2 receives a control header from node 4, it has to choose the longer VOQ between nodes 3 and 4 even if the longest VOQ may actually be for node 0 or 1. If it chooses node 3, then this node can also send its own secondary transmission in the same time slot but only to node 4. This scheme has the advantage in that it can increase dynamic bandwidth allocation property, since the node can make a decision depending on the current status of its valid VOQs. A WDM ring network consisting of 10 nodes is analyzed, with a circumference of 200 km. Each node has the same fixed burst size of 200 KB and is assumed to be equidistant from its neighboring nodes. The data rate is 1.25 Gbps per wavelength data channel. A single data channel wavelength is used along with the control channel wavelength, but the results of the simulation can be easily extended to multiple data channels or higher bit rates.

0.4 0.3 0.2 0.1 0.0 0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

1.8

2.0

Load (b) Average delay Fig 3. Performance of OBT network with and without spatial reuse

As shown in Fig. 3-(a), the throughput for the OBT network without spatial reuse is close to the theoretical maximum value of 1.25 Gbps divided equally between 10 nodes. The deviation is due to finite token propagation time between adjacent nodes. With spatial reuse the throughput of the OBT node shows a significant improvement. Due to the effectively enlarged channel capacity, the throughput per node is nearly doubled and the average delay performances at the nodes are also significantly improved. In the analysis, the delay measurement consists of the transmission delay, propagation delay and queuing delay of every packet. In Fig. 3-(b), the delay remains low when the traffic load is low, and exhibits a sharp increase when the network is fed by traffic load greater than its affordable throughput. In comparison, however, OBT with spatial reuse shows comparable packet delay at low traffic loads, and a more moderate increase when the traffic load becomes high. III. CONVENTIONAL PROTECTION MECHANISMS AND OBT

150

100

50

0

0.0

0.4

0.8

1.2

Load (a) Throughput per node

1.6

2.0

Several protection mechanisms for metro area and long-haul networks have been developed. Some protection mechanisms operate at the transport layer (SDH/SONET, ATM or IP) while others operate at the physical optical WDM layer underneath [4]. SDH/SONET protection mechanisms have been developed extensively and implemented in order to provide 99.999 % (five nines) availability from the phone network. ATM protection mechanisms were initially developed based on the basic protection mechanisms found in SDH/SONET, but some further developments were made after this. IP networks, which provide a best-effort data service, recover from failures by the slow re-convergence of their routing protocols. This process can take up to several seconds or minutes, a long time compared with the usual 60ms or less recovery time of the SDH/SONET and ATM networks. At the optical layer, several protection mechanisms have been developed as well that aim to

4 provide a continuously available optical layer service to the SDH/SONET, ATM or IP clients using them. These mechanisms have not been deployed extensively due to the slow penetration of dynamic WDM technologies in today’s networks. However, their influence in future networks will be considerable. The OBT architecture described in this paper has a higher integration between the client layer and the physical layer than the integration between SDH/SONET, ATM or IP and the Optical layer. Because of this characteristic and other particularities of OBT, its protection requirements are different and the way to implement it is different as well. In this section, we’ll briefly present the protection schemes that have been developed for SDH/SONET, ATM, IP and Optical layer and then discuss how the particularities of OBT require a different protection mechanism, presented in the following sections of this paper. Protection topologies in general can be classified as 1+1, where data is sent simultaneously along two different paths, and in case of failure of one path, the other path is immediately available; 1:1, where the protection path is available, but not actively used until it’s needed; and 1:N where one protection path is shared between N other active paths. In general 1+1 is more expensive than 1:1, and 1:N is the least expensive. Usually, the term protection is used when traffic is restored in less than hundreds of milliseconds, while the term restoration is used when it takes more time. Protection schemes may be reversible or non-reversible, depending on whether once the damaged link is fixed, the connections revert to it or stay in the protection path until the next failure. A path is one of the multiple logical connections that are going through a particular physical link, while a line usually refers to the physical link itself. A.

Conventional Protection Mechanisms 1) SDH/SONET Protection: The most common forms of SONET protection are Unidirectional Path-Switched Rings (UPSR) and Bidirectional Line-Switched Rings with 2 or 4 fibers (BLSR/2 and BLSR/4). Similar protection mechanisms, with other names, exist in SDH. In UPSR, traffic from node A to node B is transmitted simultaneously in the clockwise (working fiber) and counterclockwise (protection fiber) directions. Node B simultaneously receives the transmissions from both sides and chooses the best one. In this way, even if one of the two paths fails, Node B continuously receives the transmission from Node A. While UPSR works at the path layer, i.e. every node-to-node connection on a fiber is switched separately in case of a failure, BLSR operates at the line layer, i.e. in case of a failure a whole line is switched immediately, making it more efficient. BLSR operation requires more signaling than UPSR and provide span switching and ring switching capabilities, in which the switching needs to be only for a specific part of the end-to-end connection and not for its entirety. Ring switching can protect against node failure as well. BLSR/2 and BLSR/4 differ in the number of protection fibers. While BLSR/2 shares two fibers as working and protection fibers, BLSR/4 has two exclusive fibers for protection.

2) ATM Protection: Several protection schemes were developed for ATM in the last decade [5]. The Automatic Protection Switching (APS) protocols developed for SDH/SONET networks was implemented on ATM’s Virtual Paths (VPs) and Virtual Circuits (VCs) to restore them quickly and efficiently. Other ATM protection mechanisms include Self-Healing Network (SHN), Self-Healing Ring (SHR) and Failure Resistant Virtual Path (FRVP). These mechanisms vary in the amount of time they take to restore service (in FRVP there is no down-time) and the amount of protection resources required. This is a usual tradeoff in protection mechanism. 3) IP Protection: IP networks provide a connection-less best-effort service. Each router on the network looks at the destination IP address of every packet and based on its routing tables sends the packet to a particular neighboring router based on its next-hop information. These routing tables are updated by different routing protocols such as BGP or OSPF, in which each router propagates information about how to reach a particular IP address or subnetwork to its neighbors. When a link failure happens, both the neighbors’ connectivity and message propagation stops, and the routing tables are updated and propagated. This process, by its distributed nature and slow updating, takes some time to converge again; these times might be in the order of seconds or even a few minutes. An option to decrease this response time is to use Multi-Protocol Label Switching (MPLS) with pre-assigned recovery routes; however, this usually requires in many cases the manual setup of these alternative paths on a case-by-case basis. MPLS effectively changes the IP topology of a network by making neighbors out of distant routers. 4) Optical Layer Protection: Given the amount of multiple services that may be run over WDM networks, it is important that the physical optical layer itself provides some level of protection [6, 7, 8]. These optical layer protection schemes are based on the UPSR and BLSR mechanisms of SDH/SONET, and operate at the optical WDM path or line. Notice how in this case a path is a WDM connection, while still is the fiber itself. The physical WDM layer has very limited knowledge of the clients that are running over it (e.g., ATM, SDH or IP) and the protection schemes that they in turn are implementing. Thus, it is important to establish a protocol and fixed sequence of events that will take place in the event of a failure to avoid unstable or race conditions. The interaction of protection schemes between different layers is a very important issue. B.

OBT Protection Given the protection mechanisms that have already been developed for several network architectures and protocols why do we need a new one for OBT? We now discuss the particularities of OBT protection and the issues that are brought up by them. In OBT networks, data is not converted to electronic form at each node. Only the data that is addressed to this node or being sent by this node is in electronic form at it, plus the control messages. Most of the data in the link passes through optically. At a particular link, the data that is lost in case of a failure does not necessarily belong to the two nodes that this physical link connects; i.e., the two end nodes of the physical fiber were not

5 the source or destination of the data. It might belong to other WDM optical point-to-point links between nodes that are farther away. Once the link or path is restored, these remote nodes need to either retransmit the data that was lost or drop it. If the remote nodes are not going to drop the data, then they should either continue retransmitting continuously until they receive an acknowledgement or they should hold the data until they are informed about the restoration of the link. These alternatives and the impact that they have on the overall performance of the network in the event of a failure are investigated in this paper. Again because of the fact that multiple optical WDM point-to-point links belonging to nodes distant from a particular link traverse it at any point in time, once the damaged physical link is restored it would be complicated and very performance degrading to implement a reversible protection scheme. There would be a need to inform all nodes that this is going to happen and to halt transmission while it the reversing switching happens. Thus, the OBT protection protocol should be non-reversible, meaning that even when the damaged physical link is restored, communication will continue on the protection link until it fails. In OBT networks, we have a token that circulates through the ring for each wavelength, authorizing the node that is holding the token to transmit on it. There are multiple variables that the existence of this token brings up to the restoration issue: does the location of the token with respect to the location of the failure affect the amount of data that is lost? When should a token be regenerated and in the event of the failure? Which node should regenerate the token (i.e. should the Master node still be the Master node regardless of where the failure happen)? These are all issues that are not present in previous restoration mechanisms. Last, but not least, the layering paradigm of OBT is quite different that that of SDH/SONET, ATM or IP running over an optical services network. In OBT the burst assembly and transport directly communicates with the physical layer. Thus, we need not implement the sequential layered protection schemes that previous protocols need in order to avoid race conditions. We believe this is a considerable advantage of OBT protection. IV. PROTECTION IN OBT Network requirements for network resilience depend largely on services characteristics. For time-sensitive services such as voice, under 50 ms protection switching times are required to avoid the dropping of telephone connections. Meanwhile, for data or Internet services, relatively longer protection switching times may be acceptable. An industry wide standard for MANs however, is the 50ms protection switching time, which historically has its relation to the time a SONET/SDH payload would have to wait before it can assume a failure has happened. In practice there are several implementation specific factors that effect the actual protection time. Generally, the protection time includes the time it takes to detect the failure, any protection signaling time, the actual protection switching time,

and any additional time for signals to reacquire and stabilize. In practice some implementations with simple 1+1 or 1:1 protection over short distances can switch in a few milliseconds while implementations of more complex shared ring protection schemes on geographically large rings can take upto 50ms. For OBT networks we chose a 1:1 unidirectional protection scheme, where once a failure is detected the nodes co-ordinate with each other and switch over to the dedicated protection path (Fig. 4). We take the 1:1 approach to insure that the time required to enable protection in the ring is minimized while keeping the unidirectional approach to enable a low cost and simple implementation at the nodes. Specifically this approach requires only a pair of 1X2 optical switches at every node and no extra data channel transponder cards. It does however require an extra transponder card at each node for only the control channel.

Fig 4. OBT nodes structure with protection

The main types of failures that this protection scheme works to correct are fiber cuts and transponder failures. Entire node failures are extremely rare and are not considered here. On the other hand, the external fiber facility between nodes is considered the least reliable component of the network and as such, fiber cuts are fairly common. This category also includes any other active or passive components such as optical amplifiers that are deployed along with the fiber. The other kind of failure is the loss of transponders which are one of the more complicated (and thus error prone) parts of a WDM node. Typically 1:N equipment protection can designate spare transmitters to take over in case of failure of the equipment under use. However in OBT networks, due to the token based access of nodes for transmission on specific wavelengths, once a node is aware that its transmitter for a certain wavelength has failed, it can just ignore the token for that specific wavelength and can instead send data on another wavelength for which the corresponding token next becomes available. There is however a need for the node to be able to distinguish between a fiber cut and the loss of a transponder. This is easily accomplished by taking advantage of the unique features of OBT networks. Since the OBT ring uses a dedicated wavelength for the control channel signaling, the node that is immediately downstream from the cut can always detect a fiber cut by monitoring the power lever of the control channel. This is because the control channel is always “on” in between OBT

6 nodes as opposed to the data channels, which are only operational, when a payload is being transmitted. On the receiving end, the loss of this constant stream of control channel bits can be interpreted as a fiber cut. On the other hand, the loss of a data channel transmitter would result in the loss of a data burst but not the loss of the control header (CH) informing the destination node of the incoming data burst. Thus the destination node can signal the loss of data to the source of the CH. The destination node still cannot tell the difference between the loss of the control channel transmitter at the adjacent upstream node and the loss of the control signal due to a fiber cut. Thus in this situation, the destination node must assume that there is a fiber cut and initiate the protection algorithm to switch over to the protection fiber. The adjacent upstream node when informed of the possible fiber cut also initiates the switchover to the protection fiber. At the completion of the switch the upstream node transmits a control header (defined below), which it expects to travel the whole ring and return to it, indicating a successful bypass of the fiber cut. However if the control channel transmitter at the upstream node has failed then the node does not receive the CH within the timeout (equal to the ring latency). At this point the upstream node can re-transmit this CH but use the backup control channel transmitter instead. A single node in the OBT ring network, designated as the Master Node, deals with the loss of a token. Since the OBT protocol uses a fixed burst-length based algorithm for data transmission at any node holding the token, the token holding times (THT) at each node is fixed and is equal to the maximum burst length/ data transmission rate. Thus the maximum time the token can take to return to the Master Node is the total propagation time (ring latency) plus the product of the THT and (N-1) where N is the number of nodes in the network. Another aspect of protection switching in OBT networks is that the switching is non-revertive (as mentioned in the previous section). A part of data is always lost due to the nodes not processing all the data that traverses the node. Thus being non-revertive minimizes data loss. The maximum data lost during a fault is in part a function of the data line rate, the number of data channel wavelengths and the duration of the downtime, which includes the detection, protection switching and signaling times. It may also depend on the location of the token during the fault, the specific token regeneration algorithm used to replace lost tokens, the nature of the secondary transmissions owing to spatial reuse and the load conditions at the time of failure. We now propose three protection algorithms for OBT protection and in the next section we explore, via simulations, the performance of these algorithms under various load conditions. A. Algorithm 1 - Basic Protection: •

Fault Detection - the loss of control channel signal is detected by the node that is downstream from the failure. The loss of a control signal is interpreted as a fiber cut by this node, designated as the Node of Detection (NOD).

•

• • •

•

•

•

•

Protection Signaling - The NOD must inform it’s adjacent upstream neighbor of the failure of the link between them. It does so by sending out a Fault_CH via the unidirectional path of the ring. The adjacent upstream node is designated as the Node of Failure (NOF) Protection Switching - The NOD initiates switching of its incoming signal path to the protection fiber. The NOF receives the Fault_CH and initiates switching of its outgoing signal path to the protection fiber. The NOF, which has become aware of the fault no longer sends out secondary transmission. If it receives a token for primary transmission, it destroys the token. Stabilization Signaling - Upon completion of the switchover at the NOF, it sends out a FaultOK_CH destined to itself along the protection fiber. This serves the dual purpose of verifying that the fault has been bypassed and also serves as a means of distinguishing between a fiber cut and the loss of control channel transmitter at the NOF. If the FaultOK_CH traverses the ring and returns to the NOF, the latter assumes that the fiber cut has been successfully bypassed and it resumes operation as normal. If the FaultOK_CH does not return to the NOF within the timeout (ring latency), the NOF attempts to send out the same signal using its backup control channel transponder. Meanwhile the node designated as the Master Node continues to regenerate tokens when it detects a loss of token via its token timeouts

B. Algorithm 2 - Token Hold-Back: •

• •

The steps in this algorithm are the same as those in the Basic Protection algorithm, with the main difference being in the token regeneration process. Specifically, when the Master node becomes aware of the Fault_CH, while it is in transit through the Master Node on its way to the NOF, it holds back the regeneration of any tokens which timeout. This is done to primarily minimize data losses during the fault duration. Also if the Master Node is the NOD, then upon sending out the Fault_CH, it too holds back the regeneration of any tokens upon their timeout. In the Stabilization signaling phase, if the Master Node receives (if it is the NOF) or processes in transit the FaultOK_CH, it resumes regenerating all the tokens which it had held back previously.

7

• • •

•

• • •

•

•

Fault Detection – This is the same as in Basic Protection. The NOD assumes the role of the Master Node and starts token timeout timers for all data channel tokens. Protection Signaling and Switching – The NOD sends out the Fault_CH and initiates the switch of its incoming path to the protection fiber. The Fault_CH along its path to the NOF informs every node of the existence of the fault as the control channel is processed on a hop-by-hop basis. As nodes become “fault aware”, for any secondary transmission sent out as part of spatial reuse, they avoid sending to those destinations for which the broken link would need to be used. Also when the old Master Node becomes “fault aware” it immediately stops all its token timeout timers and ceases to be the Master Node anymore. The only exception is when the old Master Node is the NOD, in which case it continues to be the Master and does not stop its token timeout timers. Upon timeout of the tokens at the NOD, the latter regenerates intelligent tokens that require nodes to sent primary transmission sub-burst to only those destinations, which will not require the sub-burst to propagate the broken link. Ultimately when the NOF sends out the FaultOK_CH, again every node becomes aware that the fault has been bypassed and they resume sending primary and secondary transmissions to all destinations. The NOD remains as the Master Node.

1.E+06 8.E+05 Data lost (bytes)

•

sections is a little over 12 ms, which is well below the required protection time of 50 ms. We first present results for the Basic Protection scheme and analyze the results in terms of the data lost during the fault and the data successfully delivered in the same time frame.

Load=0.375 Load=0.75 Load=1.125 Load=1.875

6.E+05 4.E+05 2.E+05 0.E+00 0

1

2

3

4

Node of Failure (NOF)

1.E+06

Data delivered (bytes)

C. Algorithm 3 – Intelligent Token Regeneration:

8.E+05 6.E+05

Load=0.375 Load=0.75 Load=1.125 Load=1.875

4.E+05 2.E+05

V. SIMULATION RESULTS In order to simulate the various protection schemes, a numerical model is used. A WDM ring network consisting of 5 nodes is analyzed, with a circumference of 200 km. Each node has the same fixed burst size of 200 KB. The data rate is 1.25 Gbps per wavelength data channel. A single data channel wavelength is used along with the control channel wavelength, but the results of the simulation can be easily extended to multiple data channels or higher bit rates. For traffic arrival at a node in the network, Poisson arrival has been used, but it is easy to show that the results are quite general and would hold true for long-range dependent (LRD) traffic too, where the sessions arrive as a Poisson process and the lengths of the sessions are heavy tailed distributed as in the Pareto distribution. Data packets take the form of Ethernet frames with variable lengths ranging from 64 to 1518 bytes [9]. Additionally, the protection switching time is conservatively assumed to be 10ms and the ring latency for the above defined ring circumference is 1 ms. Thus the total protection time (defined above) for any of the schemes outlined in the previous

0.E+00 0

1

2

3

4

Node of Failure (NOF)

Fig 5. Basic Protection scheme a) data lost b) data delivered

Fig. 5 shows the data lost and delivered during the fault duration under various load conditions. It also shows the dependency of the data lost or delivered on the location of the fault. Specifically, the horizontal axes of the charts signify the location of the fiber cut defined as the link immediately downstream from the Node of Failure (NOF). In all the above simulations the Master Node is defined to be Node 0. It was found that the largest amount of data lost was when the link between node 4 (NOF) and node 0 (Master Node) went down. Alternatively the largest amount of data delivered was also for the same fault location. This can be readily understood from the fact that, with this fault location, all the other nodes use the tokens regenerated by the Master Node, before they are lost.

8 This trend can also be readily seen in Fig. 6, where we have changed the maximum allocated burst size at each node to 100 KB.

1.E+06 8.E+05

Data lost (bytes)

8.E+05 6.E+05

Data lost (bytes)

1.E+06 Load=0.375 Load=0.75 Load=1.125 Load=1.875

tkn @1 TOF=23.9 tkn @2 TOF=25.4 tkn @3 TOF=26.9 tkn @4 TOF=28.4 tkn @0 TOF=29.9

6.E+05 4.E+05 2.E+05

4.E+05

0.E+00 2.E+05

0

1

2

3

4

Node of Failure (NOF) 0.E+00 0

1

2

3

4

Fig 8. Data lost in Basic Protection Scheme for different Time of Failure

Node of Failure (NOF)

Fig 6. Data lost in Basic Protection Scheme for 100 KB Burst size

Thus we find that the data lost is not only a function of the fault location but it depends more heavily on the location of the Master Node. This was verified by changing the Master Node designation to Node 3. As shown in Fig. 7, the amount of data lost is now largest for a failure in the link between Nodes 2 and 3 and as expected the lowest amount of data lost is when the fiber cut is in the link between nodes 3 and 4.

1.E+06

Data lost (bytes)

8.E+05

Load=0.375 Load=0.75 Load=1.125 Load=1.875

6.E+05 4.E+05 2.E+05 0.E+00 0

1

2

3

4

Node of Failure (NOF)

Fig 7. Data lost in Basic Protection Scheme for Master Node 3

Another factor of interest is the location of the token at the time of failure and whether that has an impact on the amount of data lost during the duration of the fault. We can simulate this by executing different times of failure (TOF) and noting the token position at that time and the data lost during the fault. Fig. 8 shows the results of the simulation where we again designated Node 0 as the Master Node and the maximum burst size was 200 KB and the load value was fixed at 12.

We found that the amount of data lost is largely independent of where the token was at the Time of Failure. As explained before the data still shows a trend of higher data loss for a failure at the link just upstream from the Master Node 0, but the value of the data lost is almost same irrespective of which node is transmitting primary sub-bursts (i.e. which node is holding the token) at the time the link goes down. This result gives us the confidence that we can extend our conclusions to the case of multiple data channels. We note that data loss is minimized in the case where the link breaks down just after the Master Node i.e. when the Master Node is the NOF. This is because any regenerated token is also lost immediately. Another way to use this result to minimize data loss is to use the Token Hold-Back protection scheme where the Master Node does not regenerate the token as soon as it becomes aware of the fault, irrespective of where the fault is located. This way the data lost will no longer be a function of the location of the fault with respect to the Master. Clearly, these objectives are met, as verified in Fig. 9. However, we do note that in addition to minimizing the data lost during the fault, the Token Hold-Back scheme also results in very little data actually getting delivered in the same time frame. Essentially when lost tokens are no longer re-generated, none of the nodes can send primary transmissions. This holds true for the secondary transmissions as well which are a part of the spatial reuse algorithm. Essentially, as no primary transmissions occur after tokens are lost and not regenerated, the time slots necessary for secondary transmissions are also not generated. Thus when a single link is down, it ends up bringing the whole ring down for a part of the fault duration. This brings us to the third proposed protection algorithm labeled Intelligent Token Regeneration. The idea behind this scheme is to minimize the data lost during the duration of the failure while maximizing the data delivered in the same time frame. It achieves this (Fig. 10) by forcing the nodes to transmit primary data to only those destinations where the data path to the destination does not use the broken link. Similarly, the spatial reuse algorithm adjusts to account for the broken link,

9

1.4E+06

1.E+06

Load=0.375 Load=0.75 Load=1.125 Load=1.875

1.2E+06 Data delivered (bytes)

Data lost (bytes)

8.E+05

Load=0.375 Load=0.75 Load=1.125 Load=1.875

6.E+05 4.E+05 2.E+05

1.0E+06 8.0E+05 6.0E+05 4.0E+05 2.0E+05 0.0E+00

0.E+00 0

1

2

3

4

0

1

Node of Failure (NOF)

Data delivered (bytes)

6.E+05

Load=0.375 Load=0.75 Load=1.125 Load=1.875

4

We can compare all three protection schemes more easily by looking at Fig. 11. The charts in Fig. 11a and 11b are created with a constant load of 15, while the load is varied for Fig. 11c and 11d and results are shown for node 4 (NOF).

4.E+05 1.E+06

2.E+05 0.E+00

Basic Hold-Back Intelligent

8.E+05 1

2

3

4

Node of Failure (NOF)

Fig 9. Token Hold-Back Scheme a) data lost b) data delivered

Data lost

0

when a node becomes “aware” of the fault. Specifically, when a node receives a primary control header, its scheduler assigns the time slot for the secondary traffic to the longest VOQ amongst its valid destination nodes, which now excludes any destination for which the secondary transmission would have to traverse the broken link. Additionally, the dependence of the data lost on the location of the Master Node, is eliminated by re-assigning the identity of the Master Node to the node that detects the failure (NOD).

6.E+05

4.E+05 2.E+05

0.E+00 0

1.E+06

6.E+05

Data delivered

8.E+05

4.E+05

2

3

4

Basic Hold-Back Intelligent

1.2E+06

Load=0.375 Load=0.75 Load=1.125 Load=1.875

1

Node of Failure

1.4E+06

Data lost (bytes)

3

Fig 10. Intelligent Token Regeneration Scheme a) data lost b) data delivered

1.E+06 8.E+05

2

Node of Failure (NOF)

1.0E+06 8.0E+05 6.0E+05 4.0E+05 2.0E+05

2.E+05

0.0E+00

0.E+00

0

0

1

2

3

Node of Failure (NOF)

4

1

2 Node of Failure

3

4

10 VI. CONCLUSIONS 1.E+06 Basic Hold-Back Intelligent

Data lost

8.E+05 6.E+05 4.E+05 2.E+05 0.E+00 0

0.5

1

1.5

2

LOAD

1.2E+06

Basic Hold-Back Intelligent

1.0E+06 Data delivered

In this paper, we proposed an Optical Burst Transport (OBT) network with both spatial reuse and protection properties. Specifically we identified how protection in OBT can differ from conventional protection in the implementation aspects and also in its inter-working with the spatial reuse property of the WDM ring network. We suggested three different protection schemes and identified the parameters that can be used to compare the performance of these algorithms in the case of a fiber cut. Lastly we simulated the three schemes under various conditions and concluded that the Intelligent Token scheme is best suited for minimizing data loss and maximizing data delivered during the fault. A fast protection scheme with a low cost implementation has been investigated towards the goal of achieving a significant enhancement of the OBT network, which is a promising candidate for next generation MANs.

REFERENCES

8.0E+05

1

6.0E+05 2

4.0E+05 2.0E+05

3

0.0E+00 0

0.5

1

1.5

2

LOAD

4

5 Fig 11. Comparison of Protection algorithms a) & b) Data lost and delivered for load 15, c) & d) Data lost and delivered for NOF = 4

From Fig. 11a we see that the Intelligent Token scheme performs just as well as the Token Hold-Back algorithm in terms of data loss and both schemes result in far less data lost compared to the Basic protection scheme, because we can prevent burst loss by holding the token until the light path is recovered or avoid the broken link altogether. Also the former schemes do not show any dependence with respect to the Master Node. At the same time, we see from Fig. 11b that the Intelligent Token algorithm delivers the largest amount of data during the fault when compared to both the other schemes, due to he allowed primary and secondary transmissions which avoid the broken link. In Fig. 11c we see that under all load conditions, again the data loss for the last algorithm is as low if not better compared to the case where tokens are held back and lastly, from Fig. 11d, we see that the data delivery under all load conditions is best achieved in the Intelligent Token case. We thus conclude that the latter algorithm is the optimum case for OBT protection delivering minimized data loss and maximized data transferred during the fault.

6 7 8 9

L. Xu, H. G. Perros, and G. Roukas, “ Techniques for Optical Packet Switching and Optical Burst Switching,” IEEE Communications Magazine, pp. 136-142, January 2001 R. Dutta and G. N. Roukas, “Traffic Grooming in WDM Networks: Past and Future,” IEEE Network Magazine, vol. 16, no. 6, pp. 46-56, November/December 2002. Jaedon Kim, Jinwoo Cho, Mayank Jain, David Gutierrez, Leonid G. Kazovsky, Ching-Fong Su, Richard Rabbat, Takeo Hamada,"Demonstration of 2.5 Gbps Optical Burst Switched WDM Rings Network," Optical Fiber Communication Conference OFC 2006 post deadline paper, Anaheim, CA, March 2006. “Optical Networks - A Practical Perspective”, R. Ramaswami and K. N. Sivarajan, 2nd. edition, Morgan Kaufmann 2002. Kawamura, R., “Architectures for ATM Network Survivability”, IEEE Communication Surveys, vol. 1, no. 1, pp. 2-11, Fourth Quarter 1998. Gerstel, O. and Ramaswami, R, “Optical Layer Survivability: A Services Perspective”, IEEE Communications Magazine, vol. 38, no. 3, pp. 104-113, March 2000. Gerstel, O. and Ramaswami, R, “Optical Layer Survivability – An Implementation Perspective”, IEEE Journal on Selected Areas in Communications, vol. 18, no. 10, pp. 1885-1889, October 2000. Gerstel, O. and Ramaswami, R, “Optical Layer Survivability: A Post-Bubble Perspective”, IEEE Communications Magazine, vol. 41, no. 9, pp. 51-53, September 2003. National Laboratory for Applied Network Research, Measurement Operations and Analysis Team, http://pma.nlanr.net/Datacube/.