Abstract. 1. Introduction

Transcription

1 INTERCONNECTED WDM RING NETWORKS: STRATEGIES FOR INTERCONNECTION AND TRAFFIC GROOMING* Jian Wang + and Biswanath Mukherjee Networks Laboratory, Computer Science Department, University of California, Davis, CA 95616, USA jnwang@ucdavis.edu, mukherje@cs.ucdavis.edu + Correspondence Author (Tel: (530) , Fax: (530) ) Abstract SONET self-healing ring is a very popular architecture for optical communication networks today. Typically, multiple rings are interconnected together to provide large geographical coverage. Traffic grooming solves the problem of packing different low-speed traffic streams into multiple, high-speed, wavelength channels in a WDM ring with the goal of saving SONET equipment usage. The two problems of how to interconnect WDM ring networks as well as how to groom the traffic in interconnected rings are important problems that will be addressed in this study. However, not much research has been reported in this area so far. In this preliminary work, we first propose several WDM-ring interconnection strategies; then, we formulate the traffic-grooming problems for the proposed architectures. Two heuristic trafficgrooming algorithms are proposed for the cases of optically interconnected and hierarchical network architectures. We find that the optically-interconnected strategy achieves the best ADM savings; however, the hierarchical architecture is suitable for general physical network topologies and it provides moderate ADM savings. A pure digital-crossconnect scheme uses fewest wavelengths among all proposed interconnecting strategies, but it requires more network-wide ADMs. 1. Introduction SONET self-healing ring is a very popular architecture for optical communication networks today. It provides high capacity and inherent reliability. Multiple SONET rings are interconnected together to provide large geographical coverage. Traditional SONET rings use a single wavelength. But with the emergence of WDM technology, these rings need to be upgraded to WDM in order to support the increased traffic demands. Traffic grooming in a SONET ring network solves the problem of how to pack different low-speed traffic streams into a higher-speed stream, each of which may occupy a separate wavelength channel in a WDM ring. In SONET, each wavelength can carry a certain number of lower-speed streams in TDM fashion. Let s consider an example where the line rate of a wavelength channel is OC-48. So it can carry either 4 OC-12 or 16 OC-3 streams. The network operator can decide the finest granularity for the traffic stream (say OC-3); then, any traffic demand can only be an integer multiple of this basic stream (OC-3, OC-6, etc.). Each basic stream is physically carried by a timeslot and this timeslot travels through the entire ring to form a circle. Notice that a circle (timeslot) can carry different traffic flows on different links (fiber segments) because traffic can be added or dropped at intermediate nodes. The SONET equipment that is used to perform traffic-add/drop function is the SONET add/drop multiplexor (ADM). The main function of an ADM is to extract and drop some timeslots from the wavelength it is operating on, insert traffic to empty timeslots in its wavelength, and bypass some timeslots. Notice that the ADM always runs at the same datarate as that of the wavelength it is working on, and it is used at any nodes where there is at least one timeslot that needs to be added or dropped. Since different circles on the same wavelength can carry different connections, how we assign connections to circles and therefore pack them into wavelengths can significantly affect the number of ADMs needed in the network. The objective of traffic grooming is to *This work has been supported in part by Intelligent Fiber Optic Systems (IFOS) and the State of California UC MICRO Program Grant No

2 minimize the total number of ADMs in the WDM ring. Single WDM ring version of this problem has been well studied in the literature [1-7]. The two problems on (1) how to interconnect WDM ring networks as well as (2) how to groom traffic in interconnected rings are important problems that will be addressed in this study. However, not much research has been reported in this area so far. In this preliminary study, we first propose three WDM-ring interconnection strategies; then, we formulate the traffic-grooming problems for each strategy. Two trafficgrooming algorithms for two of the interconnecting strategies (optically-interconnected and hierarchical) are proposed (the third interconnection strategy, based on digital crossconnects, does not require any new traffic-grooming approach). Numerical results for the ADM usage for the different strategies are compared. 2. Interconnection of Two Rings 2.1. General Concerns about Connecting Two Rings When we design interconnected rings, at least two physical intersections are desired between two rings due to the fault recovery concern. When a node failure occurs at one intersection node, the rest of the nodes are still connected and the auto recovery mechanism can kick in to resume the traffic flows. Figure 1 shows a general case of how to physically interconnect two rings. Common node General node Common link General link Figure 1. General example of two interconnected rings. In Fig 1, several important geometric characteristics of interconnected rings are shown. Common nodes are the nodes shared by both rings. Traffic that is sourced from or sunk into a common node can be carried through the existing rings regardless of where the other end of the connection is. Traffic between general nodes (not shared by the two rings) can be classified into two categories. If the two ends of the traffic are on the same ring (intra-ring traffic), they can be carried directly on the local ring. If the two ends of the traffic are on different rings (inter-ring traffic), then either a multihop approach will be needed to bridge the traffic on different rings, or a new ring will need to be formed to cover both of the nodes. Common links are the links shared by both rings. Physically, the fiber on a shared link can be either shared or not, but the two end nodes of the link must be shared. If the fiber on shared links are also shared, then the two rings can not use the same wavelength to transmit on the same directions on that link (they can use same wavelength to transmit on different directions if there is no amplifier on the link). Because of the fault recovery property, each self-healing ring must be complete, i.e., part of the ring can be idle but can not be missing. A simplified node architecture of a general WDM ring is shown in Figure 2. Only one fiber of the ring is shown here. In this work, we assume the BLSR/4 (Bidirectional Line-Switched Rings) with 1+1 protection (traffic transmits simultaneously on both working and protection fibers and it is up to the destination to pick one of them for reception), so there will always be four fibers goes through each node. On each wavelength, a Wavelength Add/Drop Multiplexor (WADM), which is a 2 2 optical switch, is used to selectively add or drop the wavelength. The wavelength will be dropped to a ADM, in which some TDM timeslots will be dropped to local stations, and some other timeslots will be added back to the wavelength. Independent of the physical topology, the network can be logically organized differently by using crossconnect devices. In this study, if traffic gets dropped from one ring to a digital crossconnect and then gets added to another ring, we count this as two hops. However, if traffic get switched to another ring

3 Optical DMUX Fiber SONET ADM Local Add/Drop ports Optical bypass Enlarged view of SONET ADM 2 2 optical switch Local Add Incomming Outgoing Wavelength wavelength SONET frame Local drop Optical MUX Figure 2. Simplified node architecture of a general WDM/SONET ring network. through an optical switch without being demultiplexed, then we don t count the extra hop. In the following sections, we will present the three proposed interconnecting strategies Strategy 1: Using Digital Crossconnect at Intersection Nodes (cross-connect at SONET level) Fiber of first ring Low speed add/drop ports ADM Wavelength Channels Digital crossconnect To local Add/drop Fiber of second ring Figure 3. Simplified architecture of the intersection node (for the proposed case of using digital crossconnect to connect two rings).

4 The first strategy for connecting two rings is to connect them at SONET level. At each intersection node, a digital crossconnect (DXC) is used to connect the local add/drop ports of some ADMs on the two rings. Figure 3 shows the simplified architecture of the intersection node. The function of the proposed DXC is to take lower-speed streams, which are the outputs from some ADMs, as inputs, and then send them to appropriate input ports on another ring. The DXC is controlled by software and is highly configurable. For sake of simplicity, we assume that the DXC has a large number of ports so one DXC will be enough for any intersection node. We also assume that the DXC doesn t have ADM functionality. At the intersection node of either ring, part of the traffic will be dropped to local stations (which may also be higher layer network routing or switching devices like IP router, ATM switch, etc.). All the inter-ring traffic will be dropped to the DXC. The DXC sends traffic to the appropriate add-port on another ring. Notice that, for this interconnecting strategy, all the intra-ring traffic goes only on hop. All the inter-ring traffic goes two hops. The intersection nodes are special because they belong to both rings. The traffic between one of them and any other node will be carried by a local ring and it will travel only one hop Strategy 2: Using Optical Crossconnect at Intersection Nodes (cross-connect at wavelength level) The second strategy pushes the cross-connect function down to the optical layer. The basic idea is to dedicate some wavelengths to form virtual rings that go across the boundaries of the physical rings. Virtual rings are used to carry all the inter-ring traffic in the network. To do so, we need the capability to optically route some wavelengths across rings at the intersection node. We propose to use a 3 3 switch on each wavelength at the two intersection nodes that are farthest apart (here we assume that the major parts of the two rings are disjoint). The diagram of the switch and its possible transition function is shown in Figure 4. 1 Add 2 3 Ring1 Ring Drop Port States a b c d e f Figure optical switch for an intersection node Figures 5(a) ~ (c) show the details of how this switch can be configured to form virtual rings. Figure 5(a) show the configuration which the light follows the physical ring. Some wavelengths of the WDM network will be configured this way in order to carry all the intra-ring traffic and they are referred to as local rings in this study. Figure 5(b) shows the configuration where the two virtual rings carry part of the inter-ring traffic. For a certain class of physical topologies, one of these two rings will be enough to carry all of the inter-ring traffic. The physical topology of this class of networks has the following feature: every common node of the two rings is also the end of at least one common link and all the common links on the network are contiguous. Basically, the two rings share one (and only one) common part where all the nodes are shared, and there is no shared node anywhere else. Since all the nodes in the common part are shared by both rings, the traffic between one of them and another node can be carried by one of the local rings, so we only need one super-ring which covers all the non-common nodes.

5 Virtual ring 1 One big virtual ring Ring 1 Ring 2 Ring 1 Ring 2 Ring 1 Ring 2 Virtual ring 2 (a) Configuration 1: Same as physical rings (b) Configuration 2: Two virtual rings (c) Configuration 3: One big virtual ring Figure 5. Optically cross-connected double rings. For general interconnected rings as in Figure 1, even both of the virtual rings in Fig. 5(b) will not be able to carry all the inter-ring traffic. Figure 5(c) shows the configuration that one virtual ring covers all the nodes on both rings. This solves the connectivity problem; however, these two rings have only one common node and if this point fails, the virtual ring will not be recoverable. Besides the fault recovery problem, this pure optical cross-connect strategy also suffers from the ring size limitation of BLSR. The BLSR requires a maximum size of 16 nodes, so not all the rings can be interconnected by this approach. Because of the above problems, this approach is not proposed for general interconnected rings except for the case that the two rings are sharing one common part. Since one super-ring can carry all the inter-ring traffic, this case is refereed to as the super-ring case for all optically-connected double rings. Notice that, for this interconnection strategy, all of the traffic, intra-ring and inter-ring, will go only one hop and no OXC is used Strategy 3: Hierarchical Architecture (mixed SONET and optical layer crossconnection) To solve the problems encountered in Strategy 2 and still keep the advantage of using optical interconnection, we propose this hierarchical architecture. In this strategy, dedicated wavelength are still used to form virtual rings (referred to as a hyper-ring here) that cover some nodes of each ring. At the selected nodes, the local rings will be connected to the hyper-ring through DXC. Only the local stations at the logical interconnection nodes can access the hyper-ring directly. Since local stations on other nodes can not access the hyper-ring directly, inter-ring traffic from these nodes needs to travel to the logical interconnection nodes through the local ring first, and then get on to the hyper-ring. By doing so, all the intra-ring traffic travel only one hop. Inter-ring traffic may travel between 1 and 3 hops. The one-hop case occurs when the two end nods are both logical interconnection nodes. The two-hop case occurs when one of the end nodes is a logical interconnection node. The three-hop case occurs when none of the end nodes is a logical interconnection node. Notice that the intersection points of the hyper-ring and the local ring can be chosen independent of the physical connections. We can choose to put the DXC at any convenient point and just an optical (3 3) switch at the physical interconnection node. This gives us the freedom to choose the logical topology that best benefits the design goal. This strategy can also be used to interconnect more than two rings together; however, because the ring-size limitation of BLSR is 16, more than one layer of a hyper-ring may be needed for large networks. In this study, we constrain our discussion to just one hyper-ring.

6 Figure 6 illustrates all the three proposed strategies. (a) Physical topology (b) Digital cross-connected ADM Local station DXC (c) Optically connected (d) Mixed architecture Figure 6. Illustrative view of different interconnection strategies. 3. Traffic-Grooming in Interconnected Rings The traffic-grooming problem in a single ring is well studied. In [7], a formal mathematical problem definition is provided, which turns out to be a linear programming problem; also, an improved approach for arbitrary (static) traffic is discussed. In [2], arbitrary traffic on an unidirectional ring with egress node is discussed. In the three strategies studied, a specific traffic may travel on one or multiple rings. However, if the status of the traffic on one ring is independent of that in another ring, then the traffic-grooming problem can be done in the two rings separately. The problem definition provided in [7] can be used on each ring without change. Because the computational time for exhaustive search is very long, heuristics are needed for discover near-optimal result for this study. Most of the previously proposed heuristics can be used here; however, the traffic pattern for inter-ring traffic has its own characteristics that is unnatural to ordinary ring networks, so new heuristics are needed for some cases. In this section, we start with discussing the traffic pattern of different network architectures, then we propose corresponding traffic-grooming algorithms. In this study, the cross-connect cost is not counted (although its cost is not negligible) for sake of simplicity. For static traffic pattern, it is arguable since the crossconnection can be done with a fixed scheme at the design stage.

7 3.1. Traffic Pattern for Different Interconnected Double-Ring Architectures In all the following discussion, we assume arbitrary (static) traffic pattern. A randomly generated traffic matrix represents the traffic among all the nodes in the network. Each item in the traffic matrix is a linearly distributed random number between zero and a certain limit. For the case of interconnecting two rings, we assume that the number of nodes in each ring is N and M; then, statistically, the internal traffic in each ring is proportional to N 2 and M 2, respectively. The inter-ring traffic is roughly proportional to 2 N M. When N and M are comparable, the inter-ring traffic is about twice as much as the internal traffic. So, for interconnected rings, inter-ring traffic should be given enough attention Digital cross-connected rings Traffic in each ring consists of two parts. The first part is the internal traffic, which has the same characteristic as that of ordinary rings. The second part is the inter-ring traffic. Each inter-ring connection has to go through one of the intersection nodes that is decided by some global routing algorithm to get switched to the appropriate wavelength and timeslot on the other ring. All the internal and inter-ring traffic will be mixed and carried together on each ring. The traffic grooming in digital cross-connected double-rings is equivalent to two independent problems in a single ring. The traffic matrix, which represents the global traffic pattern, will be broken into two parts for the two rings, respectively. In this study, after the traffic is broken down, each ring is treated with the circle-assignment and traffic-grooming (greedy heuristic) algorithms proposed in [5]. More elaborated algorithms like the Simulated-Annealing based algorithms is also suitable for this task [7, 6]; however, in this preliminary work, we adopt just the basic approach in [5] and concentrate on the comparison and the difference between the different interconnection strategies rather than search algorithms Optical cross-connected rings As we have stated before, only the super-ring case of optically-connected rings is discussed here. This kind of interconnected rings is very common in existing networks. When this kind of network needs to be upgraded to a WDM-based network, this optical cross-connected solution may be the most cost effective one. The traffic on the three rings, i.e., the super-ring and the two local rings, are inherently independent, so the problem definition provided in [7] is suitable for each of them. The traffic pattern on the super-ring is quite different from that of the ordinary rings. This difference makes the previously proposed trafficgrooming algorithms (heuristics) perform badly. The physical intersection nodes are the bottleneck of the super-ring in the sense that all the traffic will pass through one of them. Since there is no connection between any node pair within one side (based on the border of the two rings) of each ring, when using ordinary traffic-grooming algorithm, most ADMs will only be used for half of their capacity although the other part is still needed for fault-recovery reason. We propose a special traffic-grooming algorithm for inter-ring traffic on super-ring and it will be discussed in a later section Hierarchical rings As we have discussed before, since the hyper-ring and local rings are digitally crossconnected, the rings are logically independent from each other if the traffic is properly broken down. This problem is the same as traffic grooming on three independent single rings. In this study, we only consider the case where the hyper-ring has two logical connections with each physical ring. Statistically, the traffic loads among all nodes are evenly distributed, so we only consider the farthest node pairs, which have equal chance of getting traffic from either side of the ring, as the candidates for logical interconnection nodes. This will help to avoid the potential of wasting wavelengths as well as ADMs at the interconnection nodes. Since all the inter-ring traffic goes to the logical connection nodes, and the inter-ring traffic is the dominant component in each ring, we have similar problems as those in the super-ring case. Most of the ADMs will be used for only half of their capacity if all the traffic goes to its nearest interconnection node. A new algorithm (heuristic) is proposed to help prevent this problem from happening and it will be discussed later.

8 Traffic on Hyper-Ring The traffic on hyper-ring for the hierarchical-ring strategy is discussed separately in this section. If we consider the two interconnection nodes between the hyper-ring and a local ring as one super-node, then the traffic pattern among super-nodes on the hyper-ring is the same as that of ordinary rings. This simplification is valid if no local traffic is brought up to the hyper-ring; in another words, there is no traffic between the two nodes within the same super-node. It is interesting to discuss the number of wavelength requirement on the hyper-ring for uniform traffic. There are four connections between any super-node pairs. If we normalize all the traffic between any physical node pair (except the one within the same super-node) to one, then the total number of traffic on each link will be N for odd number of nodes and N for even number of nodes (total number of super-nodes is N). Notice that the number given above is suitable not only for inter-super-node link, but also for the link within any super-node Improved Traffic Grooming Algorithm for Inter-ring Traffic on Optical Cross- Connected Rings As we have analyzed before, the proposed traffic-grooming algorithms for general traffic patterns will not work well for the super-ring. The traffic pattern differs from others in the sense that only cross border traffic exists on the ring. We propose a very intuitive algorithm for this super-ring traffic. An example is shown in Figure. 7. Ordinary traffic grooming algorithms tend to put a bidirectional link along the same path as shown in the Fig. 7(a). Obviously, ADM waste in a super-ring is caused mainly by the idle connections in the ring. Our proposed algorithm will try to construct full circles by putting the mutually opposite connections in the same ring. A connection pair, instead of being put on two fibers, can be put on one of the two fibers on the same wavelength and same timeslot. By doing this, we can increase the ADM utilization so as to decrease the total number of ADMs in the network. (a) (b) Figure 7. Traffic-grooming example for Super-Ring. The proposed algorithm is described in the following pseudocode: Begin { Put all the mutually opposite connection pairs into unidirectional full circles first; for each non-paired connection, do{ check whether there is room left for it in the existing circles { put it in the existing circle; // yes branch else { put it into a new circle along the shortest pate; // no branch

9 3.3. Improved Traffic-Grooming Algorithm for Hierarchical Architecture For hierarchically-interconnected rings, inter-ring traffic will be carried to an interconnection node by the local ring. We propose an algorithm that tries to distribute the inter-ring traffic along the four possible paths as evenly as possible. The algorithm can be described in the following pseudocode. The objective is to try to eliminate the idle connections mainly caused by the inter-ring traffic on the local ring. Begin { for all the connections from one ring to another, do { assign them to each of the four possible paths sequentially; 4. Numerical Results and Analysis 4.1. Interconnected Double-Ring To compare the three proposed architectures, we assumed the following two interconnected rings. The first ring (local ring 1) consists of 7 nodes and the second ring (local ring 2) consists of 8 nodes. They are interconnected at two consecutive nodes. The capacity of each wavelength is assumed to be 4, i.e. each wavelength can carry 4 circles each with capacity of one unit, or we view the wavelength channel rate is OC-12, and the granularity of traffic is OC-3. The traffic on the common link is equally assigned to both local rings for all of the architectures. For Strategy 3, the interconnect nodes between the hyper-ring and local rings are fixed, i.e., the effects of variations in connection node position are not intended to be studied here. Nevertheless, experiments have shown that the interconnecting position will not affect the result statistically. Figure 8 shows the average ADM usage for the three proposed architectures. The result is obtained from the average of 30 trials. In each trial, a randomly generated traffic matrix is used with each element of traffic ranging from 0 to 16 units. For each ring in each case, the greedy heuristic [5] is used on top of the basic traffic-grooming algorithms discussed above. Experiments show that the greedy heuristic will not always improve the result. No matter whether it can or not, the best result is used. In Figure 8, the local rings 1 and 2 are the 7-node and 8-node ring, respectively. The top part for Strategy 2 is the average Average Number of ADMs Strategy 1 (DXC) Strategy 2 (OXC) Strategy 3 (hierarchical) hyper-ring (super-ring) local ring 1 (7 nodes) local ring 2 (8 nodes) Figure 8. Average number of ADMs used for the three proposed inter-connecting strategies.

10 number of ADMs used by the super-ring. The top part for Strategy 3 shows the average number of ADMs used by the hyper-ring. In general, the optically-connected architecture (Strategy 2) achieves the best ADM savings. For this architecture, the number of wavelengths used in the super-ring is much more than that in any one of the local rings. The ADMs are also used mainly in the super-ring. Our proposed algorithm shows significant savings. Experiment shows that, when using shortest-path routing, the ADM usage on the super-ring will increase about 30% over our result. The hierarchical architecture (Strategy 3) can not achieve the savings reached by the optically connected rings. It still shows significant improvement when compared with the digital cross-connect strategy. If we just look at the ADM usage on the two local rings, the effect of the proposed traffic-grooming algorithm is easy to see. For both of these two cases, each local ring needs to carry all the intra-ring and inter-ring traffic (not like Strategy 2 where local ring carries only local traffic), so the traffic load are exactly the same on every local ring for strategies 1 and 3. The idea of separating the two interconnection nodes coupled with the proposed traffic-grooming algorithm managed to decrease the ADM usage to a certain level so that, even with the extra cost paid on the hyper-ring, it saves over the first strategy. The digital cross-connect strategy (Strategy 1) turns out to be the one that saves most on wavelengths. Our calculations also show that the wavelength capacity will not affect the relative ADM usage among the three architectures Interconnected Multi-Ring We now show the traffic-grooming results for another network topology. The physical topology of the network is shown in Figure 9. Thirty randomly-generated traffic matrices are used to get the average ADM usage. Only Strategy 3 (hierarchical architecture) is used. The hyper-ring (chosen arbitrarily for illustration) consists of the following eight nodes: 11, 9, 7, 6, 15, 13, 13, and 1. Notice that the nodes 7 and 6 only have Ring 1 Ring2 Ring 3 Ring Average number of ADMs Hyperring (8 nodes) Localring 1 (5 nodes) Localring 2 (4 nodes) Localring 3 (7 nodes) Localring 4 (8 nodes) Figure 9. Another test network topology. connection with ring 2, while node 13 is connected to both rings 3 and 4. Figure 10 shows the average number of ADMs that is used on each ring. Unlike the case in Figure 8, now, the number of ADMs on the ring has become the dominant part in the whole network. 5. Conclusion Figure 10. Average ADM usage for the multiring network.

11 In this paper, we proposed three strategies to inter-connect WDM/ring networks. The traffic-grooming algorithms on these different architectures were compared. The optically cross-connected strategy proved to be the most cost effective one; however, it is not applicable to general interconnected rings. The hierarchical strategy provides moderate ADM savings and it can be easily used for any physical topology. The digital cross-connected rings can achieve the minimum wavelength usage in general. The result of using hierarchical architecture to interconnect multiple rings was also shown. References [1] O. Gerstel, P. Lin, and G. Sasaki, Wavelength assignment in a WDM ring to minimize cost of embedded SONET rings, Proc., IEEE INFOCOM 98, pp , March [2] E. H. Modiano and A. L. Chiu, Traffic grooming algorithms for minimizing electronic multiplexing costs in unidirectional SONET/WDM ring networks, Proc., CISS 98, March [3] J. M. Simmons, E. L. Goldstein, and A. A. M. Saleh, Quantifying the Benefit of Wavelength Add- Drop in WDM Rings with Distance-Independent and Dependent Traffic, IEEE/OSA Journal of Lightwave Technology, vol. 17, no. 1, pp , Jan [4] X. Zhang and C. Qiao, An effective and comprehensive solution to traffic grooming and wavelength assignment in SONET/WDM rings, Proceedings of the SPIE, vol. 3531, pp , [5] G. Ellinas, G.K. Chang, M.Z. Iqbal, J. Gamelin, and M. R. Khandker, Wavelength-selective crossconnect architecture interconnecting multiwavelength self-healing rings, Proc., OFC 97, pp , Feb., [6] W. Cho, J. Wang, and B. Mukherjee. Unification and extension of the traffic grooming problem in WDM ring network, OFC [7] J. Wang, W. Cho, and B. Mukherjee. Improved Approaches for Cost-effective Traffic Grooming in WDM Ring Networks: Non-uniform Traffic and Bidirectional Ring, Submitted to ICC 2000.