Routing Information Exchange in Large-Scale WDM Networks Yong Zhu, Admela Jukan and Mostafa Ammar Georgia Institute of Technology, Atlanta, GA 30332, USA {yongzhu, ajukan, ammar}@cc.gatech.edu Abstract—In this paper, we focus on issues of routing information exchange in large-scale WDM networks. Specifically, we present two schemes, i.e., direct routing information exchange and routing information broker. We then analyze and compare these schemes based on their scalability, complexity and communication overhead. Finally, on examples of metro alloptical networks interconnected over an all-optical WDM backbone, we present the related numerical results based on simulations. I. INTRODUCTION In large-scale optical networks that generally include multiple domains, such as backbone and metro networks based on WDM, reliable routing information exchange among different parts of the network is an important issue. For multidomain networks or for networks with multiple segments [1], i.e., where an optical network is composed of a set of networking segments, each representing a sub-network that requires separate consideration for wavelength routing and administration, this function involves both intra-segment and inter-segment routing information exchange. There is very little existing work addressing and evaluating routing information exchange, in particular for multi-segment networks. Pendarakis et al., considered routing information exchange across the UNI and NNI [2]. In [3], OSPF was extended to support carrying link state information for Generalized Multi-Protocol Label Switching (GMPLS). Wang et al., proposed OSPF extensions for metro/core interworking when routing a connection across multiple subnetworks [4]. However, little work has been done to evaluate the performance of different routing information exchange schemes. In this paper, we first present two solutions for routing information exchange: direct routing information exchange and routing information broker, differing in the location where the routing information is maintained and routing decision is made. In combination with various information refreshing approaches, e.g. periodical or event-driven, these schemes are used to handle dynamic network conditions with low overhead. Furthermore, they are adapted to multi-segment networks in a hybrid way to allow gradual upgrading. We analyze and compare these schemes based on their scalability, complexity and communication overhead. Finally, we present related numerical results on examples of metro all-optical networks interconnected over an all-optical WDM backbone. II. ROUTING INFORMATION EXCHANGE ARCHITECTURE A. Network model The network we are using for evaluation is composed of nodes and links, which are collaborating in the optical control plane to provide routing information and to make the routing decision. The network under control is modeled using the multi-segment model [1], which is capable of representing the network interconnections and capturing the following three critical dimensions: i) heterogeneous, segment-specific topologies and routing objectives; ii) global, intra-segment, and local, inter-segment, traffic; iii) gateway adaptation on segment boundaries. Clustering into segments can reflect any consideration for wavelength routing such as administrative domain, management policy, traffic aggregation, or vendor specification. In this paper, we assume the existence of multiple segments, e.g. metro segments interconnected over the optical backbone (Fig. 1). The above model is generic, since every network that is not segmented can be represented as a segmented network with a single segment. Metropolitan network Metropolitan network S4 S3 S1 Backbone network Internal Link Gateway Link S2 S5 Metropolitan network Metropolitan network Fig. 1. Metro/Backbone multi-segment network Wavelength routing in optical networks, i.e., path selection and wavelength allocation, requires the control plane to provide the following routing information regarding current optical network states: • Topology: includes both connectivity within segments and segment interconnections through gateways. • Resource availability: number of wavelengths on links (i.e., wavelength capacity), wavelength continuity constraints. • Gateway adaptation information: number of gateways, their placements and corresponding adaptation properties (wavelength conversion, waveband interchangeability). • Administrative information: such as choices of segment specific routing algorithms, gateway selection rules [1], and protection/restoration considerations. Based on how the routing information is maintained and exchanged, we present two schemes, namely direct routing information exchange and routing information broker. B. Direct routing information exchange (DRIE) In the Direct Routing Information Exchange (DRIE) scheme, routing information is exchanged directly among nodes, either through in-band or out-of-band channels. DRIE can be accomplished through either flooding or hierarchical distribution. In flooding based DRIE (Fig. 2.a), each node builds a consistent view of the whole network by advertising its local information and forwarding any received information to all its neighbors, except for the node that the information comes from. This involves excessive amount of information exchange and does not scale well. A more scalable approach is hierarchical information exchange (Fig. 2.b), where the routing information is flooded within the local segment and only the aggregated information (such as segment connectivity) is distributed to other segments. The tradeoff of this approach is that each node has detailed local information and only aggregated information regarding other segments. (a) RIB have the flexibility of providing rich functionality in terms of routing and signaling, while at the same time hiding segment-specific implementations or administrative policies. D. Hybrid routing information exchange In a heterogeneous optical network, different segments can have different properties such as traffic pattern, wavelength conversion capability, and optical technology. Specifically, each segment can have its own solutions of routing information exchange, either direct exchange or through brokers. Fig. 4 shows an example of hybrid architecture, where segment A has a RIB and the other two segments implement hierarchical based direct information exchange. This flexibility is important such that it allows network carriers a smooth upgrade. (b) A Networking Segment Gateway and Gateway Link Routing Information Exchange Fig. 2. Direct routing information exchange: (a) flooding based DRIE (b) hierarchical DRIE. C. Routing information broker (RIB) The idea of broker has been proposed in differentiated services [5], where the bandwidth broker (BB) is an agent responsible for allocating preferred service to users as requested, and for configuring the network routers with the correct forwarding behavior for the defined service. We use this idea to overcome the problem of the DRIE scheme, where the basic assumption is that all segments can directly communicate with each other. This might be hard to achieve among heterogeneous segments from different vendors. This motivates us to consider the Routing Information Broker (RIB), a separate entity sitting on top of the network collecting and maintaining routing information and accomplishing a variety of routing functions. (a) (b) Networking Segment Gateway and Gateway Link RIB Broker-Broker Routing Information Interface Exchange Fig. 3. Options of optical routing information broker: (a) central RIB (b) segment specific RIB. Two RIB architectures are possible: central broker (Fig. 3.a) and segment-specific broker (Fig.3.b). The central RIB collects all the routing information of the whole network through some information exchange channels, maintains the routing information database and makes the routing decision. On the other hand, segment specific RIB architecture has multiple brokers and each of them perform local wavelength routing functions for a single segment, e.g. based on vendor-specific implementations or administrative policies. Brokers also exchange global information with each other. Segment specific B Fig. 4. C Hybrid routing information exchange. E. Refreshing the routing informaiton The state of optical network resources is not expected to be static. It can be dynamic in case of link/node up/down, gateway up/down, or based on resource availability change (wavelength capacity, free wavelengths). Outdated routing information tends to deteriorate wavelength routing performances and may even result in a sub-optimal or resource-inefficient routing decision. Keeping the routing information up-to-date is therefore of paramount importance. The refreshing of the routing information can be periodical or event-driven. Periodical refreshing may be suitable when the network information tends to be more static, in which case the information can be relatively infrequently refreshed, to reduce communication overhead. Another way is to trigger refreshing by events such as changes in topology, wavelength capacity, and wavelength utilization. For example, when wavelength utilization reaches a threshold, the link is required to advertise the current available wavelengths. Since only the link information related to the change is refreshed, event-driven refreshing has the advantage of smaller routing information exchange overhead. A special event can be a connection triggered refreshing, where some important connections (such as long term, high bandwidth connections) will solicit the network for the current routing information. In the eventdriven approach, the refreshing time depends on the traffic pattern and the routing algorithm, i.e. bursty traffic will result in more frequent refreshing of the routing information. The periodical and event-driven approaches can be also combined to achieve a more flexible solution in terms of low communication overhead and accuracy. For example, using infrequent periodical refreshing with event-driven updates can provide quick responds to bursty local network state changes and capture smoother fluctuations with low communication overhead. F. Interactions with wavelength routing Routing in the optical control plane includes two critical dimensions: routing information exchange and wavelength routing (i.e., path selection and wavelength allocation). On the one dimension, the routing algorithm performs the path selection and wavelength allocation based on the available routing information. On the other dimension, wavelength routed networks need to provide up-to-date and consistent routing information to the routing algorithm. We have developed three wavelength routing algorithms for multi-segment optical networks: end to end shortest path (E2E), concatenated shortest path routing (CSR) and hierarchical routing (HIR), differing in the way the global traffic is accommodated and the routing information is maintained [1]. E2E routing assumes every node maintains full information of the network and selects the end-to-end path using global shortest path algorithm. In CSR routing, each segment decides the route and allocates wavelengths only based on local information. Gateways, on the other hand, make the decision regarding to the next segment towards the destination based on the segment interconnection information. HIR routing is between E2E and CSR in the sense that all nodes maintain local information and some inter-segment connectivity information such that they can directly choose the right gateway towards the next segment to the destination. Routing Algorithm E2E CSR HIR PNNI Flooding DRIE HIR DRIE Central RIB Segment RIB Routing Information Exchange Schemes Fig. 5. Relationship between routing algorithms and routing information exchange scheme. Both flooding based DRIE and central RIB schemes can provide the routing algorithm full routing information of the whole network, therefore, they can accommodate any routing algorithms including E2E, CSR and HIR. On the other hand, both hierarchical DRIE and segment specific routing schemes can only provide detailed routing information regarding the local segment and aggregated information regarding other segments. This is sufficient for both CSR and HIR routing but inadequate for end-to-end path selection and wavelength allocation. PNNI-based routing is similar to the CSR routing since it requires internal nodes to maintain local information and peer group leaders maintain global information. Therefore, PNNI can be accommodated by all the routing information exchange schemes we proposed. Fig. 5 summarizes the relationship of routing algorithms and routing information exchange schemes. Wavelength routing is based on the approximate information (due to dynamic conditions) given by routing information exchange schemes and the traffic pattern and routing algorithms will affect the fluctuation of resource utilization which will in turn trigger routing information refreshing. Therefore, problems of routing information exchange and the wavelength routing are coupled. In this scenario, how to evaluate the performance of the routing function as a whole is an interesting research issue. III. ANALYSIS OF ROUTING INFORMATION EXCHANGE SCHEMES To be able to quantitatively analyze how different routing information exchange schemes will perform and scale, we identify the following parameters related to the scalability, complexity and communication overhead. • Bandwidth requirements (BR): In all routing information exchange schemes, the routing information is disseminated in forms of Routing Information Advertisement packets (RIA packets). Since the RIA packets are handled by individual nodes (or brokers) and transmitted through links, the amount of RIA packets determines requirements of both the process power at individual nodes and bandwidth on links in the control plane. • Memory requirements (MR): All routing information exchange schemes need to maintain databases regarding current network states. The size of the database determines the memory requirements at network nodes and brokers. In the rest of this section, we will analyze each routing information exchange scheme based on these issues. As described previously, routing information can be refreshed periodically or triggered by events. The bandwidth requirement analysis here is for periodical refreshing. For the event-driven refreshing, bandwidth requirements cannot be estimated without the knowledge of the traffic pattern and wavelength routing algorithms. A. Analysis of the direct information exchange In flooding based DRIE, since the routing information regarding each optical link is flooded throughout the control plane, any control plane link carries at least one RIA packets for each optical link during any refreshing period. This determines the lower bound of the link bandwidth requirement. If a node has seen the incoming packet before, it silently discards the packet. Therefore, there are at most two copies of the same RIA packet carried by any single link, this refers to the situation where both ends of the link forward the same RIA packet at simultaneously. Assume the network under control has a total of L links, the size of the RIA packet is S, and the refreshing period is T, then the link bandwidth requirement can be expressed as follows: L⋅S L⋅S (1) ≤ B < 2⋅ T R T The flooding based DRIE scheme results in synchronized routing information database at each node. The size of such database is given by: (2) L⋅R R is the size of routing information record for a single link. In hierarchical DRIE, detailed intra-segment information is flooded within the local segment and inter-segment information is exchanged among all segments in the network. We differentiate between two kinds of links: internal link (connecting nodes within the same segment) and gateway link (connecting nodes from consecutive segments). Internal links carry both intra-segment RIA packets and inter-segment RIA packets and gateway links carry only inter-segment RIA packets. For the kth segment, the bandwidth needed for intrasegment RIA is: Lk ⋅ S k Lk ⋅ S k (3) k Intra k T Intra ≤ B Intra < 2 ⋅ Intra k T Intra where Lk is the total number of links in the kth segment, SkIntra and TkIntra are intra-segment RIA packet size and intra-segment refreshing period of the kth segment respectively. The bandwidth required for inter-segment advertisement is: L ⋅S LG ⋅ S Inter (4) ≤B < 2 ⋅ G Inter R _ Inter TInter TInter where SInter and TInter are inter-segment RIA packet size and inter-segment refreshing period respectively, LG is the total number of gateway links. Therefore, the bandwidth requirement for an internal link in the kth segment is: Lk ⋅ S k Lk ⋅ S k LG ⋅ S Inter (5) LG ⋅ S Inter k Intra k T Intra + TInter ≤ BR < 2 ⋅ ( Intra k T Intra + TInter ) And the bandwidth requirement on the gateway link is BR_Inter given in (4). In hierarchical DRIE, every node is only required to store intra-segment information and aggregated inter-segment information. Therefore, memory requirements for nodes in the kth segment is: (6) Lk ⋅ R k + LG ⋅ R Inter In which Rk is the routing information record size of the kth segment and RInter is the inter-segment routing information record size. Comparing results of hierarchical DRIE and flooding based DRIE, we can see that by using multiple segments (i. e., Lk<L), hierarchical exchange can significantly reduce the bandwidth requirements and memory requirements. The improvement depends on the clustering granularity such that using small segments can achieve low intra-segment communication overhead and database size. However, this will increase the total number of segments for the same network, which will in turn increase inter-segment exchanges. Furthermore, small segment means more and more connections will travel across multiple segments and this will increase the blocking probability [1][6]. Equations (1) and (3) also show that the bandwidth requirements are inversely proportion to the refreshing period, which is consistent with the refreshing analysis in the previous section. It should be noted that the information exchanged through hierarchical DRIE can be tuned to accommodate different routing algorithms, for example, in the CSR routing [1] and PNNI routing, the inter-segment information only needs to be maintained at the gateway or the peer group leader (PGL) instead of having the inter-segment information disseminated into the segment. Consequently, the second part of (5) should be removed and (6) becomes the requirements for the gateway/PGL. B. Analysis of the routing information broker scheme In the routing information broker scheme, the routing information is exchanged through channels between the network nodes and the broker. Bandwidth requirements on these channels depend on the physical location of the broker and the topology of the network. Therefore, instead of evaluating the bandwidth requirement on a specific link, we estimate the total bandwidth required to perform routing information exchange. In central RIB, all links periodically report their states to the central broker. Therefore, only one RIA packet per link is needed during any refreshing period and the total bandwidth requirement to the broker is: L⋅S (7) B = R T Unlike the DRIE scheme, which has duplicated routing information databases, a single database is maintained by the broker. The size of the database at the central broker is the same as that of the flooding based DRIE at individual node in (2). In segment specific RIB, routing information in each segment is directly reported to the local broker and we assume inter-segment information is flooded among brokers through broker-to-broker channels. Therefore, there are bandwidth requirements for both collecting local information and brokerto-broker communications, and is given by: Q Q Lk ⋅ S k Lk ⋅ S k ′ LG ⋅ S Inter ′ LG ⋅ S Inter (8) ∑ k =1 Intra k T Intra + LG ⋅ TInter ≤ BR < ∑ k =1 Intra k T Intra + 2 ⋅ LG ⋅ TInter Where Q is the total number of segments and LG′ is the number of gateway links in the control plane. Memory requirement of the kth broker is the same as that of a single node in the kth segment in the hierarchical DRIE scheme given in (6). Comparing above results with corresponding DRIE results, we can see that the total bandwidth for RIB scheme is close to the bandwidth on a single link in DRIE scheme. Furthermore, RIB has memory requirements only on brokers, the number of which is usually much smaller than the total number of nodes. Therefore, we can significantly improve both bandwidth requirements and memory requirements by using broker. C. Effects of wavelength capacity The existence of wavelengths differs the routing information exchange in the WDM networks to IP networks. Wavelength capacity affects the scalability of routing information exchange in the following aspects. Associated with each wavelength, there is a set of routing information, such as service type, active connections carried by the wavelength and wavelength conversion set. All of them need to be advertised throughout the network. Therefore, it is reasonable to assume that the size of the RIA packet for any link is proportional to the wavelengths capacity at that link. Also, under dynamic network conditions, information refreshing can be triggered by the change of wavelength utilization. Therefore, given the same data traffic load, higher wavelength capacity can tolerant more changes and result in less frequent refreshing. IV. EXPERIMENTS AND NUMERICAL RESULTS We perform our simulation on a multi-segment network where a 6 × 6 mesh-torus backbone connects 4 equal-sized metro networks with ring topology (Fig. 1). In the following experiments, we assume all segments have the same packet size and refreshing period and measure the number of packets instead of bandwidth to avoid selections of these parameters. The first experiment illustrates the scalability of different schemes by showing the relationship between the total number of RIA packets and the network size. While fixing the backbone network as well as gateway locations, we increase the network size by increasing the metro segment size. As we can see from Fig. 6, the number of packets goes up with the increasing metro size in all four schemes. DRIE schemes disseminate routing information through flooding so that they generate much more packets than RIB schemes. Hierarchical DRIE is more scalable than the flooding based DRIE since the flooding is limited within the segment instead of the whole network. Since there are only 4 segments, the segment specific RIB requires slightly more packets than the central RIB for inter-segment exchange. Fig. 6 also shows results for hybrid information exchange where the backbone is assumed to use RIB scheme (e.g. Hybrid 2 DRIE means 2 metro segments perform DRIE and others use routing information broker). The total number of packets for hybrid exchange is between that of DRIE and RIB schemes. considered together, both the backbone link change and metro link change require the same number of RIA packets and this number is increasing with the increasing metro size. However, in hierarchical DRIE, the amount of RIA packets for backbone link change is fixed regardless the size of metro segment. This shows an advantage of the multi-segment model where effects of state change is restricted within the local segment and does not affect other segments. Compared with DRIE schemes, RIB schemes are more scalable to the network change such that only one packet is needed to report single link change to the broker (not shown here). The last experiment shows the effect of clustering (Fig. 8). The simulation is performed on a 200-node bi-directional ring. All nodes are clustered into a number of equal-sized segments from 2 nodes/segment (100 segments) to 200 nodes/segment (1 segment). Since both flooding based DRIE and central RIB treat the multiple segments as a whole network, clustering does not have any effect on them. In segment specific RIB scheme, the total intra-segment bandwidth is fixed, but the number of segments is decreasing with the increasing segment size. Therefore, the total number of packets is decreasing due to reduced inter-segment exchange. For hierarchical DRIE, as we explained in previous section, increasing the segment size will increase the intra-segment exchange but decrease intersegment exchange. Therefore, there is an optimal segment size where the best tradeoff between intra-segment flooding and inter-segment exchange can be achieved and the total number of packets is minimized. 5 Fig. 6. Total number of packets vs. metro segment size 1000 Flooding Metro Link Change HIR Metro Link Change Flooding Backbone Link Change HIR Backbone Link Change 900 800 Total Number of Packets 10 4 10 Flooding DRIE HIR DRIE Central RIB Segment RIB 3 10 Num of RIA packets 700 2 10 600 500 40 60 200 100 0 20 40 60 80 100 120 Metro size 140 160 180 200 220 Number of RIA for single link change vs. metro segment size To quickly respond to dynamic network conditions, information regarding changes should be timely refreshed. Fig. 7 shows the number of packets to advertise the state change (e.g. link up/down, wavelength capacity/utilization change) of a single link (metro link or backbone link). To illustrate the scalability, we plot results under increasing metro segment sizes while keeping the backbone fixed. In flooding based DRIE, where both the backbone and metro network are 80 100 120 Segment Size 140 160 180 200 Effects of segment granularity V. 300 Fig. 7. 20 Fig. 8. 400 0 0 CONCLUSIONS We have presented and compared two routing information exchange schemes in this paper: direct routing information exchange and routing information broker. To handle dynamic network conditions, routing information needs to be refreshed which can be either periodically or triggered by events. Both analytical and simulation results showed that using the multisegment routing information exchange can reduce the communication overhead and memory requirements. Furthermore, using multiple segments allows isolation of network changes and providing flexible segment specific control functions. Finally, our simulation also gives the guideline for choosing suitable segment size for a network to reduce the routing information exchange overhead. REFERENCES [1] [2] [3] [4] [5] [6] Y. Zhu, A. Jukan and M. Ammar, “Multi-segment Wavelength Routing in Large-scale Optical Networks,” ICC2003. D. Pendarakis, B. Rajagopalan, amd D. Saha, “Routing Information Exchange in Optical Networks,” IETF Internet draft. K. Kompella et al., “OSPF extensiions in support of generalized MPLS,” IETF Internet draft. D. Wang, J. Strand and J. Yates, “OSPF for Routing Information Exchange Across Metro/Core Optical Networks,” Optical Network Magazine, Sept. 2002. K. Nichols, V. Jacobson and L. Zhang, “A Two-bit Differentiated Services Architecture for the Internet”, RFC2638, July 1999. Y. Zhu, A. Jukan and M. Ammar, "Performance analy-sis of multisegment wavelength routing", IEEE/LEOS Summer Topical Meetings, 2002, Quebec, Canada.