> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 1 On Shared Risk Link Group Optimization (Invited) Guangzhi Li, Dongmei Wang, Timothy Gallivan, and Robert Doverspike AT&T Labs, New Jersey, USA, {guangzhi.li,dongmei.wang,timothy.gallivan,rdoverspike}@att.com Abstract— Shared Risk Link Groups (SRLGs) have been investigated during the past 10 years due to optical network innovation and deployment. Dense Wavelength Division Multiplexing (DWDM) technology has significantly increased the data transmission capability per fiber. Service providers build their overlay networks, such as IP networks, on top of optical networks and all optical networks are built over some combination of DWDM equipment and fibers. If there is a single DWDM system outage or fiber outage, the set of overlay network links dependent on the failed resource would all fail at the same time. The set of links failed by a common resource outage is called a shared risk link group (SRLG). SRLGs have been designed and implemented in many network planning tools and some network routing protocols. A single SRLG represents one potential outage (or failure mode) and a large service provider’s network could easily contain tens of thousands of failure modes. The greater the number of SRLGs, the more difficult it is to attain good performance from planning tools (like routers) whose computations are dependent on the number of SRLGs. For some routing protocols using SRLG information, the situation becomes even worse because a routing protocol may have space constraints to hold a limited number of SRLGs. These issues create a challenge to optimize the SRLG calculations such that the SRLG-related functions are not impacted or the impacts to the SRLG-related functions are limited. This paper takes a closer look at the SRLG optimization issue and proposes algorithms for how to reduce the number of SRLGs for different applications. Index Terms—Shared Risk Link Group, Optimization, Algorithm, Case Study. I. INTRODUCTION service provider overlay networks are built on Most top of optical networks and all optical networks are built over some combination of DWDM equipment and/or fibers. If there is a single DWDM system outage or fiber cut, the set of overlay links routed over the failure risk (the DWDM system or fiber) would fail simultaneously. This set of simultaneously failed links is called a shared risk link group (SRLG). SRLGs are typically represented by identifiers that are associated with all of the links in the set. These IDs are used by network planning tools or network routing protocols to represent the different ways that the network can fail, i.e., the failure modes. The SRLG IDs can also be viewed as a representation of a network’s diversity, since two links that are not simultaneously failed by an SRLG ID can be considered to be diverse under that particular mode of failure. For example, an IP link (part of an overlay network) between two routers may have multiple SRLG IDs associated with it and a single SRLG ID could be shared by multiple IP links. Thus the SRLG information for each IP link describes a list of SRLG IDs to which the link belongs. An SRLG can also represent a potential node outage, such as a total or partial router outage or outages due to router maintenance or software upgrade procedures in the IP network. To manage the total number of SRLG IDs, the lower-layer topology information is often consolidated. For example, for the purpose of restoration planning, a long path of fiber cables that do not bifurcate at intermediate locations can be aggregated into a single SRLG ID (i.e., the path does not encounter a splice location where some fibers are spliced into a different cable or end at an fiber patch panel in a central office). Conversely, for a given SRLG ID, we can list all the IP links that route over that SRLG. Each SRLG ID is unique within a network routing domain. For diverse routing and protection purposes, IGP (interior gateway protocol) routing protocols in both the standardized specification and in commercial products would support the assignment of SRLG information to the network links. IETF RFC 4203 [1] specifies an SRLG as a sub-TLV (type, length, value) of the link TLV. The value is an unordered list of integer SRLG IDs that the link belongs to. > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < Low Rate Private Line Services Mid Rate Private Line Services IP Services Wideband DCS Layer High Rate Private Line Services IP Layer IOS Layer SONET /SDH Ring Layer OTN Layer Layer with automatic restoration Layer without automatic restoration DWDM = Dense Wavelength Division Multiplexing ROADM = Reconfigurable Optical Add/Drop Multiplexer Coherent/ ROADM /Pt-to-pt DWDM Layer DCS = Digital Cross-Connect System IOS = Intelligent Optical Switch Fiber Layer OTN = Optical Transport Network Coherent: More advanced optical network Figure 1: A Sample of Overlay Networks Figure 1 shows a typical service provider’s overly networks. The bottom of the network layers is the fiber network, different DWDM sub-networks, including early point-to-point DWDM sub-networks, large reconfigurable optical add/drop multiplexing (ROADM) sub-networks and recent colorless/directionless coherent optical sub-networks. In order to better use optical network capacity and support relatively lower-speed private line services, service providers may build an electronically-switched optical network layer, including SONET/SDH ring networks, intelligent optical switching networks, or optical transport networks. As more and more services and applications are IP based, the IP network is necessary for a service provider and it can be built on top of the electronically-switched optical network layer and/or on top of the DWDM optical network layer. A link in the IP network is a circuit in the optical networks. Of course there are digital crossconnect networks on top of the electronically-switched optical network to support very low rate services. Thus for all wired services, the data transmission must ultimately go down to the fibers. If a fiber is cut, or an optical device fails, a large number of upperlayer services may go down simultaneously. In order to support protection/restoration at the upper layer overlay networks, we need to know which overlay network links share a common failure risk and which overlay network links are unlikely to fail at the same time. This information is provided by associating SRLG IDs (called bundle IDs in some products) with the overlay network links. In the AT&T Intelligent Optical Switch (IOS) network, the equipment node is the Ciena Core Director. The nodes are connected by “lines” (SONET OC-48s or OC-192s) and multiple lines between the same switch pair are aggregated into “links”. The Ciena OSRP (Optical Switching and Routing Protocol) 2 protocol defines a list of SRLG bundle IDs for each OSRP link [2,3]. From the point of view of the Core Director and its Element Management System, these bundle IDs are simply integers and constraints associated with each link, i.e., they have no actual topological graph model for lower layer networks. Each bundle ID may represent a portion of the underlying fiber path or it may represent some other failure risk (like an intra-office tie cable or some piece of equipment). IEFT RFC 4203 does not specify a maximum length for the list of SRLGs per link, but some commercial products implement maxima. For example, the Ciena OSRP routing protocol enforces a maximum list length for the number of bundle IDs per link [3]. However, in reality, there are links whose SRLGs exceed the maximum number. As mentioned above, if each SRLG represents the smallest unit of an individual fiber span (i.e., cable between two cable splice locations1), the number of SRLG IDs could easily exceed 40,000 IDs in a large carrier’s fiber network. One solution is to combine multiple fiber spans (without bifurcation) into one super fiber span to reduce the number of SRLGs. B C D A E G F (a) Overlay network 3 B A G 1 11 12 16 4 C 2 15 18 10 13 14 19 9 F 5 6 8 24 20 21 7 23 22 D 17 E (a) Fiber network Figure 2: Overlay Network with Fiber Spans Figure 2 shows an example of one simple overlay 1This fiber span definition is over-simplified to make it clear. A fiber span can encompass multiple fiber cables and may not have splice points at its ends. Two cables travelling diversely could converge in a man-hole cover (no splice) and travel together for some distance before separating again. In such a case, a single fiber span would contain both cables for the distance that they travel together. That is, fiber spans are really defined in terms of physical proximity of multiple cables. > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < network with associated fiber spans. In this example, link AB routes over span 1, span2, span3; link BC over span4; link CD routes over span5, span6; link AD routes over span1, span12, span11, span10, span9, span8, and span7; link AG routes over span1, span12, span11, span10, span13, span14, span15, and span16; link DE routes over span17; link GE routes over span16, span15, span14, span13, span9, span24, span23; link EF routes over span20, span21, span22; link GF routes over span16, span18, span19. There are a total of 24 spans. If each span is one SRLG, there are total 24 SRLGs. Both link AD and link GE have 7 SRLGs and link AG has 8 SRLGs. If we combine multiple spans into one super span, we could combine span2, span3 into span2-3; span5, span6 into span5-6; span10, span11, span12 into span10-11-12; span 7, span8 into span7-8; span13, span14, span15 into span13-14-15; span18, span19 into span18-19; and span20, span21, span22 into span20-21-22, span23, span24 into span23-24. After combination, we can reduce SRLG number from 24 into 13 without impacting the information carried by the SRLGs. The number of SRLGs per link is also reduced, for example, link AD has only 4 SRLGs, and link AG only has 4 SRLGs. SRLG has been proposed as a fundamental concept for failure management in networks [4]. In general, SRLG information is maintained manually by the network operator with the knowledge of the physical fiber routes of the network. In some cases, SRLG auto discovery schemes can also be used [5]. SRLG is an important component in survivable network design. In order to provide the failure-independent protection to a customer demand from any single SRLG failure, a pair of SRLG-disjoint paths should be determined. That is, the links along service path must not share any common SRLGs with the links along the restoration path. According to research work [6], finding a pair of SRLG disjoint paths is an NP-complete problem. Thus some network management tools may need to rely on integer linear programing to find SRLG diverse paths. During the past decade, almost all of research papers related on SRLGs are focusing on either restoration capacity planning [7,8,9], or distributed signaling [10,11], or SRLG diverse path algorithms [12,13,14], or trap avoidance algorithm [15]. To the best of our knowledge, we have not seen any research work focusing on how to optimize SRLG size in a network, which is our objective of this paper [16]. The paper is organized as follows: In section II, we describe several SRLG-related functions. Our objective is to reduce the number of SRLGs without impacting those SRLG-related functions. Then we present SRLG optimization algorithms in section III. Section IV proves the correctness of our algorithm and discusses why our optimization does not impact the SRLG-related functions. In section V, we present a few case studies of implementations optimization algorithms in network applications. Section VI is our conclusion. 3 of SRLG operation II. SRLG FUNCTIONS SRLGs were invented for diverse routing, including protection and restoration. If two paths do not share a common SRLG, these two paths won’t fail simultaneously by a single SRLG failure. In practical network management, we classify SRLG functions into two classes—completely diverse routing and maximally diverse routing. Completely Diverse Routing is useful in a number of different situations. Following are a few examples. In a network with IGP supporting SRLG information, each node has the view of the entire network, including the list of SRLG IDs in each link. Then each node or the element management system is able to provide a Fast Reroute Computation. In link-based MPLS FRR (multi-protocol label switching fast reroute), each backup LSP (label switched path) is a list of links that are diversely routed from a given link. In node-based FRR, a backup LSP is also a list of links, but further depends on the next hop of the primary LSP at each node along its path (the backup LSP skips the next node). Each node in the LSP is required to create an SRLG diverse backup LSP to its next-next hop node except the last two nodes. The second last node is required to create a SRLG diverse backup LSP to the last node. During any outage, the right upstream node detects the outage and it switches the LSP traffic to the SRLG-diverse backup LSP immediately. In any single-SRLG outage, such a scheme provides the fastest recovery to the failed LSPs. Fully-diverse routing is also used in end-to-end Protection/Restoration Path Computations. In end-toend protection/restoration schemes, each source node or the element management system needs to find one service path and one protection/restoration path for each demand, and the two end-to-end paths have to be SRLG diverse. A Diverse Routes Computation may be necessary when a customer wants to create his own overlay network by provisioning several mutual diverse LSPs. Since finding two SRLG diverse paths is NP-complete problem, the network will use either a heuristic algorithm or integer linear programming to find the diverse routing paths. Either way, we do not want any two paths to share any SRLGs. Fullydiverse routing plays a useful role in Maximum Restoration Capacity Calculations. To enable rapid restoration, some intelligent networks pre-calculate the restoration path for each service path and store the restoration path in the source node of the service path [3]. Once an outage occurs, the source node > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < detects (or is notified of) the outage and starts the restoration process using the pre-calculated SRLGdiverse restoration path. To reduce required restoration capacity, the network is usually designed to consider only single SRLG outages and thus restoration paths could share restoration capacity over non-simultaneous SRLG outages. This is called shared mesh restoration [10]. Network management can use the fully-diverse restoration paths to calculate the maximum restoration capacity required on each link for any single SRLG failure. Maximal Diverse Routing is an option when it is impossible to find completely diverse routing paths for a demand set in the network. In maximally-diverse routing, a shared SRLG penalty is minimized, where each SRLG is associated with some penalty number. For example, each fiber span has distance and the distance could be the penalty number when routes overlap on that span. Each overlapping node may also be associated with some positive penalty number. III. OPTIMIZATION ALGORITHMS One may define any potential outage as one SRLG, such as a city, a building, a switch or cross-connect component, a conduit, a fiber span, etc. Although there are SRLGs in upper layers, such as router common equipment outages, the most common SRLG represents the potential outage of some sort of fiber spans, or spans for simplicity. As mentioned previously, in this case we need to explore methods to consolidate the size of the SRLG set for large carriers, but without affecting the network restoration metrics or network SRLG functions. In this paper, we mainly consider how to optimize SRLGs from spans only and the optimization algorithms provided in this paper can be easily extended to other type of SRLGs. Our span-based SRLGs have following properties: (1) each span is included in only one SRLG; (2) two links sharing no SRLG are span diverse (not necessarily node diverse); (3) two links sharing a SRLG are not span diverse, and they share all of the spans in the shared SRLG; (4) SRLGs are the smallest possible set of quantities that fully represent the diversity of overlay network; (5) spans comprising a given SRLG are not required to be geographically adjacent, which is more general than the situation in Figure 2. The diagram in Figure 3 illustrates the general relationship among links of an overlay network (such as the IP layer or intelligent network layer), SRLGs, and spans. Each circle represents an overlay link and contains the spans over which it routes. By examining the areas of overlap, for the purposes of restoration calculation of this overlay network, we could group the 16 fiber spans into 7 SRLGs. For example, the link associated with the red bubble route (the top 4 circle) over SRLGs 1, 2, 3 and 4, the blue route (the right circle) over SRLGs, 3, 4, 6, and 7, the green route (the left circle) over SRLGs, 2, 3, 5, and 6. Figure 3: Example of Grouping Spans into SRLGs Thus, for network capacity design, we independently consider the failure of each SRLG and how rerouting is accomplished. For example, the outage of SRLG-2 represents an outage of either fiberspan 6 or 10. Next, we will formally describe how to combine spans2 into SRLGs for a specific network G(V,E), where V is the set of overlay network nodes and E is the set of overlay network links (which we will refer to simply as links). Assume we know the specific fiber routes of each link. Then each link l has a list of fiber spans that it routes over, denoted as Fl , and each fiber span f has a list of links that route over it (called the dependent set of links), denoted as Lf. Now we take a close look at following cases: The dependent set of fiber span x equals the dependent set of another fiber span y, Lx = Ly: in this case, we can combine these two fiber spans into one single SRLG because if two upper-layer links are diverse from one fiber span x, the same two upper-layer links must also be diverse from another fiber span y. Actually we can combine all fiber spans with the same dependent set into one single SRLG. This is exactly what we showed in Figure 3. The dependent set of fiber span x is a subset of the dependent set of fiber span y: if two upper-layer links do not share span y, the same two upperlayer links must not share span x. For the first three SRLG related functions in section II of complete diverse routing, we can drop fiber span x and keep span y as one SRLG. However for maximal diverse routing, we need to keep both 2 We use span consolidation as one example. Other failure modes can be consolidated similarly. > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < span x and span y and cannot drop the SRLG associated with span x. For example, in Figure 3, for complete diverse routing, we may only need to maintain SRLG 3 per link and drop all other SRLGs. However for maximal diverse routing, we need to maintain all SRLGs for a single link, shared by two links, and shared by all three links, i.e., all the 7 SRLGs as well as the SRLG penalty associated with each SRLG overlap, say the total fiber span distance of each SRLG. The above observations can be easily cast into two algorithms to quickly compute a reduced set of SLRGs for each network link that are adequate to accurately perform the complete diverse routing and maximal diverse routing functions. Input: network G(V,E) and span route of each link Output: SRLG info for each link Algorithm 1: for complete diverse routing [1] read network [2] find the list of spans for each link Fl [3] for each span f, find the list of links over it Lf and set span[f].remove = 0 [4] for i=0 to n-1, where n is the number of spans [5] for j=i+1 to n [6] if Li = Lj , mark span[j].removed = 1 [7] else if Li Lj , mark span[i].removed = 1 [8] else if Lj Li , mark span[j].removed = 1 [9] end j [10] end i [11] for each span f, associate SRLG[sf] with span f if span[f].remove = 0 [12] for each link l, define reduced SRLG set Sl = [13] for each span f in Fl [14] if span[f].remove = 0, add SRLG[sf] to Sl [15] end for each span f [16] end for each link l [17] for each link l, report SRLG info of Sl. Algorithm 2: for maximal diverse routing [1] read network [2] find the list of spans for each link Fl [3] for each span f, find the list of links over it Lf, record span[f].length, and set span[f].remove = 0 [4] for i=0 to n-1, where n is the number of spans [5] for j=i+1 to n if Li = Lj , mark span[j].removed = 1, and record span[i].length += span[j].length [6] end j [7] end i [8] for each span f, associate SRLG[sf] with span f if span[f].remove = 0 and set SRLG[sf].penalty = span[f].length [9] for each link l, define reduced SRLG set Sl = [10] for each span f in Fl 5 [11] if span[f].remove = 0, add SRLG[sf] to Sl [12] end for each span f [13] end for each link l [14] for each link l, report SRLG info of Sl. IV. ALGORITHM CORRECTNESS In above SRLG optimization algorithm 1, we mark following two types of fiber spans as removed: (1) if two spans have exactly the same set of dependent links, we combine them together and leave only one span to represent them. Basically the network separates the set of spans into different groups based on their dependent links. Each group is assigned one SRLG ID. This idea is illustrated in Figure 2 and Figure 3; (2) if the dependent links of one group is a subset of the dependent links of another group, we can drop the first group and only keep the second group. The reason is that when two upper-layer links are diverse, they must be diverse on all SRLG groups, i.e., they do not share any common SRLG. If two upper-layer links are diverse on the second group, they must be diverse on the first group. So we are safe to drop the first group for completely-diverse routing. In Figure 3, SRLG1 only supports link1, SRLG2 supports both link1 and link2, while SRLG3 supports link1, link2, and link3. In this case, we can drop SRLG1 and SRLG2, and only keep SRLG3. Similarly we can also drop SRLG4, SRLG5, SRLG6, and SRLG7 without losing essential information about link diversity. Thus it is easy to verify the algorithm correctness for completely-diverse routing functions listed in section II. Next we look at the maximal restoration capacity calculation function. In a shared mesh restoration scheme [4], we define a matrix called failneed[s][l], where s is the SRLG index and l is the link index. Matrix failneed[s][l] maintains the restoration capacity needed in link l if SRLG s fails. The maximal restoration capacity is defined as: R[l] = maxs failneed[s][l] over all SRLGs. For any two SRLG s1 and s2, assume Ls1 Ls2. If we can prove failneed[s1][l] failneed[s2][l] for any link l, then we are safe to drop SRLG s1 without impacting the calculation of R[l]. For any outage s and a set of paths P, we define Ps = {pP, p∩Ls ≠}, i.e., the subset of paths routed over SRLG s. For any path pP, we define Kp as the set of links of p, Cp as the capacity of p, and p* as the pre-calculated completely-diverse restoration path of p. We further define Vs = {l: lp*, pPs}. Then we have failneed[s][l]= p Ps, l p* Cp, i.e, failneed[s][l] is the sum of the bandwidth of all demands failed by s whose restoration paths use link l. Assume Ls1 Ls2, then for any p Ps1, p∩Ls1 ≠, we have p∩Ls2 ≠. So we have Ps1 Ps2. Thus for any l Vs1, we have failneed[s1][l] failneed[s2][l]. For any lVs1, > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < failneed[s1]l]=0. So for any link l, we have failneed[s1][l] failneed[s2][l] when Ls1 Ls2. According to the definition of maximal restoration capacity calculation formula, we can drop SRLG s1 without impacting the maximal restoration capacity calculation for any link. Algorithm 1 can be used for completely-diverse routing, but cannot be used for maximally-diverse routing. For example, if we apply algorithm 1 to the network of Figure 2, we have link AB with SRLG1; link BC with SRLG4; link CD with SRLG5; link AD with SRLG1, SRLG9; link AG with SRLG1, SRLG16; link GE with SRLG16, SRLG9; link DE with SRLG17; link GF with SRLG16; and link EF with SRLG20. If we need to find two maximally-diverse routes between A and G, the solution could be AG, and AD-DE-GE, which only share common SRLG1 and SRLG16. In fact, these two routes also share SRLG10-11-12 and SRLG13-14-15. These two routes are not maximal diverse routes. On the other hand, if we apply algorithm 2 to the same network, we only combine spans with same set of dependent set and accumulate the combined distance. In this way, we do not lose any failure related information. The maximal diverse routes between A and G will be route AG and route AB-BC-CD-DE-EF-FG, which only share common SRLG1 and SRLG16. V. CASE STUDIES We have used the SRLG optimization processes described above in many of AT&T’s internal management tools. In this case study section, we describe a DWDM network planning tool with completely-diverse routing and maximally-diverse routing using SRLG optimization. We also describe a process based on SRLG optimization to drop extra SRLGs due to protocol limitations for the intelligent optical network. AT&T’s DWDM layer is a highly heterogeneous network, including DWDM systems from early pointto-point DWDM systems to recent reconfigurable optical add/drop multiplexing (ROADM) systems; from 2.5Gbps per wavelength systems to 40Gbps per wavelength systems; from optical switching to electronic switching; from wavelength routing to demultiplexing sub-channel routing; et al. In order to help planners make cost-effective service provisioning decisions, we built a web-based interactive tool with integrated optimization algorithms to route high speed circuits over heterogeneous networks through a graphic user interface. The tool provides visualization of the heterogeneous DWDM network and the status of its network links (sub-network type, distance, maximum and available wavelengths number, available 10G and 40G channels, deployment date, 6 and other related information). It can show the whole network or any particular set of sub-networks filtered with the sub-network types and/or the supported maximum speeds. It can also predict and highlight hot-spot links, i.e., DWDM links that will be exhausted by a specified future date based on historical circuit requests and wavelength usage information. The tool also has the capability to generate fullydiverse and maximally-diverse routes for groups of circuits. Since finding two SRLG-diverse routes is NPhard [6], we developed a built-in integer linear programming (ILP) model to find multiple diverse routes, including complete diverse routing and maximal diverse routing. In order to speed up the ILP processing time, we implemented algorithm 1 for complete diverse routing and algorithm 2 for maximal diverse routing. Working experience shows that our optimized ILP can complete most task computations within 1 minute and provide visualization of several candidate paths for interactive navigation. The tool has been deployed since 2009 and is actively used by network planners for their daily circuit planning tasks. AT&T has a large intelligent optical switched network [3]. The Ciena Core Director defines a list of SRLG IDs (known as bundles in Ciena’s terminology) for each link with a limited maximal number of bundle IDs. This is typically less than the number of fiber spans needed to describe the link's diversity. If the number of real bundle IDs is larger than the maximal number, the list of bundle IDs must be truncated. In this case, which bundle ID should be dropped becomes a critical question and bundle ID optimization is required. Bundles are re-computed periodically to keep current with ongoing changes in the network link and fiber span data. When changes are required, an attempt is made to minimize changes to existing bundle IDs. In the intelligent optical switched network, we are required to provide both complete diverse routing and maximal diverse routing if complete diverse routing is unavailable. Thus we cannot completely drop bundles consisting of subset link groups (group i if Li Lj). Instead we use a numeric number to measure the importance of each bundle related to the nature of the link overlaps and its total mileage. We consider three factors for each bundle: (1) the mileage associated with the bundle; (2) the number of simultaneous outages that the bundle represents; (3) whether the links failed by the bundle are a subset of those failed by another bundle. When the number of bundles exceeds the maximum allowed, bundles having lesser importance are dropped until the desired number is achieved. The following formula is used to rank the > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < importance of every bundle based on the above three criteria: Importance(bundle) = 1000000*isSuperSet(bundle) + 10000*linkCount(bundle) + mileage(bundle) where isSuperSet(bundle) = 1 if the links failed by the bundle are not a sub-set of any other bundle failure; it is zero otherwise. linkCount(bundle) = number of links simultaneously failed by the bundle, mileage(bundle) = the sum of the mileages of the fiber spans in the bundle. This solution basically combines algorithm 1 and algorithm 2 together and provides similar results to complete diverse routing and maximal diverse routing. VI. CONCLUSION In this paper, we studied the SRLG optimization issue in detail. After considering the relationship and importance of individual SRLGs, we proposed algorithms on how to reduce the number of SRLGs for completely-diverse routing and maximally-diverse routing. These SRLG-reduction strategies are useful for improving the performance of network planning tools that depend on SRLG information and in selecting which SRLGs are most important when network management systems impose a limit. REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] K. Kompella ed., “OSPF extensions in support of Generalized Multi-Protocol Lable Switching(GMPLS)”, RFC 4203 B. Ramamurthy et al, “CoreDirector CI system description”, http://groups.geni.net/geni/wiki/Ciena%20Core%20Director%2 0switch%20component%20manager%20interface. B. Cortez, “The emerging intelligent optical network: now a reality”, OFC 2002, WH1. G. Ellinas et al., “Routing and restoration architectures in mesh optical networks,” Opt. Network Mag., vol. 4, no. 1, pp. 91–106, 2003. P. Sebos et al., “Effectiveness of shared risk link group auto-discovery in optical networks,” in OFC’02, 2002, p. Th05. J. Hu, “Diverse routing in optical mesh networks,” IEEE Trans. Commun., vol. 51, pp. 489–494, 2003. J. Doucette and W. D. Grover, “Capacity design studies of spanrestorable mesh transport networks with shared-risk link group (SRLG) effects,” in Proc. SPIE Opticomm 2002, vol. 4874, 2002, pp. 25–38. Y. Liu, D. Tipper, and P. Siripongwutikorn, “Approximating optimal spare capacity allocation by successive survivable routing,” in Proc. INFOCOM’01, 2001, pp. 699–708. G. Li, B. Doverspike, and C. Kalmanek, “Fiber span failure protection in mesh optical networks,” Opt. Networks Mag., vol. 3, no. 3, May/June 2002. G. Li et al, “Efficient distributed restoration path selection for shared mesh restoration”, IEEE/ACM ToN, 11(5), October 2003, pages 761-771. C. Qiao and D. Xu, “Distributed partial information management (DPIM) schemes for survivable networks—Part I,” in Proc. INFOCOM’ 02, June 2002, pp. 302–311. R. Bhandari, Survivable Networks: Algorithms for Diverse Routing. Norwell, MA: Kluwer, 1999. K. Lee and K. Siu, “An algorithmic framework for protection switching WDM networks,” in NFOEC’01, July 2001, pp. 402–410. 7 [14] M. Kodialam and T. V. Lakshman, “Dynamic routing of bandwidth guaranteed tunnels with restoration,” in Proc. INFOCOM’00, 2000, pp. 902–911. [15] D. Xu et al, “Trap avoidance and protection schemes in networks with shared risk link groups,” Journal of lightwave technology, vol 21, no 11, Nov 2003. [16] G. Li et al, “On shared risk link group optimization,” OFC 2012, WH2. Guangzhi Li (M’00) is a principle member of technical staff researcher at AT&T labsresearch. He got his PhD and MS in computer science from the College of William and Mary. His research interests include IP-based control plane for optical networks, restoration schemes and algorithms, network simulation and performance evaluation as well as network related applications. He holds about 30 patents and has published more than 90 research papers in journals and conferences. Dongmei Wang (M’00) received her PhD in physics from the college of William and Mary and joined AT&T research lab at 2000. During the past 12 years, she has worked on areas of network design and optimization. Recently she is leading an important IPAG network design for mobility access. She holds about 20 patents and has published about 60 journal and conference papers. Timothy Gallivan is a Senior Network Engineer at AT&T. He obtained a PhD in physics from the University of Texas, Austin TX USA (1989) and an MS in Electrical Engineering from the Georgia Institute of Technology, Atlanta GA USA (1984). Robert Doverspike received his undergraduate degree from the University of Colorado and Masters and Ph.D. degrees from Rensselaer Polytechnic Institute (RPI). He began his career with Bell Labs and, upon divestiture of the Bell System, went to Bellcore (now Telcordia). Later, he returned to AT&T Labs (Research) where he is now Executive Director of Network Evolution Research. Dr. Doverspike has made extensive contributions to the field of optimization of multi-layered transmission and switching networks and pioneered the concept of packet transport in metro and long distance networks. He also > REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < pioneered work in spearheading the deployment of new architectures for transport and IP networks, network restoration, and integrated network management of IP-over-optical-layer networks. He has over 1500 citations to his books and articles over diverse areas/publications such as Telecommunications, Optical Networking, Mathematical Programming, IEEE Magazine, IEEE Communications Society, Operations Research, Applied Probability, and Network Management. Dr. Doverspike holds many professional leadership positions and awards, such as INFORMS Fellow, IEEE Fellow, member of Optical Society of America (OSA), co-founder of the INFORMS Technical Section on Telecommunications, OFC (Optical Fiber Communications) Network Technologies and Applications Committee, DRCN (Design of Reliable Communications Networks) Steering Committee, and Associate Editor of the Journal of Heuristics. 8