due to fierce competition in the telecom industry, traditional. 1+1 Automatic Protection Switching is still the most widespread answer to the issue of ensuring high ...
Comparative Study of Fully Pre-Cross-Connected Protection Architectures for Transparent Optical Networks Aden Grue1,2, Wayne D. Grover1,2, Fellow, IEEE, Matthieu Clouqueur4, Dominic A. Schupke4, Member, IEEE, John Doucette1,3, Brian Forst1, 2, Student Member, IEEE, Diane Onguetou1, 2, Dimitri Baloukov1, 2 1 2
TRLabs, 7th floor, 9107 - 116St NW, Edmonton, Alberta, Canada T6G 2V4 Department of Electrical and Computer Engineering, University of Alberta 3 Department of Mechanical Engineering, University of Alberta 4 Nokia Siemens Networks GmbH & Co. KG, Munich, Germany
Abstract—Protection architectures that have the property of precross-connection are advantageous to the implementation of transparent optical paths. A pre-cross-connected protection path can be in a known-good working condition before use, whereas on-the-fly assembly of a protection path through transparent concatenation of optical channels may not rapidly satisfy the optical path integrity objective. In this study, several pre-crossconnected architectures were compared on the basis of spare capacity cost for 100% single failure restorability, the dual failure restorability of these designs, and the ability of each architecture to support a feasible wavelength assignment and limits on the maximum length of transparent optical paths. The architectures considered were p-cycles, failure independent pathprotecting (FIPP) p-cycles, demand-wise shared protection (DSP), pre-cross-connected trails (PXTs), span-protecting p-trees, and path-protecting p-trees. The results give insight into the relative merits and demerits of these architectures. Keywords-PXT; p-cycle; FIPP; p-tree; demand-wise shared protection; pre-cross-connected; pre-connected; transparent optical networks
I. INTRODUCTION While operators are pressed to reduce their network costs due to fierce competition in the telecom industry, traditional 1+1 Automatic Protection Switching is still the most widespread answer to the issue of ensuring high service availability. Significant cost savings could be obtained by allowing the sharing of protection capacity, but this is commonly believed to be detrimental to availability. The present study, whose first results are presented here, investigates new pre-cross-connected protection schemes from the viewpoint of their ability to reduce network costs while maintaining a high level of availability through the continued guarantee of full single-failure protection. Beyond that, it evaluates their ability to protect against dual-failures, a feature needed to enable future classes of service with very high availability. A. Motivation for Full Pre-Cross-Connection Because they reduce processing load on intermediate optical switches, pre-connected survivability schemes have the
potential to offer a speed advantage over schemes in which restoration paths are cross-connected on-demand, after the failure. But more importantly, they offer the possibility of simplifying the engineering of protection paths, which may be pre-tested and therefore known to be in an already stable working condition prior to their use. This assurance of transmission integrity may not be available if the backup lightpaths are only assembled dynamically after the failure, in which case it may take some time for adaptive transmission techniques to bring the new end-to-end path into alignment and obtain the very low bit error rate (BER) usually required. Indeed, polarization, dispersion, transient power level variations, and several other noise and nonlinear impairment processes must all be precisely engineered end-to-end to be under the dispersion or loss budget for a link in order for a light signal to be reliably transmitted end to end. In these circumstances, the full pre-connection of protection paths may not be only desirable; it may be necessary to ensure their timely activation. B. Study Methodology The study consisted of a sequence of investigations into the properties of multiple architectures, each study focused on a different property. The four main investigations focused on the creation of low cost, single-failure protected network designs, followed by the analysis of the dual failure restorability of these designs, the problem of making wavelength assignments to the optical paths in the designs, and finally the process of altering the designs to comply with real-world limitations on optical path lengths (where necessary). Throughout the description of the experiments and results the reader may find various special cases and exceptions that disrupt the uniformity of the study and dilute the value of a straight comparison between the experimental results. This observation should be tempered with the understanding that the ultimate goal of the study was not so much to obtain a rigorous theoretical comparison of the different architectural concepts as it was to gain experience in the design of restorable networks using these architectures, as well as to gain understanding into their respective strengths and the unique challenges that may be faced by each of them individually.
Page 1 of 8
The most interesting of the studied architectures are those that provide end-to-end protection of entire paths, as opposed to restoration between the endpoints of single failed spans. This is because span protection cannot be said to be “truly” precross-connected, as cross-connections must be changed at the endpoints of failed spans. The span protecting architectures are included as comparative references; p-cycles in particular are well studied and known to be an effective design alternative. II.
As all routes are span-disjoint from one another, any single failure can only damage a single route. Thus when any single span failure occurs, the receiving node signals the transmitting node to perform a switchover operation to recover from the failure. This occurs in tandem for every demand that crosses
a)
ARCHITECTURES UNDER STUDY
Examples of the 6 architectures that were studied are illustrated in Figure 1. A. p-Cycles A p-cycle is a ring-like pre-configured structure of spare capacity used to protect against failures of on-cycle spans (spans that are crossed by the p-cycle) and straddling spans (spans not crossed by the cycle but whose end-nodes are). Upon failure of an on-cycle span, the p-cycle is “broken into” by the restoration mechanism at each end-node of the failed span, and an affected lightpath is rerouted in the other direction around the p-cycle. Therefore, each unit-sized p-cycle (consisting of one pre-cross-connected wavelength over the whole cycle) can protect a single wavelength on each on-cycle span. In the case of a straddling span failure, the restoration mechanism can reroute one wavelength on the failed span in one direction around the p-cycle and a second wavelength in the other direction, so a unit-sized p-cycle can protect two wavelengths on each straddling span. p-Cycle restoration is explained in more detail in numerous other sources, such as [1], [3], and Chapter 10 of [4]. B. Failure Independent Path-Protecting p-Cycles Like regular p-cycles, FIPP p-cycles are cyclical protection structures that are fully pre-planned and pre-cross-connected. Unlike regular p-cycles, FIPP p-cycles protect entire end-toend working paths as opposed to single spans. Paths protected by the same unit capacity FIPP p-cycle must be mutually disjoint. Path protection is achieved as long as these paths are span-disjoint. Naturally, a FIPP p-cycle can provide protection only to paths with end-nodes on the cycle. More background on FIPP p-cycles can be found in [5] and [6]. C. Demand-Wise Shared Protection Demand-wise Shared Protection (DSP) combines principles from both dedicated and shared protection and is similar in operation to M:N Automatic Protection Switching (APS). DSP requires there to be at least two span-disjoint routes between the nodes to be protected. From the set of span-disjoint routes, one is then used as a protection route. The working capacity is then diversified across the remaining disjoint routes, which become working routes. The demand exchanged between the pair of nodes is divided as equally as possible into an integer number of lightpaths on each of these routes. The capacity placed upon the protection route (spare capacity) is then the largest number of lightpaths placed on any single working route between the node pair. This occurs for each set of demands exchanged between nodes. Thus, each demand pair has its own dedicated, pre-connected restoration path.
p-Cycles
b)
FIPP p-Cycles
c)
d)
DSP
PXTs
e)
f)
Span pTrees
Path pTrees
Protection structure
Protection path
Working path
Failed working path
Figure 1. Demonstration of protection actions for considered architectures (a) p-cycles (b) FIPP p-cycles (c) DSP (d) PXTs (e) span p-trees (f) path p-trees
Page 2 of 8
the failed span, and because each demand set is protected separately in the same manner, they are always fully restorable. More information on DSP is given in [7][8][9][10]. D. Pre-Cross-Connected Trails Pre-Cross-Connected Trails (PXTs) are path-protecting non-cyclical pre-connected structures. The concept is functionally identical to FIPP p-cycles except that PXTs are non-closed “trails” instead of cycles. This approach has been advocated primarily in [11]. The design assurance is that a PXT can protect against the failure of its protected working routes caused by the failure of a single span. A PXT may cross nodes and spans more than once. It may protect multiple demands with overlapping working routes and protection paths as long as it is not possible for spare capacity contention on the PXT to occur as a result of a single span failure. A PXT can protect working paths that have end nodes on the PXT, but that are disjoint from the protection path that they would use within the PXT. PXT protection is described in more detail in [11], [12], and [13]. E. Span-Protecting p-Trees A span-protecting p-tree behaves as a protection structure in much the same way as a p-cycle. Each p-tree consists of a number of spare wavelength paths that are pre-connected to each other at the nodes of the network. For the sake of simplicity, it is assumed that a p-tree is not allowed to cross the same node or span more than once. In addition, a p-tree may not contain any pre-connected paths of spare wavelengths that are cyclical (i.e., they are strictly tree-like structures). For the sake of p-tree protection structures, it is assumed that multipledegree pre-connections are possible, regardless of whether or not this theoretical construction has any real-life meaning1.
The pre-connected paths of spare links in the p-tree are used to protect against span failures. Any p-tree is able to protect against the failure of any span that has both its end nodes on the p-tree but is not itself on the p-tree. Unlike pcycles, on-structure spans are not protected. In a standard p-tree network design, the on-tree spans of one p-tree are protected by other, complementary p-trees. To our knowledge, this type of tree protection has not to date been investigated in the literature, although a different type of span-protecting treebased protection scheme has been investigated in [14].
the protection path that is formed must be disjoint from the failed working path. As with span-protecting p-trees, we do not believe this exact concept has yet been used in the literature, although similar path-protecting tree structures have been studied (i.e. [15]). The above-described architectures are summarized in TABLE I, which breaks down the architectures according to the topology of their protection structures and whether they protect individual spans or end-to-end paths. TABLE I. Trails Spanprotecting Pathprotecting
SUMMARY OF THE TEST ARCHITECTURES Cycles
1+1 Principle
p-Cycles PXTs
FIPP p-Cycles
III.
DSP
Trees Span-protecting p-trees Path-protecting p-trees
TEST DATA
A. Network Topology and Demand Data This study used a single network topology and demand data set representing a characteristic European inter-exchange regional network model, shown in Figure 2. The demand pattern is a relatively sparse, with nonzero demand quantities between only 58 out of the 136 total node pairs in the network. Of the nonzero demands, there are 37 with 1 unit, 10 with 2 units, 6 with 3 units, 3 with 4 units, and 2 with 5 units.
IV.
EXPERIMENTS AND RESULTS
A. Minimum Capacity Designs A first question for any restorable architecture is “how much spare capacity is required to achieve basic single failure restorability?” Therefore our first effort was to create a 100% single failure protected design for each architecture, both to support a comparative study of design costs and also to establish initial “reference designs” on which our further investigations would be based.
We created most of the reference designs by solving the corresponding ILP (integer linear programming) representations of the design problems to obtain designs of minimum or near-minimum capacity. Most of our design
F. Path-Protecting p-Trees A path-protecting p-tree is a p-tree that protects entire working paths instead of single spans. As with all pathprotecting structures, a path-protecting p-tree protects a working path by forming a protection path through its spare capacity between the end nodes of the failed path. Obviously this working path must have its end nodes on the p-tree. Also, 1
Strictly speaking, any tree-based scheme cannot be fully pre-cross-connected in the same robust sense that, say, p-cycles are pre-connected before failure. In the outgoing direction at a tree connected (i.e. degree 3 or greater) node, signal splitting can be argued as the means of pre-connection, but on the incoming direction, a selection function is needed to achieve something close to pre-connectedness. We include the architecture for thoroughness, to understand how tree architectures compare with the others, assuming that this technical challenge can be overcome.
Figure 2. Topology of the “Germany” test network
Page 3 of 8
approaches are SCP (spare capacity placement) methods that take a static working capacity routing plan as input. DSP is necessarily treated differently because the entire concept is focused around splitting and routing the working flow in such a way that the sharing of protection capacity becomes efficient. An inherently joint optimization approach (where the choice of working routes and protection routes used in the final design is jointly optimized) is therefore necessary for DSP. For the other architectures, the working routing was always based on shortest path routing with small modifications made to take the specific restrictions of each protection architecture into account. In addition, we considered two interpretations of “shortest path” routing, one being shortest path by total path length and the other being shortest path in terms of the number of hops in the path.
4) PXTs PXT designs were created using the same DRS-based method used for FIPP p-cycles described in [6] (its use for PXTs described in [13]). The only difference is that the candidate structures are trails instead of cycles. We chose to generate designs using 30 DRSs per demand, 30 candidate trails per DRS, and a maximum DRS size of 15 demands, as these values were found experimentally to produce good results. The set of all simple network trails (trails that do not repeat nodes or spans) was used as the candidate trail set. This set contains 13,640 trails.
In some cases, heuristic methods were used to reduce the complexity of our ILP design problems by limiting the size of the input data set, so it is important to note that these designs are “minimum capacity” only in the context of the data set provided, and not in the sense of a global minimum over the entirety of the design space. In practice, however, this distinction is often only one of provable optimality, as our past experience shows that such designs are usually within a few percent of globally optimal solutions.
5) Span p-Trees Span p-tree designs were created using a model that is functionally equivalent to the p-cycle model. The only difference is that the structures used are trees rather than cycles. Unfortunately the set of trees in the network is much too large to include all trees as candidates, so the set of candidate trees was limited to trees with the maximum size of 9 spans (i.e. number of contained spans) and maximum tree node degree of 4 (i.e. maximum number of tree spans incident on any tree node). These constraints produce a candidate set containing 136,690 trees.
Finally, in the PXT case, a fully heuristic-based PXT design method recently proposed in the literature was also used to generate designs for further comparison. 1) p-Cycles The model used to create the p-cycle reference designs is well documented in [4] and other sources. This model was used to solve the network design problem to full optimality (i.e. all 135 cycles were considered as candidate p-cycles). 2) FIPP p-Cycles FIPP p-cycle designs were created using the DRS (disjoint route set) method described in [6]. In brief, this method reduces the complexity of the design problem by only considering for protection a small, randomly-generated selection of sets of disjoint working routes (i.e. DRSs), instead of considering all possible combinations of demands that may be protected by each possible cycle. Designs were generated using 30 DRSs per demand and 2 eligible cycles per DRS, with a maximum DRS size of 11 routes. These parameters are explained in more detail in [6]. 3) DSP As mentioned above, the DSP design process is intrinsically joint. However, because there are no sharing interdependencies between different demands, the optimization can be performed on a demand-by-demand basis only. The ILP model that was used is presented in [17], and is conceptually similar to the one described in [8]. For each demand, the constraints ensure that the demand is split between geographically diverse working routes, and that they share a single disjoint backup route that contains enough capacity to protect against the failure of any single working route. The objective function then minimizes the capacity cost of the DSP configuration.
A design was also created using the heuristic method described in [11] and [12], as an implementation of the algorithm was readily at hand.
6) Path p-Trees Path p-tree designs were created using a DRS-based model that is functionally equivalent to both the PXT and FIPP pcycle models. The tree set was limited in the same way as the tree set for span p-trees, although because the path p-tree model is more complex, the size of the set had to be even more tightly restricted. The model was given the set of all trees with maximum size of 8 spans and maximum degree of 3, a total of 45,997 trees. As for the DRS parameters, we chose to use 20 DRSs per demand, 20 candidate trees per DRS, and a maximum DRS size of 10 demands. 7) Results The total capacity costs (spare plus working) for each of the designs are given in TABLE III. These costs are given as total capacity channel counts (in the case of hop-based routing), and distance-channel counts (in the case of distance-based routing, i.e. if distances are assumed to be in units of km, then costs are in units of channel-km). The counter-intuitive fact that the PXT heuristic performs better than the ILP model can be explained by the fact that the heuristic places fewer restrictions on the types of PXTs that are used than the ILP model does, as explained in [13]. B. Dual Failure Restorability Analysis We next investigated how our 100% single failure restorable designs would respond to dual span failures. Intuitively, designs with lower redundancy might be expected to show lower dual failure restorability, but results will also be affected by innate properties of each individual architecture and by how each architecture handles the limitations that the nodal equipment model places on real-time switching agility.
Page 4 of 8
TABLE III.
CAPACITY COSTS OF THE INITIAL DESIGNS
p-cycles FIPP p-cycles DSP PXTs (ILP-based) PXTs (heuristic) Span p-trees Path p-trees
Under leasthops working routing 286 342 445 375 334 469 437
Under leastdistance working routing 44,198 45,990 63,617 52,849 47,332 69,247 61,467
The tests involved simulating each design’s response to every possible dual failure scenario and recording the total numbers of damaged working routes, as well as the total number of these working routes that could be restored. To calculate the R2 value, we divided the total number of restored routes over all dual failure scenarios by the overall total number of routes affected. For the purposes of this analysis, it is assumed that nodes are only capable of performing a single, pre-determined switching action in the event of a failure. In other words, we assume that nodes have no ability to adaptively respond to multiple failures; whenever a failure is detected on a working path, the same restoration response is attempted regardless of possible prior failures and recovery actions. If this attempt fails due to a failure on the protection path, then the demand simply remains unrestored. Note that, because of this restriction based on limitations to equipment functionality, the R2 values obtained are more conservative than those attainable due to topological limitations alone. To simplify the number of test cases, only the hop-based working routing designs were used for this investigation and continuing forward from this point in the study. 1) Results The dual failure restorability values of the network designs are summarized in TABLE II. The results roughly follow the prediction that designs that use more capacity are more resilient to dual failures. The anomaly with the ILP-based and heuristic PXT designs may be explained by the fact that, although the heuristic design uses less total capacity, the self-crossing of the PXTs may create more available backup paths to use in the case of dual failures.
At this point, as the p-tree designs were the most computationally intensive yet also the least capacity efficient of all the designs, and we felt that their R2 properties were not exceptional enough to justify these costs, they were dropped from further consideration in the study.
TABLE II.
DUAL FAILURE RESTORABILITY RESULTS
p-cycles FIPP p-cycles DSP PXTs (ILP-based) PXTs (heuristic) Span p-trees Path p-trees
R2 Value 65% 53% 86% 67% 77% 81% 83%
C. Wavelength Assignment Having a fully pre-cross-connected capacity plan for a 100% single failure restorable network is useful, but it still does not quite answer the question “how do we implement this design in a real network?” One of the reasons is that, for the network system under study, the cross-connected paths in question are optical paths, which must be assigned carrier wavelengths. A wavelength can only be used by one path on any given fiber, so the wavelength assignment problem is nontrivial. The problem can be rendered easier by making use of wavelength conversion hardware, but in order to fulfill the main objective of developing cost-efficient network designs, it was assumed that it is preferable to avoid the use of such expensive hardware.
Therefore, our next step was to attempt to assign wavelengths to both the pre-connected working paths and the structures (p-cycles/PXTs/etc.) in the network designs, assuming no wavelength conversion capability. The intent was to find solutions that could also be implemented in a single fiber network (i.e., each wavelength used at most once on each span). Our first goal was to do so using fewer than the 40 wavelengths that were assumed to be available on a single fiber in the initial project specification. It was also assumed for the study that the available wavelengths are organized in two bands of 20 wavelengths each. Therefore we also pursued the more ambitious goal of attempting to create assignments using 20 or fewer wavelengths, as to do so would only require the use of a single 20-wavelength band, presumably benefiting system costs considerably [16]. The first step before attempting to perform wavelength assignment was to inspect the initial designs to determine the maximum number of total wavelengths (working and spare combined) that were required on any one span. If this number is greater than 40 (or 20), it is impossible for a wavelength assignment using 40 (20) wavelengths to exist, as not enough wavelengths exist to support the span in question. In the cases where this number was too great, we attempted to modify the design to reduce this number to a level at which a feasible wavelength assignment could exist. In cases where these modifications could be made, we then used simple ILP models to perform wavelength assignment.2 It turns out that, for the reference designs used in this study, the number of capacity units (used wavelengths) per span was always well below 40, and there was no problem in producing corresponding 40-wavelength assignments. Problems were only encountered in some cases when the problem was further restricted to 20 wavelengths. 1) p-Cycles The wavelength assignment problem for p-cycles has an additional twist due to the fact that p-cycles are span protecting instead of end-to-end path protecting. For a span to be protected by a p-cycle, it must break into that p-cycle’s spare capacity and continue transmission on the backup path found 2
Note that in general a maximum span wavelength usage of less than 40/20 is only a necessary, not a sufficient, condition for a 40/20 wavelength assignment to exist. However, this constraint alone proved to be sufficient for all of the test cases encountered in the study.
Page 5 of 8
therein. But if wavelength conversion is not available, transmission must continue on the same wavelength as the original path. Furthermore, p-cycles are pre-connected and thus must use the same wavelength throughout the cycle. Therefore the p-cycle and the path must share wavelengths. This has serious consequences for p-cycles, because if only a single fiber is available, then protection of on-cycle spans becomes impossible, as two copies of the wavelength (one for the path, one for the p-cycle) cannot be provided on the same span. Furthermore, straddling span protection is reduced from 2 protection paths to 1, because a p-cycle cannot transmit on the two different wavelengths that would be required in order to provide two protection paths for two different units of capacity on the same span. For the sake of being able to produce p-cycle reference designs, it was assumed that enough resources were available (either wavelength conversion or 2 fibers on each span) to avoid the above-mentioned problems. This assumption was made for the p-cycle case only, which should be kept in mind when making comparisons with the other results. The p-cycle reference design was found to use up to 22 wavelengths on some spans. Therefore a 20-wavelength assignment could not be computed without modifying the pcycle design. An ILP model was run to find the absolute minimum value for the maximum number of wavelengths required by any single span. This value was found to also be 22. Therefore a 20-wavelength assignment could not be found even if the p-cycle spare capacity arrangement was modified. If such a p-cycle design exists for this network, it must be attained by modifying the working routing plan as well (which had been assumed to this point to be static). Rather than take this approach, we devoted resources to other investigations, and so it remains as future work. 2) FIPP p-Cycles For FIPP p-cycles as well, we found it very difficult to attain the 20 wavelength limit. It was found, using a method similar to that used for p-cycles, that the minimum value that could be attained for any one span was 22 wavelengths, if rearrangement of the protection plan was allowed. Using a joint capacity placement (JCP) model, we were then able to attain a design using at most 20 wavelengths on any one span, as well as a corresponding 20 wavelength assignment. The penalty for achieving this design was both the vastly increased complexity of the JCP model versus the SCP model, as well as a 3% increase in the overall capacity required by the design. 3) DSP The original DSP design contained up to 39 units of capacity on some spans. This is much higher than for other architectures. This is not wholly unexpected, as the overall amount of capacity in the DSP designs is higher, as shown in Section IV.A. It was found through ILP methods that this maximum value could be reduced to 34 without affecting the capacity cost of the design, and further reduced to an absolute minimum of 22 at the cost of a 10% increase in total capacity. As the DSP design problem is a fundamentally joint working and capacity placement problem, this shows that a 20 wavelength solution is infeasible for a 100% single span failure restorable DSP design in this network.
4) PXTs At this point in the study, the heuristically generated PXT design was omitted from consideration, due to the fact that the heuristic produces designs with PXTs that cross the same span several times (this tendency is elaborated upon in [12]). If wavelength conversion is not used, such a PXT cannot be implemented under the single fiber per span assumption, as this would require a single fiber to provide two copies of the same wavelength for the same PXT on the same span, an obvious impossibility. This was identified as another demotivating factor for the use of this heuristic to generate PXT designs, as opposed to more easily controlled DRS-based ILP method.
In comparison to the other architectures, it was relatively easy to obtain a 20 wavelength assignment for the ILPgenerated PXT design. The original design used up to 24 wavelengths on a single span, so a modified design model was run that restricted the number of wavelengths to 20. This model produced a feasible solution that used only 2% more total capacity than the original. A corresponding 20-wavelength assignment was then found using ILP methods. D. Transparent Optical Reach Considerations Our final efforts in this study were focused on addressing the practical reality that, in real transparent optical networks, there is a limit to the reach of an optical signal, as for the channel to remain transparent the signal cannot undergo regeneration at any point. The network can only transmit an optical signal transparently for a certain distance before losses in the fiber and at the node transitions degrade the signal unacceptably. Therefore some pre-connected paths from our original network designs may not be able to function properly in practice if they are too long. The question answered in this section of this study is “do our designs satisfy this network’s limits on optical path length, and if not, how can we modify them so that they do?”
The equation defining the limitations on pre-connected path length due to optical losses is as follows: length + (hops − 1) Deq ≤ Dmax
(1)
In this equation, the length of the optical path is expressed in km of fiber. The term involving hops is included to represent the fact that each node transition is assumed to have insertion losses equivalent to Deq km of fiber. We term the left hand side of equation (1) the “equivalent optically transparent length” of a path, which should not exceed a certain maximum distance Dmax. For this study, we used the values 80 km and 2000 km respectively for Deq and Dmax. Our first step was to characterize the initial reference designs in terms of the equivalent transparent lengths of their paths (both working and protection), to determine which architectures exceeded the 2000 km limit, and to what degree. In cases where this limit was exceeded, we then attempted to modify the designs such that they would both satisfy the limit and also stay as close as possible to the original amounts of spare capacity.
Page 6 of 8
Note that, because our working routing plan is done via shortest paths, and the 2000 km restriction for this network is large enough to exceed the shortest path length between each node pair, all working routes are well within the limit. Therefore it is only the length of the protection paths that may cause problems. 1) p-Cycles The exercise of characterizing path lengths has an interesting twist for p-cycles, again because they protect spans instead of end-to-end paths. Therefore the length of the protection path will depend both on the length of the original working path and the length of the section of the cycle used to protect the failed span. Furthermore, a single span will likely contain many different working links that are protected by several different p-cycles. Therefore, the length of the failure state paths in the network will depend on which p-cycle is assigned for the protection of each working path that crosses over that span. Depending on how intelligently this assignment is made, the 2000 km limit may or may not be satisfied.
Therefore the p-cycle investigation included the additional first step of assigning the p-cycles that provide the longest backup paths for a given span to the shortest working paths that cross that span, and vice-versa, in order to maximize the chances of satisfying the 2000 km length limit. Even following this approach, however, 49% of the protection paths in the network exceeded the 2000 km limit. Therefore, a new ILP model was developed that would generate minimum capacity designs while at the same time not producing any paths that were too long. Using this model, a path length limited design was found at the cost of a 10% increase in spare capacity. 2) FIPP p-Cycles As FIPP p-cycles are end-to-end path protecting structures, no explicit assignment of FIPP p-cycles to protected working paths was required as in the p-cycle experiment. However, a related effect that does apply to path-protecting architectures is overprotection: even in a minimum capacity network design, some working paths may have a choice between several possible protection paths in the installed protection structures. Among these choices, we must avoid those that exceed 2000 km. Therefore, before beginning the analysis, a shortestprotection-path-first assignment of protection structures to working paths was performed. This was also done for PXTs, under which overprotection may also occur.
Of the backup paths in the original design, 45% were found to exceed the 2000 km length limit. Therefore a modification was made to the original ILP to restrict the FIPP p-cycles used in the solution such that none of them would provide protection paths longer than 2000 km. The new design produced by this modified approach had an increase in capacity redundancy of 25% over the original design. 3) DSP One advantage of DSP is that, because the backup paths of different demands do not share capacity, the backup paths do not need to be deviated much from shortest path routing in order to facilitate sharing. Because of this, the original DSP design used no paths with effective lengths greater than 1450km, easily satisfying the optical reach restriction.
4) PXTs The original DRS ILP-based PXT design did not have as much trouble with long protection paths as either p-cycles or FIPP p-cycles; 91% of protection paths were below the 2000 km limit. To force all of the protection paths below this limit, the same approach was taken as with FIPP p-cycles: restricting the PXTs used in the design to those that only provide protection paths shorter than 2000 km. The resulting design, once restricted also to ensure the feasibility of a 20 wavelength assignment, was only 3% more expensive than a 20 wavelength-feasible design without the path length restriction.
The heuristic-based PXT design was again omitted here, as the design heuristic fundamentally “grows” PXTs to protect successive demands, resulting in PXTs that are very long. In this case, the resulting PXTs greatly exceed the 2000 km limit. Also, it is fundamentally much harder to alter the heuristic to control PXT size than it is to simply modify the PXT set used by the ILP model; this is another demotivating factor for the use of the custom heuristic to perform PXT network design. V.
SUMMARY AND CONCLUSION
A. Summary Table of Results The major output of the study was a final set of additional minimum or low capacity designs that, according to our best efforts to date, also satisfied as many constraints of the originally defined problem as possible. These designs are to be carried forward as the reference designs for the next stage of the project. The relevant characteristics of these designs are summarized in TABLE IV. Naturally, these results vary somewhat from the values reported previously, as these designs are refined versions of those that were analyzed above. Once again, the R2 values given are dependent on equipmentspecific limitations, and are more conservative than those that would be attained in an abstracted framework where the only limitations to R2 are topological. B. Concluding Comments The results comprise a multi-faceted comparison of important characteristics of the p-cycle, FIPP p-cycle, DSP, PXT, and span/path p-tree architectures on a representative European network model. Instead of studying a breadth of networks under one or two central objectives, our philosophy was to learn by studying a single network in depth, considering the problem in ever increasing levels of detail, leading finally to comparable optically transparent designs that are as near as possible to being "actually buildable" with low-cost WDM add-drop multiplexers and terminal equipment.
The p-tree architectures were removed from the study early on because of the high inefficiency of their designs and the questionable feasibility of their implementation. Among the remaining four architectures, each has strengths and weaknesses that can usually be intuitively understood. DSP is conceptually the simplest, but is quite inefficient, although its R2 value benefits from the excess spare capacity. PXTs are more complex, both conceptually and computationally, but improve on efficiency. PXTs must also be constrained to avoid span self-looping, otherwise fiber counts must be multiplied to
Page 7 of 8
TABLE IV.
CHARACTERIZATION OF FINALIZED ARCHITECTURE DESIGNS
Architecture
Notes
DSP
• DSP does not distinguish explicitly between working and spare capacity • Redundancy in this case is standard redundancy, assuming shortest path working routing of 166 units • Wavelength feasibility and reach feasibility not considered • Wavelength feasibility and reach feasibility not considered • R2 calculated for pure min-capacity (non-wavelength/reach feasible) design and static nodal model; probable underestimate of best achievable R2
Span-Protecting p-Trees Path-Protecting p-Trees PXTs
p-Cycles FIPP p-Cycles
• R2 calculated for pure min-capacity (non-wavelength/reach feasible) design and static nodal model; probable underestimate of best achievable R2
Length Limit Satisfied?
20-λ Design?
Yes
No
Not considered Not considered
Not considered Not considered
Yes
[4]
[5]
Redundancy
R2
445
168%
86%
469
183%
81%
166
271
437
163%
83%
Yes
166
220
386
133%
67%
Yes
No
166
151
317
91%
66%
Yes
No
166
231
397
139%
51%
REFERENCES
[3]
Total
303
Evidently, a single best architecture cannot be identified. Rather, the benefits of the study lie in the insights gained into network design using these architectures. The design methods that were developed are valuable because they address important real-world details, such as wavelength continuity, functional limitations of equipment, and optical reach limits. Future work will look at further characteristics, such as restorability against node failures, support for multiple quality of protection schemes, and heuristics or other techniques for solving the more difficult design problems.
[2]
Spare
166
permit wavelength continuity. Span-protecting p-cycles present an easy design problem, and are conceptually and operationally easy to understand. They easily reach the highest spare capacity efficiency, but are not path-protecting structures. Finally, our FIPP p-Cycles designs have efficiency similar to PXTs, but lower R2 and a design problem that is still quite challenging. The lower R2 is attributable in part to limitations on the agility of the failure switching response, imposed by the static, pre-planned nature of the nodal equipment model.
[1]
Working
W. D. Grover, D. Stamatelakis, “Bridging the ring-mesh dichotomy with p-cycles,” Proc. Second International Workshop on the Design of Reliable Communication Networks (DRCN 2000), Munich, Germany, 912 Apr. 2000, pp. 92-104. C. G. Gruber, D. A. Schupke, “Capacity-efficient planning of resilient networks with p-cycles,” Proc. Tenth International Telecommunication Network Strategy and Planning Symposium (Networks 2002), Munich, Germany, 23-27 Jun. 2002. D. A. Schupke, W. D. Grover, M. Clouqueur, “Strategies for enhanced dual failure restorability with static or reconfigurable p-cycle networks,” Proc. IEEE International Conference on Communications (ICC) 2004, 20-24 Jun. 2004, pp. 1628-1633. W. Grover, Mesh-Based Survivable Networks: Options and Strategies for Optical, MPLS, SONET and ATM Networking. Upper Saddle River, NJ: Prentice-Hall PTR, 2003. A. Kodian, W.D. Grover, “Failure-independent path-protecting p-cycles: efficient and simple fully preconnected optical-path protection,” Journal of Lightwave Technology, Vol. 23, Iss. 10, Oct. 2005 pp. 3241 - 3259.
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16] [17]
J. Doucette, A. Kodian, W.D. Grover, “A Disjoint Route-Sets Approach to Design of Failure-Independent Path-Protecting p-Cycle Networks,” 5th International Workshop on Design of Reliable Communication Networks (DRCN) 2005, Naples, Italy, 16 Oct.-19 Oct. 2005. A. M. C. A. Koster, A. Zymolka, M. Jager, R. Hulsermann, “Demandwise Shared Protection for Meshed Optical Networks,” Journal of Network and Systems Management, vol. 13, no. 1, pp. 35-55, March 2005. R. Wessäly, S. Orlowski, A. Zymolka, A. M. C. A. Koster, and C. G. Gruber, “Demand-wise Shared Protection revisited: A new model for survivable network design,” Proc. International Network Optimization Conference (INOC 2005), Lisbon, Portugal, March 2005, pp. 100-105. C. G. Gruber, A. M. C. A. Koster, S. Orlowski, R. Wessäly, and A. Zymolka, “A Computational Study for Demand-wise Shared Protection,” Proc. Design of Reliable Communication Networks (DRCN 2005), Naples, Italy, October 2005, pp. 421-428. R. Hülsermann, M. Jäger, A. M. C. A. Koster, S. Orlowski, R. Wessäly, and A. Zymolka, “Availability and cost based evaluation of demandwise shared protection,” Proceedings 7th ITG-Workshop on Photonic Networks, Leipzig, Germany, 2006, pp. 161–168. T. Y. Chow, F. Chudak, A. M. Ffrench, "Fast Optical Layer Mesh Protection Using Pre-Cross-Connected Trails," IEEE/ACM Trans. Netw., vol. 12, no. 3, pp. 539-547, June 2004. A. Grue, W. D. Grover, "Characterization of pre-cross-connected trails for optical mesh network protection," J. Opt. Netw., vol. 5, no. 6, pp. 493-508, June 2006. A. Grue, W. D. Grover, "Improved method for survivable network design based on pre-cross-connected trails," J. Opt. Netw., vol. 6, no. 2, pp. 200-216, February 2007. S. Shah-Heydri and O. Yang, "Hierarchical protection tree scheme for failure recovery in mesh networks," J. Photon. Netw. Commun. 7, 145159 (2004). A. Groebbens, D. Colle, S. D. Maesschalck, I. Lievens, M. Pickavet, and P. Demeester, "Efficient protection in MPλS networks using backup trees: part one--concepts and heuristics," J. Photon. Netw. Commun. 6, 191-206 (2003). A. L. Chiu, G. Li, and D.-M. Hwang, “New problems on wavelength assignment in ULH networks,” Proc. OFC-NFOEC 2006, March 2006. B. Forst, W. D. Grover, “Factors Affecting the Efficiency of Demandwise Shared Protection,” to appear in Proc. Design of Reliable Communication Networks (DRCN 2007), La Rochelle, France, October 2007.
Page 8 of 8
