1. IntroductionDense wavelength division multiplexing (DWDM) transmission technology allows anincreasing number of wavelengths (data channels) to be multiplexed in a single opti-cal fiber, each channel transporting a huge amount of traffic. Moreover, in transportnetworks most of the carried traffic must be provided with a guaranteed high avail-ability. The way to improve availability in optical transport networks is by means ofrecovery mechanisms. Recovery is provided by either protection or restoration mecha-nisms. The former is based on the substitution of a failed resource [e.g., a link or alabel-switched path (LSP)] with a preassigned backup resource; the latter is based onrerouting the LSP using spare capacity. Protection (or spare) resources can be eitherdedicated, in which case the spare resource is reserved to a single working LSP, orshared, in which case the same spare resource may be used to provide protection tomultiple working LSPs.
Because shared-path protection (SPP) provides better resource utilization thandedicated-path protection (DPP), it has been widely studied in the literature [1,2]. Theworking routes of two connections sharing spare resources must be link-disjoint and,as a consequence, SPP has been traditionally applied to mesh networks. In [1] theproblem of computing working and protection routes under SPP constraints for a con-nection request was proved to be NP-complete. In [2] a SPP routing algorithm withshared-risk link groups (SRLGs) disjoint protection for mesh networks is presented.In this paper we propose a method to implement SPP on dynamic optical rings, wherethe routing and wavelength assignment (RWA) problem can be computed in polyno-mial time. Moreover, we assume that the optical topology has been designed at thenetwork planning phase in a way that no common infrastructure (optical cables, con-duits, etc.) is used by any two links in the network. Therefore, SRLGs are not used inthis paper.
Legacy Synchronous Optical Networking (SONET)/Synchronous Digital Hierarchy(SDH) metropolitan transport networks were designed as ring-based networks due totheir inherently fast protection switching capabilities and because they provide highconnection availability. However, mesh-based networks have been extensively used in
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 131
packet-based networks due to their high efficiency and flexibility. However, with theintroduction of the p-cycles concept [3], fast protection is also possible in mesh net-works. The p-cycles concept can be applied to a wide range of technologies—such asWDM, SONET/SDH, or IP–multiprotocol label switching (MPLS) networks—and pro-tection schemes, such as path and link protection [4]. The aim of this paper is to pro-vide a protection mechanism for automatically switched optical network (ASON)–generalized MPLS (GMPLS) optical rings [5,6], allowing the migration of existingring-based metropolitan transport networks to the ASON–GMPLS paradigm.
An ASON [5] is an optical transport network that has dynamic connection set-up–tear-down capability. This functionality is accomplished by means of a control planethat carries out, among others, routing and signaling functions. GMPLS provides asuitable control plane for dynamic optical networks [6]. It includes the traffic engi-neering (TE) extensions of resource reservation protocol (RSVP-TE) [7] for signalingand of open shortest path first (OSPF-TE) [8] for intradomain routing. Moreover, theuse of DWDM technology implies a very large number of parallel links (data links)between two adjacent nodes. For scalability purposes, multiple data links can be com-bined to form a single TE link, and the link management protocol (LMP) [9] has beenspecified to manage this issue. The efficiency of SPP can be improved by supportingextra traffic. In this case protection resources are used to transport this extra trafficunder normal conditions, and it will be preempted in case of failure.
Protection mechanisms can work both at the link layer [optical multiplex section(OMS)], and at the path (LSP) layer. Protecting at the OMS layer allows the completebundle of multiplexed optical channels in a fiber to be recovered with only one protec-tion action. In previous work we designed the GMPLS automatic protection switching(GAPS) mechanism [10], which is based on extensions to the GMPLS LMP [9]. TheGAPS mechanism, which is based on the well-known OMS shared protection [alsoknown as two-fiber bidirectional line-switched ring–multiplex section-shared protec-tion ring (2F-BLSR–MS-SPRing) in the SONET/SDH] scheme, is designed to provideservice recovery within 50 ms after fault detection, even in large GMPLS-controlledrings. The OMS shared scheme is deployed over two-fiber bidirectional rings, wherethe total capacity of each fiber is divided in two sets of wavelengths: one set isreserved to transport working channels, whereas the other set is used to transportprotection channels. When a link failure occurs, the two optical nodes adjacent to thefailure loop back the bundle of working channels on the protection channels in theopposite direction, being that all the extra traffic is preempted. Similarly, we proposeto implement SPP on rings also dividing the fiber capacity into two sets of wave-lengths. However, contrary to OMS shared protection, when a link failure occurs, onlythe affected LSPs are protected, thus preempting only a part of the extra traffic.
The authors of [11] propose a protection scheme that records the unavailability timeof paths. When a failed path approaches its service level agreement (SLA) limit, it canpreempt other paths to avoid violating the SLA. This scheme requires memory forpath state information. Note that in distributed controlled networks, as is the case forASON–GMPLS networks, this per-path memory requirement involves additional pro-cessing and memory requirements and the design of signaling extensions. In ourapproach, however, paths are assigned to a class of service. Paths belonging to pro-tected classes can preempt paths belonging to the lower preemptable class of service.As such, no path state information is needed, thus making the overhead introduced byour approach much lower than that of [11].
The remainder of the paper is organized as follows: Section 2 describes our proposalto implement SPP with extra traffic in GMPLS-based optical rings. Experimentalresults, obtained over the ASON–GMPLS CARISMA network test bed [12], are shown.In Section 3 we design an optical add–drop multiplexer (OADM) to support SPP and,based on this design, we model the protection time as a function of the switching time.In Section 4 a mechanism to provide protection times under 50 ms is presented. Itchooses the layer (OMS or LSP) in which the protection will be performed as a func-tion of the number of LSPs to protect. Finally, in Section 5 we draw the main conclu-sions of this work.
2. Shared-Path Protection with Extra Traffic in ASON–GMPLS RingsSPP has usually been proposed to be implemented over mesh networks, as resourcesharing for protection LSPs is only performed among link-disjoint working LSPs. Fig-
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 132
ure 1 shows two shared-protected (SP) connections where the working LSPs, w1 andw2, are strictly link-disjoint and the protection LSPs, p1 and p2, share the link 3–4.
We propose to implement SPP in GMPLS-based DWDM rings. In this case, two dif-ferent wavelengths are used: one to support working and the other to support protec-tion LSPs. Figure 2(a) shows two dedicated-protected (DP) connections, 5–7 and 1–4,whose working LSPs do not overlap. If working and protection LSPs for both connec-tions were routed using the same wavelength, SPP could not be implemented in ringssince the protection LSP for one connection would use resources already allocated forthe working LSP of the second connection. However, when the working LSPs arerouted using a common wavelength, �i in Fig. 2(b), protection LSPs can shareresources in wavelength �k. Note that protection LSPs are sharing two commonresources: links 1–7 and 4–5. Figure 2(b) also suggests that two additional SP connec-tions, 1–7 and 4–5, could be allocated, increasing the resource usage ratio. Generaliz-ing this, all the SP working LSPs using �i share �k for the protection LSP.
Let us denote W as the number of wavelengths available in each link of the ring. Tounivocally determine which wavelength has to be used for the route of a protectionLSP, we propose to split the set of wavelengths into two bands: wavelengths in the setSPWL= �1, . . . ,W /2� support working LSPs, whereas wavelengths in BEWL= �W /2+1, . . . ,W� support protection LSPs. Therefore, if the routing algorithm chooses �i forthe working LSP of a connection, then the protection LSP will be routed through�W−i+1.
In the absence of failures, protection resources can be used to transport best-effort(BE) extra traffic, which will be preempted in case of failure. As resources in �W−i+1have been assigned to protect resources in �i, once the working LSP of a SP connec-tion is routed using �i, all resources (not only those in the protection LSP of that path)in �W−i+1 are available to be used for BE traffic.
The GMPLS recovery framework [13,14] specifies three protection schemes: 1+1dedicated protection, 1:N �N� =1� LSP protection with extra traffic, and preplannedrerouting without extra traffic. In the 1:N scheme, N working LSPs (having the sameorigin and destination) are protected by one LSP. In the preplanned LSP rerouting,two disjoint LSPs are established between the end nodes: the working and the protec-tion LSP. The working LSP is implemented in the transport plane, while the resourcesof the protection LSP are only prereserved in the control plane, and, therefore, anexplicit signaling is required to instantiate them in the transport plane. This gives theopportunity of reusing the protection reserved resources to accommodate extra trafficwithout the constraint of sharing the same origin and destination nodes.
In our implementation the SPP scheme uses the preplanned LSP rerouting. Beingthe working LSP of a SP connection affected by a link failure, the protection LSP issignaled and activated in the transport plane. Taking advantage of its shared natureand that the protection LSPs are created in a failure-driven way, it is possible to reusethe protection capacity to transport extra traffic. In the event of a link failure, theextra traffic can be preempted to accommodate the working traffic to be protected.
Several approaches to transport extra traffic can be implemented using the pre-planned LSP rerouting: the first approach consists of only transport extra traffic overthe resources reserved for protection LSPs of already-established SP connections. Forexample, when connection 1 is established in the ring in Fig. 2(b), only resources usedby the protection LSP �p1� could be used for extra traffic. When the SP connection istorn down, the extra traffic must also be torn down. Moreover, as several SP connec-tions share protection resources, specific per-resource reserve counters must be main-tained to know whether a resource can be used for extra traffic. Therefore, the over-head introduced by this approach appears to be too high.
Fig. 1. Shared-path protection.
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 133
The second approach that can be considered consists of reserving all resources inthe whole ring in wavelength �W−i+1 to accommodate extra traffic, when �i transportsat least one working LSP. As an example of this, when connection 1 is established inthe ring in Fig. 2(b), the working LSP �w1� is established in the transport plane usingresources in �i. Resources of the protection LSP �p1� in �k �k=W− i+1� are reserved inthe control plane to be used in case of failure of w1, and thus all resources in �k can beused for extra traffic. At the time when no resources in �i are used for SP workingLSPs, all extra traffic in �k must be torn down. Note that in this approach, only spe-cific per wavelength reserve counters need to be maintained. The overhead introducedby this approach is much lower since only W /2 reserve counters are needed.
Finally, considering that resources in BEWL are all reserved for protection LSPs,all these resources could be used to transport extra traffic. In such a case, extra traf-fic would be completely decoupled from SP traffic; the extra traffic would use all thoseresources, and reserve counters would not be required. Therefore this approach doesnot introduce any overhead.
In this paper, we implement both the second (per-wavelength) and third (full-band)approaches. A network without wavelength conversion that is able to deal with twoclasses of service (the SP and the BE preemptable services) is deployed, and the block-ing probabilities obtained with both approaches (and with DPP) are compared.
2.A. RWA with Extra TrafficEvery optical node connection controller (OCC) in the GMPLS control plane floods thestate of the local outgoing data links using OSPF-TE opaque link state advertise-ments (OLSAs) [8,15] to its control plane neighbor OCCs. OLSAs include the unre-served bandwidth sub-type-length-value (sub-TLV), which specifies the amount ofbandwidth not reserved yet in every class of traffic. As stated above, in this work twoclasses of traffic are considered (SP and BE). OLSA flooding is performed every time adata link is used by a LSP or is released.
Being the network deployed without wavelength converters, its topology can bedescribed using a different graph for each wavelength. Let us denote Gi�V ,Ei� as thegraph describing the resources at the wavelength i. The set of graphs is dynamicallyupdated to represent the current allocation state of the data links in the network.Moreover, we use an additional graph G�V ,E�, which represents the physical networktopology independently of the allocation state of the resources. Graph G is used tocompute the disjoint route, allowing reserving resources currently used for extra traf-fic.
As previously described, each SP connection consists of two disjoint LSPs, sup-ported by two different wavelengths. We have developed the precomputed RWA (PC-RWA) algorithm to solve the RWA problem. First, we calculate the distance betweenorigin and destination nodes over the graph G. This is the minimum distance. Then,we search for a wavelength i providing this minimum. If there is no wavelength pro-viding the overall minimum distance, the wavelength with the minimum distanceamong all wavelengths is chosen. This search is performed with wavelengths in SPWLin increasing order. The working route is then computed over the graph G . Second,
Fig. 2. Dedicated protection and shared-path protection in optical rings.
i
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 134
we search for a disjoint route in G and we move that route to GW−i+1. We use thegraph G for the protection route because some resources in GW−i+1 may be allocatedfor extra traffic. Table 1 shows the pseudocode for the PC-RWA algorithm.
Figure 3 shows an example of three graphs representing a network: graph G repre-sents the physical topology of the network, whereas G1 and G2 represent the currentstatus of the resources at wavelengths 1 and 2, respectively. The minimum distancebetween nodes 6 and 3 is 3 hops, as can be observed in graph G. The distance betweenthose nodes is 4 in G1 and 3 in G2. Note that our PC-RWA algorithm will compute theroute over G2 (3 hops), while the well-known first-fit (FF) heuristic [16] would com-pute the route over G1 (4 hops).
The performance of the FF heuristic and the PC-RWA algorithm have been experi-mentally evaluated over the ASON–GMPLS CARISMA network test bed [12]. TheCARISMA GMPLS control plane uses the RSVP-TE protocol for signaling, theOSPF-TE protocol for routing and the LMP protocol for control channel managementand link property correlation. The OCCs have been implemented using Linux-basedrouters. Each pair of OCCs communicates through a single IP control channel imple-mented with full duplex fast Ethernet links. Finally, each OCC communicates withthe local optical cross-connect (OXC) through the connection controller interface(CCI).
Unprotected connection requests arrive at each OCC according to a Poisson processwith a mean interarrival time �iat�. The connection’s holding time is exponentiallydistributed with a predefined mean �ht�. The offered traffic can then be calculated asE=ht / iat.
Figure 4 compares the performance of the FF heuristic against the PC-RWA interms of blocking probability as a function of the offered traffic for a five-node ringwith 20 wavelengths per link (C band). The offered traffic ranges from 1 to 10Erlangs/node. The blocking probability for both algorithms is negligible when the
Table 1. PC-RWA Algorithm
Procedure PC-RWA (IN Node source, destination)begin
Route w, pShortestRouteSP (source, destination, w)If length �w�= =0 then
No route found;Look for single route p link-disjoint with w in G
Move p to wavelength W-wavelength �w�+1Use route w for the working LSP androute p for the protection LSP
end
Procedure ShortestRouteSP (IN Node source, destination; OUT Route r�begin
distance=get distance from source to destination in GminDistance=INFINITEminWL=0For each wavelength i in SPWL do
If Gi is not updated thenUpdate source’s shortest path tree in Gi
distanceWL=get distance from source to destination in Gi
If �distanceWL�minDistance� thenminDistance=distanceWLminWL= iend IfIf �minDistance=distance� then
Break loopend ForIf minWL�0 then
Create the route r from source to destination in GminWL
End
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 135
offered traffic is low. However, when the offered traffic increases, the blocking prob-ability for the FF heuristic becomes slightly higher than for the PC-RWA algorithm.However, the aim of the PC-RWA algorithm is to compute disjoint paths when theresources used by the protection route may be in use by BE traffic.
The PC-RWA algorithm is accelerated by having the shortest path tree precomputedfor every wavelength graph. On the reception of OLSAs updating or deleting datalinks, each OCC recalculates the shortest path tree in the corresponding wavelengthgraph. As a consequence of a LSP signaling (set-up or tear-down) a group of OLSAs isgenerated and, thus, a short time of convergence is needed. Therefore the shortestpath computation is not performed immediately after the reception of a single OLSA,but it is delayed a short time, allowing a group of OLSAs for a hypothetical LSP toarrive. If a request arrives and the shortest path tree is not updated in any graph Gi,the shortest path tree is computed at that time. Note that, if the shortest path treesare updated at the time of computing a disjoint pair of routes, the computational com-plexity of the PC-RWA algorithm is given by the computational complexity of themodified Dijkstra algorithm, O��V� · log�V�� [17], which is used to compute the protec-tion route. Therefore, the PC-RWA algorithm provides constant computation timesindependently of the wavelength assigned.
The shortest route for a BE connection is computed in a similar way as workingroutes for SP connections. In this case, the search is performed within BEWL.
2.B. SPP with Extra-Traffic ImplementationIn the per-wavelength approach, data links belonging to wavelengths in SPWL areadvertised as free and with available bandwidth for SP traffic. However, data links inBEWL are also advertised as free but with no available bandwidth for any traffic. This
Fig. 3. Example of set of three graphs representing a network.
Fig. 4. Performance of PC-WRA versus FF.
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 136
allows the routing algorithm in the OCCs to know the network topology at everywavelength, while it prevents those resources from being used to route BE traffic.
When a new SP connection is signaled, the origin OCC sends two RSVP-TE Pathmessages: one for the working LSP and another for the protection LSP. The workingLSP is signaled with label �i, whereas the protection LSP is signaled with �W−i+1.When the Path messages arrive at the destination node, Resv messages are originatedand the resources are allocated for the working LSP and reserved for the protectionLSP. At this point, all OCCs in the ring know that local resources in �W−i+1 can beused for extra traffic and are advertised with available bandwidth for the BE classthrough OLSAs to all nodes in the ring.
To properly manage whether resources in �W−i+1 can be used for extra traffic, weuse a per wavelength reserve counter. Every SP connection establishment will incre-ment that counter for local resources in �W−i+1. On the contrary, the counter will bedecremented in the tear-down process. When a SP connection is torn down, the originOCC sends RSVP-TE Path-Tear messages for the working and protection LSPs. If thisSP connection is the last one using �i, BE traffic using �W−i+1 must be deallocated.Upon the reception of a Path-Tear message for a working LSP in �i or a protectionLSP in �W−i+1, each OCC decrements the reserve counter for resources in �W−i+1.When this counter is 0 and the resource is free, it will be advertised with no availablebandwidth, preventing this way to be further used for BE traffic. If the resource isallocated, a notification will be sent to the origin OCC containing the LSP’s sessionID. Upon the reception of this message, the origin OCC should send a tear-downrequest for that BE-class connection and notify the client through the user networkinterface (UNI) [18]. At the end of this process all resources in �W−i+1 will be releasedand advertised with no available bandwidth for BE traffic.
Note that the per-wavelength approach introduces an overhead in the GMPLS con-trol plane. On the one hand, a RSVP-TE notification is needed when protectionresources are used for BE and the per-wavelength reserve counter is 0. In such a case,a tear-down RSVP-TE signaling for those LSPs is also needed. However, consideringthat these forced tear-downs are very infrequent with respect to the number of normalset-ups and tear-downs, the introduced overhead is negligible. On the other hand, anadditional overhead is introduced in the routing process since OSPF-TE OLSAs needto be originated to advertise the data links in a �k with available bandwidth for theBE class. Note in this regard that the PC-RWA algorithm must recalculate the short-est path tree upon the arrival of a group of OLSAs. However, since the shortest pathtree is computed upon the reception of the OLSAs, and not when a new route needs tobe computed, its performance is not affected. No additional OLSAs are generatedwhen a reserve counter is decreased to 0. Therefore, we can conclude that the per-wavelength introduced overhead is limited to the OSPF-TE advertisement producedwhen the first SP connection using resources in a �i is set up. The number of OLSAsgenerated in such a case is 2*n, with n standing for the number of nodes in the ring.
In the full-band approach, all data links are advertised as free and with availablebandwidth. The PC-RWA algorithm looks for available resources in the set of wave-lengths corresponding to the class of service requested. Since BE traffic can use allresources in BEWL, no reserve counters are needed, and no additional flooding forresources in BEWL is needed. As a consequence, the above discussion about the intro-duced overhead is not applied to this approach.
2.C. Performance EvaluationFor comparison reasons, we have also implemented the DPP scheme with extra traf-fic, using the 1:N �N=1� LSP protection with extra traffic, specified in [13,14]. In thisscheme, the working and the protection LSPs are signaled and effectively activated inthe transport plane. When a link failure affects the working LSP of a DP connection,the nodes adjacent to the failure send RSVP-TE Notify messages to the end nodes.Upon the reception of a Notify message the end nodes switch the traffic on the work-ing LSP to the protection LSP and the extra traffic is preempted.
Several differences between SPP and DPP can thus be found: (1) as a consequenceof the fact that only the end nodes have to switch traffic on the event of a failure, DPPprovides faster protection times than SPP; (2) working and protection LSPs are routedusing the same wavelength in the dedicated scheme, whereas different wavelengthsare used in the shared scheme; (3) in the DPP scheme the end points of the extra traf-
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 137
fic are predetermined by the protection LSP—in the SPP scheme extra traffic and pro-tected traffic are decoupled; (4) due to its shared nature, SPP provides better resourceusage than DPP. The performance of DPP and SPP with extra traffic (both the per-wavelength and the full-band approaches) has been experimentally evaluated over theASON–GMPLS CARISMA network test bed previously described.
Figure 5 shows the blocking probability as a function of the offered traffic for a five-node ring with 40 wavelengths per link (C band). The offered traffic ranges from 1 to10 Erlangs/node of SP (or DP) and BE traffic. The results are plotted with its 95% con-fidence interval, so that the accuracy of the results can be appreciated.
The blocking probability for DP and SP traffic is negligible when the offered trafficis low. However, when the offered traffic increases, the blocking probability for DPtraffic becomes higher than for the SP traffic. Note that in the DP scheme one pro-tected connection uses all resources in a wavelength. In the case of SPP, although onlyhalf of the wavelengths are used for the working LSPs, it is possible that more thantwo working LSPs are being simultaneously transported in one wavelength. Thisclearly shows the advantage of the SPP scheme.
Moreover, the SPP scheme (even the per-wavelength approach) provides higher flex-ibility in the use of spare resources to transport best-effort traffic. Served BE traffic inthe per-wavelength approach of the SPP scheme is conditioned by the availability ofresources in the wavelengths used for protection; served BE traffic in the DPP schemeis conditioned by the availability of protected connections with the same origin–destination nodes. This flexibility provided by the SPP scheme marks the difference interms of blocking probability. Therefore, when the offered traffic is low, the BE trafficpresents high blocking probability in both schemes. However, the DPP scheme pro-vides the worst blocking probability since it is highly improbable to find protected con-nections with the same origin–destination nodes as the BE connection requested,which are not being used for extra traffic. When the offered traffic is high, the block-ing probability decreases to be closer to the corresponding protected (SP or DP) traffic,as more resources are available to be used for extra traffic, following the upward ten-dency of their protected traffic under highly loaded traffic.
The blocking probability obtained with BE traffic under the per-wavelength SPPapproach, although much lower than with DPP, is still high. However, when consider-ing the full-band SPP approach, the blocking probability for the BE traffic is coinci-dent with the SP blocking probability, since the same resources are available to beused for both traffic classes. The remainder of the paper will focus on the full-bandapproach since it shows clear advantages with respect to the per-wavelengthapproach.
2.D. Availability ComparisonIn this subsection we analyze the availability provided by SPP for SP and BE classesof service, compared with that of the DPP. Availability is the probability that a systemwill be found in the operating state at a random time in the future. Steady-state avail-ability can be expressed as [4]
Fig. 5. Blocking probability against SP traffic load.
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 138
A =UpTime
UpTime + DownTime�
MTTF
MTTF + MTTR, �1�
where the mean time to failure (MTTF) represents the expected time to the next fail-ure of the network component, following completion of the repair; the mean time torepair (MTTR) is the expected time needed to repair the network component. Theprobabilistic complement of the availability A is the unavailability �U�: U=1−A.
For purposes of availability analysis, let us consider only link failures. Note thatthis is an accurate approach since the system components with the highest failurerate are the optical cables [4]. According to [19,20] we use MTTF=311 FITs/km andMTTR=12 h, where 1 FIT represents one failure in 109 hours.
We define the following sets of links: let R be the set of all links in the ring and letP be the set of links transporting a particular LSP. The availability of a SP connectionis given by the union of two disjoint groups of events; namely, (1) all links i in theworking LSP are available and (2) one link in the working LSP is unavailable, whilethe links in the protection LSP are available and can be used for protection. In thepresent study we consider the links in the ring as being mutually failure independent.Thus, the availability can be expressed as
ASP = �∀i�P
Ai + �∀i�P
Ui �∀j�R\P
Aj. �2�
Note that Eq. (2) models the availability for a protected connection without anyresource sharing (i.e., DPP). Let us analyze the behavior of SPP in a multiple failurescenario. As an example, let us consider the ring in Fig. 2(b). Three cases can be dis-tinguished: (1) Two links transporting one working LSP fail simultaneously; the con-nection will be protected using the protection LSP. For example, if links 5–6 and 6–7fail simultaneously, connection 1 continues working using the protection LSP p1. Con-nection 2 is not affected by the failure. (2) Two links, one transporting the workingLSP and the other transporting the protection LSP of the same protected connection,fail simultaneously; the path is cut. For example, if links 5–6 and 1–7 fail simulta-neously, connection 1 is cut. Connection 2 is not affected by the failure. (3) Exactly thesame as (2), but the link transporting the protection LSP also transports the workingLSP of another SP connection; both connections are cut. For example, if links 5–6 and2–3 fail simultaneously, connections 1 and 2 are cut. This is exactly the same behav-ior as DPP where both connections have been routed using different wavelengths as inFig. 2(a). Therefore, we can conclude that SPP in rings presents the same availabilityas dedicated protection, showing a better resource usage ratio.
On the other hand, a BE-class connection will be available when all links i in P areavailable, but will be preempted when one link not in P is unavailable while the restof the links not in P are available and can be used for the protection of a SP-class con-nection. This is equivalent to saying that a BE-class connection will be available whenall links i in the ring are available:
ABE = �∀i�R
Ai. �3�
Note that, in the case of DPP, the BE-class traffic is torn down when the associatedprotected connection is torn down. These forced tear-downs provide additionalunavailability to the BE class under the DPP scheme. In the full-band SPP scheme noforced tear-downs are needed, and thus Eq. (3) can be used to compute the BE-classavailability.
Figure 6 shows the unavailability for the longest possible routes for both connectionclasses as a function of the number of nodes in the ring �n�. The unprotected connec-tion availability is also plotted for reference. We assume that the average link length�L� is 30 km. We can observe how preemption over BE-class connections leads tounavailability values, which double those of unprotected connections. Obviously, theSP class presents the best availability, which is provided by the protection.
3. ROADM Design and Protection Time ModelTo support SPP with extra traffic, we have designed the OADM shown in Fig. 7. Thebasic components are splitters–couplers (S) and wavelength-selective switches (WSS).
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 139
The incoming optical signal in the east and west ports can either pass through or bedropped to any port. The local traffic can be added either to the east or to the westoutgoing signals. Note that additional hardware is required to monitor the incomingoptical power in order to detect failure conditions.
In the case of a link failure, the adjacent OADMs, once they have detected the lossof light (LoL), send notice of the failure to the OCCs in the GMPLS control plane.Then, for each LSP to be protected, the OCCs send notice of the failure to the OCC ofthe closest end node (origin or destination) from the failure by sending a RSVP-TENotify message. The address of the node to be notified was received in theNOTIFY�REQ object in the RSVP-TE Path/Resv message. When the source OCCreceives the Notify message, the signaling of the protection LSP starts. It consists ofsending a Path message to eliminate the extra traffic from the resources required bythe protection LSP and sending Path/Resv messages to effectively activate the protec-tion LSP. One command to create–eliminate a connection in the OADM (Fig. 7)implies a set of commands to the WSSs, which will be sequentially processed in caseof simultaneous requests.
Let us define the protection time �tSPP� in an n-node ring with preplanned rerout-ing, as the interval from the failure detection to the completion of the switching opera-tion (for each single connection to be protected). Let us denote tCCI as the communica-tion time between the OADM and the OCC, tOCC as the time to process a single
Fig. 6. Connection unavailability.
Fig. 7. OADM design to support SPP with extra traffic.
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 140
RSVP-TE message, tlink as the link propagation delay, and tswitch as the OADM switch-ing time. We determine the expression of tSPP considering two extreme cases for theLSPs to be protected: (a) the origin and destination nodes are adjacent to the failure(hereafter adjacent LSP); (b) the LSP origin node is adjacent to the failure, while thedestination node is the farthest one (maximum number of hops). Note that the OCCsadjacent to the failure are notified by their associated OADMs after a tCCI interval.
Figure 8 presents an example of the signaling involved after a failure. In thisexample a five-node ring is considered, where four LSPs need to be protected after afailure in the link C–D. The LSP C–D is an adjacent LSP. One of the end nodes of theLSPs A–C and F–C is adjacent to the failure. The LSP E–B is a pass-through LSP. InFig. 8 the length of each bar is proportional to its processing time. Finally, the left-hand side dotted line below every node represents activity in the control plane whilethe right-hand side dotted line represents activity in the optical node.
For the adjacent LSPs, the RSVP-TE signaling has to travel from one OCC to theadjacent one using the opposite side of the ring. Each OCC has to process the Pathand Resv messages and send configuration messages to its OADM to perform theswitch. Let us denote ra as the number of adjacent LSPs to be protected. To obtain theaverage case, let us assume that half of these LSPs have their origin in one of theadjacent nodes, while the rest have their origin node in the other node adjacent to thefailure. Hence, �ra /2� connections will be serially processed on each of the adjacentOADMs.
For the LSPs with one end in one node adjacent to the failure, the RSVP-TE signal-ing messages travel from the origin OCC to the destination OCC through the oppositeside of the ring. Let us denote r as the total number of LSPs to be protected.
All LSPs, independently from their destination nodes, have their protection routethrough the nodes, which are �n /2�−1 hops distant from the failure adjacent nodes.Therefore, those OADMs have to perform r connections in a serial basis. Dependingon the tswitch and r values, the effect of the serial processing of connections can beslower than the propagation delay around the ring. We can express the time to protectas
The first two terms are the time needed for the detection of the failure and to sendthe first switching command from the control to the transport plane �2tCCI�, and theprocess time in the OCC adjacent to the failure. The max{ } function captures themaximum of two terms: the time to protect the adjacent LSPs around the ring, andthe time to protect all the affected traffic due to the coincidence of multiple connec-tions in some nodes.
The performance of SPP with extra traffic in terms of protection time has beenexperimentally evaluated over the ASON–GMPLS CARISMA network test bed [12].In our implementation, we have obtained the following times: tCCI=1 ms, tOCC=0.5 ms.
Fig. 8. SPP with extra traffic time model.
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 141
Figure 9 shows the time needed to protect 20 LSPs (we assume links with 40 wave-lengths), as a function of tswitch, for different link lengths �L� and number of nodes �n�.As shown, when tswitch is low, propagation and control plane processing times aredominant on the protection time. However, when tswitch increases, the connections’sserial processing is the dominant effect. Thus, to protect 20 LSPs we need tswitch to belower than 1.8 ms, even for small rings (where propagation time becomes negligible).
Figure 10 shows the protection time as a function of the number of LSPs to protect,assuming a medium-size long-haul 15 node ring with 100 km links, and for severaltswitch values. Note that tswitch includes the physical WSS switching time and the timeto process a request in the OADM. Currently available WSSs provide physical switch-ing time close to 2 ms, and the latest technology will provide commercial componentswith a submillisecond physical WSS switching time [21,22].
As shown, the higher is tswitch the smaller is the number of LSPs that can be pro-tected within 50 ms after the failure detection. For example, with tswitch=4 ms it ispossible to protect only nine LSPs within 50 ms.
However, telecom equipment is usually based on cards, where one card representsthe interface with the control plane, and another card includes the WSS component.Cards are interconnected through buses. In the card holding the WSS component, onespecific command has to be generated. In our implementation we have measured thetime from the reception of the command from the control plane to when the WSS com-mand is generated as about 1.5 ms [10]. Thus, assuming switching times of currentlyavailable WSS, the actual OADM switching time is about 3.5 ms. As a conclusion, thenumber of LSPs that can be protected within 50 ms is limited to 11.
On the basis of the previous experimental results, two alternatives can be consid-ered. On the one hand, classes of protection can be defined as a function of the protec-tion time (Table 2) to be guaranteed. This is similar to those defined in [23] for a mesh
Fig. 9. Protection times against switching time.
Fig. 10. Protection times against the number of LSPs to protect.
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 142
network. Due to the switching time restriction, we can dedicate up to 11 wavelengthsto the SP-50 class, 9 wavelengths to the SP-100 class, and 20 wavelengths to the BEclass.
Additionally, the number of sub-50-ms protected connections can be increased byusing OMS protection. This is evaluated in Section 4.
4. Shared-Path Protection and OMS Shared ProtectionThe objective of the OMS schemes is to provide protection times under 50 ms to thecomplete set of protected LSPs assuming OADM switching times of 3.5 ms. In [10] wepresented the GAPS mechanism providing the needed OMS protection coordination atthe GMPLS control plane. OMS shared protection also allows the protection resourcesto be used to transport extra traffic. However, conversely to SPP, in OMS shared pro-tection, all extra traffic is preempted when recovering the protected traffic after a linkfailure.
OMS shared protection and SPP schemes can coexist in the same ring. In the sameway of our proposal for SPP, OMS shared protection divides the total capacity of eachfiber in two wavebands: the working and the protection wavebands. Therefore, wave-bands can be used to support SPP and OMS shared protection. To support the SPPscheme with extra traffic, in Section 3 we designed the OADM block depicted in Fig.7. To also support the OMS shared protection scheme we have added to that OADMblock a set of optical switches, splitters, couplers, and band splitters (BSs) and theadditional hardware to monitor the incoming optical power (Fig. 11).
Using the designed OADM supporting both levels of protection, we propose a real-time mechanism to decide which protection scheme is the most appropriate to applyupon the detection of a failure, based on the number of currently established LSPs �r�to be restored: if r�11, then SPP can be performed, keeping the protection time under50 ms, while maximizing the total traffic transported by the ring, as some preempt-
Table 2. Classes of Protection
Service Description Protection Time
SP-50 (Shared-path) Protected service �50 msSP-100 (Shared-path) Protected service 50–100 ms
BE Best-effort service. Preemptable when resourcesare needed for protection of the SP-x class.
Repair time
Fig. 11. OADM design supporting both SPP and OMS SP, with extra traffic.
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 143
able LSPs can continue working; on the contrary, if r�11 the protection time cannotbe kept under the limit, so OMS shared protection is performed instead although allextra traffic will be preempted.
To illustrate the impact of the protection method chosen over the spare resources,let us assume uniformly distributed traffic to be transported over the ring. We assumealso that the shortest route is always used for the working LSP. In the seven-nodering example of Fig. 12, a particular link (e.g., 3–4) can transport working LSPs of 1…�n /2� (1, 2, and 3) hops long. Depending on the end nodes, several distinct LSPs for thesame hop count may exist as shown in Fig. 12. In the case of failure, we can calculatethe average number of spare data links �dl� that will be used to protect a single work-ing LSP as the product of the number of hops used for protection and the probabilityof each distinct LSP:
dl =�i=1
�n/2��n − i�i
�i=1
�n/2�i
. �5�
In our example the average number of data links used for the protection of one LSPwhen SPP is applied is 4.67 data links/LSP. Using OMS shared protection, the num-ber of data links used for the protection of the link 3–4 is �7−1��20=120, which cor-responds to the whole network spare capacity.
Figure 13 shows the number of spare data links that are preempted, on average, asa function of the number of protected connections for different ring sizes �n�. Whenthe number of LSPs to protect is low �r�11�, SPP is performed; the spare capacityused grows linearly with the number of LSPs. When the number of LSPs to protect isgreater than the threshold �r�11�, OMS shared protection has to be applied in orderto keep the protecting time within 50 ms; in this case, the whole spare capacity is
Fig. 12. All distinct LSPs routed through link 3–4.
Fig. 13. Spare capacity used �dl� against the number of protected connections �r�. When r�11,SPP is applied, or else OMS SP is applied.
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 144
used. Figure 13 also shows, in percentage, the evolution of the spare data links thatcontinue working after the protection is performed. The right Y axis is used in thiscase.
To maximize the amount of extra traffic that is not preempted after a failure, theorder to fill the working and protection bands should be the one shown in Fig. 14. Theworking band should start filling data links in an ascending way. Recall that, if therouting algorithm chooses �i for the working LSP of a connection, then the protectionLSP will be routed through �W−i+1. Thus, in the protection band, the lower is the data-link index, the lower is the probability of being preempted after a failure. In the pro-tection band, as a consequence, the same order should be used to assign the wave-length for the extra traffic.
5. ConclusionsIn this paper we have discussed the shared-path protection with extra traffic forASON–GMPLS optical rings. Two different approaches have been implemented: theper-wavelength and the full-band approaches. In the former, extra traffic is trans-ported by resources in wavelengths being used by any SP connection for the protectionLSP. In the latter, all resources in half of the wavelengths of the links are reserved forprotection, and thus can be used for extra traffic.
From our study different conclusions can be drawn. The first one is that the SPP inrings presents the same connection availability (even better in the case of thefull-band SPP approach) as dedicated protection, while it shows a better resourceusage ratio. To implement SPP in GMPLS-controlled optical rings, the proper designof the optical nodes (i.e., OADM) has been done. As a second conclusion, we experi-mentally found 1.8 ms as the maximum switching time that provides SPP protectionwithin 50 ms to the maximum of LSPs per fiber when extra traffic is supported. More-over, it has been demonstrated that the number of LSPs that SPP can protect within50 ms is 11 for the switching times provided by currently available WSS. This opensthe possibility to define two classes of protection: the SP-50 class where connectionsare recovered in less than 50 ms, and the SP-100 for protection times under 100 ms.
In case the number of LSPs to be protected is higher than the threshold, the OMSscheme must be applied. In line with this, as a further contribution, a real-timemechanism that, on the basis of the number of protected connections, decides whichprotection scheme has to be applied (OMS shared protection or SPP) has been definedand evaluated. Both methods can coexist using a new OADM design that supportsprotection at both layers. As a final conclusion, on the basis of the experimentalresults, in using SPP only a portion of the spare capacity is used, whereas in usingOMS shared protection all spare capacity needs to be used.
AcknowledgmentsThe work described in this paper was carried out with the support of the BONEproject (Building the Future Optical Network in Europe), a Network of Excellencefunded by the European Commission through the 7th ICT-Framework Program; thei2Cat Foundation through the TRILOGY project; and the Spanish Science Ministrythrough the TEC-2005-08051-C03-02 RINGING project.
Fig. 14. Spare capacity used against the number of protected connections.
Vol. 8, No. 2 / February 2009 / JOURNAL OF OPTICAL NETWORKING 145
References1. C. Ou, J. Zhang, H. Zang, L. Sahasrabuddhe, and B. Mukherjee, “New and improved
approaches for shared-path-protection in WDM mesh networks,” J. Lightwave Technol. 22,1223–1232 (2004).
2. L. Guo, H. Yu, and L. Li, “A new shared-path protection algorithm under shared risk linkgroup constraints for survivable WDM mesh networks,” Opt. Commun. 246, 285–295(2005).
3. W. D. Grover and D. Stamatelakis, “Cycle-oriented distributed preconfiguration: ring-likespeed with mesh-like capacity for self planning network restoration,” in IEEE InternationalConference on Communications (ICC) (IEEE, 1998), pp. 537–543.
4. W. Grover, Mesh-Based Survivable Networks (Prentice Hall, 2004).5. Architecture for the Automatically Switched Optical Network (ASON), ITU-T G.8080, (ITU,
2001).6. E. Mannie, Generalized Multi-Protocol Label Switching (GMPLS) Architecture, RFC-3945
ReserVation Protocol-Traffic Engineering (RSVP-TE) Extensions, RFC-3473 (IETF, 2003).8. D. Katz, K. Kompella, and D. Yeung, Traffic Engineering (TE) Extensions to OSPF Version 2,
RFC-3630 (IETF, 2003).9. J. Lang, Link Management Protocol (LMP), RFC 4204 (IETF, 2005).
10. L. Velasco, S. Spadaro, J. Comellas, and G. Junyent, “Introducing OMS protection inGMPLS-based optical ring networks,” Comput. Netw. 52, 1975–1987 (2008).
11. O. Gerstel and G. Sasaki, “Meeting SLAs by design: a protection scheme with memory,” inOptical Fiber Communication Conference and Exposition and the National Fiber OpticEngineers Conference, OSA Technical Digest Series (CD) (Optical Society of America, 2001),paper OThJ2.
12. J. Perelló, E. Escalona, S. Spadaro, J. Comellas, and G. Junyent, “Resource discovery inASON/GMPLS transport networks,” IEEE Commun. Mag. 45(10), 86–92 (2007).
13. J. P. Lang, Y. Rekhter, and D. Papadimitriou, RSVP-TE Extensions in Support of End-to-End Generalized Multi-Protocol Label Switching (GMPLS) Recovery, RFC 4872 (IETF,2007).
14. L. Berger, I. Bryskin, D. Papadimitriou, and A. Farrel, GMPLS Segment Recovery, RFC4873 (IETF, 2007).
15. K. Kompella and Y. Rekhter, OSPF Extensions in Support of Generalized Multi-ProtocolLabel Switching (GMPLS), RFC-4203 (IETF, 2005).
16. H. Zang, J. P. Juez, and B. Mukherjee, “A review of routing and wavelength assignmentapproaches for wavelength-routed optical WDM networks,” Opt. Networks Mag. 1(2), 47–60(2000).
17. R. Bhandari, Survivable Networks: Algorithms for Diverse Routing (Kluwer Academic,1999).
18. G. Swallow, J. Drake, H. Ishimatsu, and Y. Rekhter, Generalized Multiprotocol LabelSwitching (GMPLS) User-Network Interface (UNI): Resource ReserVation Protocol-TrafficEngineering (RSVP-TE) Support for the Overlay Model, RFC-4208 (IETF, 2005).
19. M. To and P. Neusy, “Unavailability analysis of long-haul networks,” IEEE J. Sel. AreasCommun. 12, 100–109 (1994).
20. S. Verbrugge, D. Colle, P. Demeester, R. Huelsermann, and M. Jaeger, General availabilitymodel for multilayer transport networks,” in Proceedings of the 5th International Workshopon Design Reliable Communication Networks (DRCN 2005) (IEEE, 2005), pp. 85–92.
21. Y. Goebuchi, T. Kato, and Y. Kokubun, “Fast and stable wavelength-selective switch usingdouble-series coupled dielectric microring resonator,” IEEE Photon. Technol. Lett. 18,538–540 (2006).
22. Y. Goebuchi, M. Hisada, T. Kato, and Y. Kokubun, “Optical cross-connect circuit usinghitless wavelength selective switch,” Opt. Express 16, 535–548 (2008).
23. E. Bouillet, K. Kumaran, G. Liu, and I. Saniee, “Wavelength usage efficiency versusrecovery time in path-protected DWDM mesh networks,” in Optical Fiber CommunicationConference, 2001 OSA Technical Digest Series (Optical Society of America, 2001), paperTuG1.
![OG for SDH ASON Network Management-(V100R002C01 02)[1]](https://static.documents.pub/doc/80x56/553cb1564a7959eb2d8b49f9/og-for-sdh-ason-network-management-v100r002c01-021.jpg)
