1. Introduction

Optical networks are evolving from a fixed ITU-T DWDM wavelength grid to a flexible grid [1] in which the optical spectrum is partitioned into fixed-sized spectrum slices identified by a nominal central frequency and the granularity (e.g. 6.25 GHz) [2]. The required spectral resource (frequency slot) for an optical channel is determined by the used modulation format and data rate. It is dynamically and adaptively allocated by assigning the necessary number of contiguous nominal central frequencies [3], according to the traffic demand and network conditions. The problem of computing a physical route (nodes and links), assigning the needed set of modulation parameters (e.g., format, bits per symbol, Forward Error Correction - FEC), and allocating the required spectrum resources is known as Routing, Modulation and Spectrum Allocation (RMSA) or simply RSA [4]. The main constraint of the RSA is that the allocated nominal central frequencies of a frequency slot must be consecutive (in frequency), and the same frequency slot must be available across all links along the route, allowing an end-to-end continuous spectrum (Spectrum Continuity Constraint, SCC).

Dynamic RSA can rely on either a centralized or a distributed control approach. In a centralized control (e.g., based on OpenFlow [5]), a single controller maintains a detailed, global and unique view of the network topology and the available frequency ranges on each link (i.e., available nominal central frequencies) and physical impairment parameters (e.g., OSNR, CD, PMD, etc). This single controller is responsible for performing combined RSA (i.e., a single entity performs both routing, modulation and frequency slot assignment) and the connection provisioning. Under distributed control (e.g., based on GMPLS), each node has its own controller and maintains its own network state information, collected in a local Traffic Engineering Database (TED) repository. The OSPF-TE routing protocol is then responsible for disseminating any change occurring in the network state information, allowing the synchronization of the nodes’ TED repositories to have a unified view. Dynamic RSA can be performed at either the nodes’ controllers (i.e., source routing) or externalized in a dedicated and centralized Path Computation Element (PCE), using in both cases the collected TED as input (the PCE can construct a local copy of the TED by sniffing OSPF-TE traffic). Finally, the provisioning of end-to-end connections requires horizontal coordination among the nodes controllers. It is performed by the RSVP-TE signaling protocol, employing a label request within a RSVP Path message from the source to the destination node (i.e., forward direction), and a generalized label assignment (reservation) sent in a RSVP Resv message which travels backwards to the source node (i.e., backward direction).

OSPF-TE can both operate on an aggregated basis (flooding resource information on the unreserved optical spectrum) or enabling the dissemination of the status of nominal central frequencies (detailed information) [6]. Detailed information allows to perform a combined RSA, that is, the source node/PCE performs both routing, modulation and frequency slot assignment. Note that the source node/PCE also needs information on the physical transmission impairment constraints (e.g., PMD, OSNR, CD, etc.). This information could also be disseminated through OSPF-TE, alternatively, for PCE-based computation, the physical impairment information could be stored in an extended TED (centralized). The dissemination of the status of the nominal central frequencies on each link through the OSPF-TE protocol is still under consideration by the IETF, with specific initial drafts being submitted as in [6]. However, given the potentially high number of nominal central frequencies with low granularities (e.g., 6,25 GHz as defined by the ITU-T), and the fact that there are still open issues such as efficient encodings or the applicability in link bundling cases, an alternative approach is based on the dissemination of aggregated information which, although admittedly limited by design, does not present the problems bound to the dissemination of large amounts of network link-state information which may cause scalability problems in terms of control bandwidth and processing.

Alternatively, when OSPF-TE operates with aggregated resource information, dynamic RSA should rely on a routing and distributed spectrum allocation approach. In [7,8], the authors proposed and experimentally assessed a routing and distributed spectrum allocation approach for dynamic RSA in GMPLS-enabled elastic optical networks, performing the routing (i.e., spatial path computation) and modulation assignment at the source node/PCE, and the spectrum allocation (i.e., frequency slot assignment) at the destination node. To satisfy the SCC, the RSVP-TE Label Set object was extended to collect the available frequency range information hop-by-hop during the signaling phase from the source to the destination node. This approach is currently considered for standardization in the IETF [9]. However, the lack of available frequency ranges information on each link at the time of the routing impacts negatively on the network performance: using aggregated values results in selecting links that will likely not satisfy the spectrum continuity constraint. The focus of this paper is to present and experimentally evaluate efficient strategies for dynamic source/PCE routing with aggregated information and advanced distributed spectrum allocation algorithms aiming at increasing the network performance in terms of the global spectral resource utilization. The assignment of a modulation for the computed spatial path is out of the scope of this work. An efficient modulation assignment according to the distance length and hop count for elastic CO-OFDM networks is addressed by the authors in [10].

2. Efficient strategies for dynamic routing algorithms with aggregated information

In GMPLS OSPF-TE protocol, resource information is aggregated on a per link basis (e.g., TE link unreserved bandwidth and maximum bandwidth in bytes per second). In elastic optical networks, link resources must be considered in terms of optical spectrum (e.g., Hz), instead of bandwidth (bytes per second). The relationship between the optical spectrum and the bandwidth depends on the chosen modulation format, and since optical channels can support different modulations, the resource information aggregated in bytes per second is no longer useful. Thus, in this work, the TE link unreserved bandwidth attribute conveys the total amount of free optical spectrum for the link, and the maximum bandwidth attribute refers to the maximum optical spectrum supported by the link. Upon reception of a connection request at the source node, such a node (or, alternatively a PCE), executes a path computation algorithm to find a feasible end-to-end path. Each incoming connection request specifies the pair of source and destination nodes and the requested bandwidth in bytes/s.

A common dynamic path computation algorithm operating with aggregated unreserved link resource information computes the constrained shortest hop count path (i.e., link TE metric set to 1) using a modified Dijkstra algorithm. Applied to elastic optical networks, the shortest path computation discards a link for relaxation as long as the unreserved optical spectrum attribute (in GHz) is lower than the estimated optical spectrum. A conservative/maximum optical spectrum estimation approach (worst-case modulation) is to assume 1 bit per symbol (e.g., for a client data rate of 40 Gbps, the estimated optical spectrum is 40 GHz). This approach is sub-optimal for the following reasons: first, the considered modulation formats have a spectral efficiency greater than one bit per second per Hz, so assuming an efficiency of 1 implies discarding potential links which could, in fact, be used for the shortest path computation (that is, the algorithm is overestimating the actual required optical spectrum). Second, it does not take into account fragmentation of the optical spectrum caused by dynamic establishment and release of data plane connections: using aggregated information, the algorithm tends to concentrate spectrum requests in the same spatial shortest paths as long as enough aggregated resources are available. Thus, it is quite unlikely that the computed shortest path can meet the SCC which is checked during the signaling process.

The first problem can be minimized by introducing iterative algorithms, as proposed by the authors in [11]: a path is computed assuming the most efficient modulation format. The path is later validated against the set of feasible combinations, and the step is repeated with a less efficient modulation format if the computed path was not feasible. Such iterative algorithms are more efficient than algorithms that assume worst-case modulation formats that overestimate the required optical spectrum.

The second problem is due to the lack of detailed link information (dissemination of available frequency range information). In [11], the authors proposed and compared a path computation algorithm operating either with aggregated and detailed information. In this work, we propose novel strategies to minimize the effect of the fragmentation of the optical spectrum when using aggregated information at the path computation.

First, we calculate the occupied optical spectrum for each link, subtracting the unreserved optical spectrum from the maximum optical spectrum. Then, a link is considered (for relaxation) if its aggregated occupied bandwidth is above a percentage, γ, of it nominal maximum spectrum. For example, if we consider a maximum optical spectrum of 4000GHz, an unreserved bandwidth of 200GHz, with a γ = 90%, the link will be discarded since the occupied bandwidth (3800GHz) is above the defined threshold (3600GHz). As usual, a link is also discarded if no enough unreserved optical spectrum is available to accommodate the connection request. If after applying this new strategy the path computation cannot find an end-to-end spatial path, a second path computation is applied considering only the unreserved optical spectrum restriction to discard a link. This new strategy aims at balancing the optical spectrum requests through the overall links. It avoids exhausting the shortest-path links, which favors to satisfy the SCC.

3. Distributed Spectrum Allocation Algorithms

The spectrum resource allocation algorithm assigns, from the set of available frequency slots collected from the source to the destination node through the Label Set object, a subset of n nominal central frequencies, based on the requested optical spectrum. In general, the Label Set collects the nominal central frequencies that are available at each outgoing link from the source to the destination node, forming the set of (continuous) frequency slots to allocate from. A basic set of distributed spectrum allocation algorithms were evaluated in [7], showing that First-Fit performed better than (or similar to) the rest of analyzed policies.

Herein, we also propose a new distributed spectrum allocation algorithm, by extending the collected information, in order to perform a more efficient spectrum allocation. Specifically, the proposed extended Label Set collects, at each outgoing link, the status of all nominal central frequencies, in such a way that the destination node has a whole view of the status of the nominal central frequencies on each traversed link (Fig. 1 ). First, the destination node identifies the available frequency slots which meet the SCC (e.g., frequency slots [-1,0] and [5,6] in Fig. 1). Then, for each available frequency slot, we identify the gaps, defined as the number of remaining available nominal central frequencies on each link before and after the first and last nominal central frequency respectively. For example in Fig. 1, the frequency slot [-1,0] has a gap of 0 (before) / 0 (after) available nominal central frequencies in link 1, and 2/0 in link 2. After that, we compute, for each frequency slot, the number of empty gaps with 0 available nominal central frequencies. Finally, the destination node selects the frequency slot with a higher number of empty gaps. Ties are broken with First-Fit. This algorithm aims at performing an efficient resource allocation by minimizing the number of available nominal central frequencies before/after the candidate frequency slot.

 

Fig. 1 Example of the procedure with the standard Label Set and the proposed Extended Label Set to collect the status of all nominal central frequencies.

Download Full Size | PDF

4. Testbed setup and performance evaluation

Experimental performance evaluation of the routing algorithms has been carried out in the 14-node Japanese network topology (Fig. 2(a) ) using the GMPLS/PCE control plane platform of the ADRENALINE testbed (Fig. 2(b)). The usable optical spectrum considers 128 nominal central frequencies, and we evaluate the blocking probability in function of the offered traffic load. We consider a dynamic stochastic model of elastic connection requests, in which the inter-arrival process is Poisson, and the holding time follows a negative exponential distribution, with connection request events uniformly distributed among all distinct source-destination node pairs. For each connection request, the number of requested nominal central frequencies are uniformly distributed between 4 and 32 (scenario A), 2-16 (scenario B), and 1-8 (scenario C). Each data point is obtained requesting 10⁴ elastic connections.

 

Fig. 2 a) Deployed 14-node Japanese topology. b) View of the GMPLS/PCE control plane platform of the ADRENALINE testbed in the lab.

Download Full Size | PDF

The developed GMPLS/PCE control plane for elastic networks considers the following extensions [7,8]: i) a new LSP encoding type and a new switching type, ii) new label semantics and encodings, iii) an extended Label Set to collect available frequency ranges, and iv) a new Explicit Route sub-object, which conveys the PCE assigned optical parameters (i.e., modulation, FEC and requested spectrum) as TLVs.

Figures 3(a) , 3(b), and 3(c) plot the blocking probability in function of the offered traffic load for the presented routing strategies. We consider 5 γ thresholds (90%, 80%, 70%, 60% and 50%) for the occupied optical spectrum, and First-fit as the spectrum allocation algorithm at the destination node. In general, we can observe a very high reduction of the blocking probability when γ = 60%, with a relative difference (i.e., gain) in comparison with standard shortest path between 18% and 67% in scenario A (Fig. 3(a)), 56% and 97% in scenario B (Fig. 3(b)), and 98.48% and 99.97% in scenario C (Fig. 3(c)). Thus, this results shows the selecting links that will likely not satisfy the spectrum continuity constraint

 

Fig. 3 a,b,c):Blocking probability vs. traffic load for different routing strategies for scenario A (a), B (b) and C (c). d,e,f): Absolute and relative difference in Blocking probability between First-Fit and the proposed distributed SA algorithm vs. traffic load for scenario A (d), B (e) and C (f)

Download Full Size | PDF

As for the advanced distributed spectrum allocation algorithms, performance is evaluated by means of simulations with a C + + event-driven simulator. Figures 3(d), 3(e), and 3(f) show the relative and absolute differences in blocking probability between First-fit and the proposed advanced spectrum allocation algorithm in function of the offered traffic load, considering the shortest path algorithm. As it can be observed, the gain in the blocking probability (relative diff.) is quite moderate for scenario A (Fig. 3(d)) and B (Fig. 3(e)), with a quite lineal gain around 2% and 4% respectively. For scenario C (Fig. 3(f)), the gain of the new spectrum allocation algorithm is more significant, rising up to 16%. In any case, these values are quite far from the improvements achieved through the proposed routing strategies.

5. Conclusions

This work has shown that the proposed strategies for dynamic routing algorithms in elastic optical networks are extremely effective to mitigate the effects of the optical spectrum fragmentation when considering aggregated resource information. The proposed algorithm avoids selecting those links that will likely not satisfy the spectrum continuity constraint. The experimental results have shown a reduction of the blocking probability in some orders of magnitude. On the other hand, advanced distributed spectrum allocation policies have marginal benefits with regard to simpler ones such as first-fit, that does not justify the implementation effort required to extend the RSVP-TE signaling protocol.