This invention relates to apparatus, systems and methods relating to optical data transmission, particularly but not exclusively in respect of optimising the use of optical spectrum resource in a telecommunications network.

The way optical networks are used is undergoing significant change, driven in part by the huge growth of traffic such as multimedia services and by the increased uncertainty in predicting the sources of this traffic due to the ever changing models of content providers over the Internet. Sophisticated modulation schemes for higher bandwidth 100 Gb/s services and beyond are known and come into commercial use in optical networks of large and increasing link and node numbers.

A bottleneck to widespread deployment of such schemes is the "fixed" wavelength grid approach specified by the International Telecommunication Union (ITU), in which the relevant optical spectrum range in the C-band is divided into fixed-sized spectrum slots. As used herein, a "slot", "wavelength" or "channel" is defined as a wavelength or a spectrum of wavelengths associated with a certain signal size. A connection is made by assigning spectral (i.e. wavelength) slots on the optical links comprising the path between source and destination. A problem arises with bit rates greater than 100 Gb/s, which occupy spectral widths too broad to fit into such fixed-sized spectrum slots or channels; forcing to fit by adopting higher spectral-efficiency modulation methods which tend to compromise transmission distances. It seems clear that this decade-old ITU "fixed grid" approach will not work for bit rates significantly greater than 100 Gb/s (e.g. 400 Gb/s and above), highlighting the need for a more flexible approach to be developed. Work on such a new "elastic optical networking" approach is underway, based on a "flexgrid" WDM (wavelength division multiplexing) approach, in which the optical spectrum can be divided up in a way to form spectral slots of arbitrary widths which are selectable depending on, e.g., the bit rates being used, tailored to the requirements of the optical transceiver and accommodating new bit rate services.

Representations of the fixed and flexible grid approaches are depicted in the example graphs shown in Figure 1 , in which graphs (a), (c) and (d) depict examples of fixed grid implementations, while (b) and (e) illustrates the flexgrid approach. Graph (a) illustrates the strict guard bands partitioning adjoining optical channels in a fixed grid implementation, and demands or wavelengths at a particular bit rate. Graph (b) shows how the channels for the demand can be grouped into a single "superchannel" which can be transported as one entity in a flexgrid system. In graph (c), five demands (of varying bit rates and distances) and their spectrum needs are shown on a fixed grid, assuming quaternary phase shift keying (QPSK) modulation. Graph (d) depicts the same service demands of graph (c) _; with adaptive modulation optimized for the required bit rate and reach. Arbitrary-size spectrum slot allocation is shown in the flexgrid implementation of graph (e) which has the same demands as graphs (c) and (d).

In brief, there is typically only one way in fixed grid networks to implement a given demand as the wavelength bit rate, optical reach, and spectrum parameters are highly constrained in their allowable implementation solutions so as to ultimately align with the fixed grid architecture. This means that the demand can occupy less than a full slot resulting in wasted spectrum capacity like that shown in graph (c). On the other hand, a superchannel width wider than the fixed slot width as illustrated in (b) cannot be accommodated in a fixed grid network. In the flexgrid scheme, a choice can be made when implementing a demand, by assigning a modulation format that gives sufficient performance to reach the required distance, while relaxing the requirements on the actual width of the spectral bandwidth occupied by the optical path. The spectrum savings that may be achieved using a flexgrid approach may be seen by a comparison between the fixed grid scheme shown in (d) and an identical scenario when used under the flexgrid scheme shown in (e). In other words, moving away from the use of fixed- position guard bands defining channel widths in the fixed grid approach, can achieve efficiencies by spacing channels contiguously or at least closer to each other, across the spectrum. This can result in the freeing up of spectrum resource for other demands. In this way, the flexgrid scheme allows greater flexibility and choice in allocating spectrum.

A drawback suffered particularly by flexgrid systems however, is that the optical spectrum can become "fragmented", consisting of non-contiguous used spectrum sections in a manner akin to a computer hard disk including fragmented disk blocks. This is because when a signal or demand reaches its destination node, the optical connection terminates and the wavelength or channel "vacates" the spectrum slot. One way of preventing spectrum fragmentation is to find a new resource request having an identical or near-identical slot width, to occupy that slot just at the point when the slot becomes available. The chances of such coincidences occurring are, however, not high, and as may be expected, such vacated slots are likely to remain wholly or partly unfilled. In other words, even if a wavelength channel is found which is capable of fitting into the vacated slot, this is likely to be narrower in width to the previous wavelength channel, resulting in the creation of unused sections of the spectrum so that the level of fragmentation tends to increase over time. The state in which a spectrum comprises non-contiguous used or unused sections results in a state of "entropy" (randomness or disorder). Matters become especially problematic when the unused spectrum parts are so narrowly splintered that they cannot be used to accommodate a demand, even when the total (summed) amount of actual unused spectrum might otherwise have been usable. By way of illustration, graph (e) of Figure 1 depicts an optimal situation in a flexgrid environment, in which the demands essentially occupy contiguous positions in the spectrum. In a less ideal case, fragments of unused "stranded" spectrum fragments can be represented as gaps between neighbouring used slots, as shown in e.g. Figure 3 (discussed below). It can be expected that failure to address this problem may eventually result in significant inefficiencies in the use of the precious spectrum resource, possibly resulting in the need to build expensive new links to cope with traffic levels.

As might be expected, this is much less of a problem in fixed grid systems owing to the standardised slot size arranged contiguously to each other. Indeed, splintering of the spectrum is simply not an issue in a fixed grid system, since all slots are of the same standard size, such that when a channel is terminated and frees up spectral space, then any new demand that arises will automatically be of the same (standardised) width to fit into the available slot. There is therefore a need to address the above issues, especially in connection with the routing of optical data traffic in flexgrid implementations in the elastic optical networking paradigm. Some approaches referring to an entropy measure along optical links are known. For example, "Utilization Entropy for Assessing Resource Fragmentation in Optical Networks" (W. Xi et al., Optical Fiber Communication Conference, OSA Technical Digest (Optical Society of America, 2012)) discusses spectral fragmentation using qualitative and algorithmic descriptions, which as an approach may not provide a predictive measure of the overall spectral fragmentation. This is because algorithmic approaches often implies a degree of non-linearity (e.g. the presence of binary logic steps, i.e. XOR, OR, AND operations etc.), which may cause non-monotonic (and/or non-linear) behaviour of the output result of the algorithm. Nor does an algorithm necessarily offer the degree of 'sensitivity' or ability to distinguish between subtle differences of fragmentation - it depends on the construction of the algorithm, and its quantitative behaviour. Methods based on this approach might not be sufficiently reliable for deployment in the field. Another document titled "Planning and Provisioning of Elastic O-OFDM Networks with Fragmentation-Aware Routing and Spectrum Assignment (RSA) Algorithms" (M. Zhang et al., Asia Communications and Photonics Conference, OSA Technical Digest (online) (Optical Society of America, 2012)) describes a fragmentation-aware RSA, which uses a fragmentation ratio when making resource allocations. It does not however use an entropy-based metric for choosing routes, and the paper explicitly notes that the utilization entropy approach of the W. Xi document above does not sufficiently quantify bandwidth fragmentation to be helpful.

"Dynamic Routing and Frequency Slot Assignment for Elastic Optical Path Networks that Adopt Distance Adaptive Modulation" (T. Takagi et al., Optical Fiber Communication Conference/National Fiber Optic Engineers Conference 2011 , OSA Technical Digest (CD) (Optical Society of America, 2011 )) describes a classical style RSA applied to elastic networks and does not take into account spectrum entropy. As such, the idea of assessing resource fragmentation and fragmentation-aware routing are known, as is the application of entropic concepts in the area of optical fibre spectrum. However, a stochastically well-ordered study of optical networking on statistical mechanical principals using information-theoretic definitions of entropy has not yet been known to have been conducted in connection with optical networks, whose size with their ever-larger link and node numbers, and featuring an ever- increasingly exploitable optical spectrum (atomized to an ever-smaller quantum, 100GHz→6.25GHz) now makes them particularly amenable to such statistical mechanical (thermodynamical) analyses. In particular, there is currently no known approach which allows calculation of an entropy measure in manner which is essentially quantitative in behaviour, (i.e. it can be expressed as a compact mathematical equation) which has highly predictive behaviour, is monotonic as a measure of entropy, and can sensitively distinguish between arbitrarily fine differences of fragmentation. There is also no method allowing for such a measure to be used in an optical path routing choice in a manner which allows overall network resource (i.e. utilisation of overall network capacity) to be more greatly exploited before overall networking blocking probabilities start exceeding certain thresholds (e.g. at 5%). In particular, it would be useful to reduce or stop the spectrum becoming fragmented in the first place. With the aim of full and efficient use of the spectrum resource, network operators would wish to have tools to address the above problems.

According to a first aspect of the invention, there is provided a routing and wavelength assignment method for use in an optical system, comprising:

(i) in an optical path between a source node and a destination node, selecting a block of spectrum resource comprising one or more adjacent wavelength channels, the block being of a width sufficient accommodate a demand having a spectral width occupying one or more adjacent wavelength channels, and

(ii) determining an entropy value of the spectrum entropy level of the selected block based on a logarithm of the ratio of the number of wavelength channels in the block, to the total number of wavelength channels across the spectrum band of the optical path.

A path can comprises an optical link between neighbouring nodes, or an end-to-end optical path between source and destination nodes which are not neighbouring nodes. The calculation of the spectrum fragmentation (or entropy) levels can be expressed as a value for such a link or end-to end-path, which usefully allows for a comparison to be made between paths having different entropy values, where a number of potential paths are available on the path are available. This enables a user such as a network operator to choose to admit demands selectively, based e.g. on a policy of minimising spectrum entropy.

A block of spectrum resource can, as will be discussed below, comprise one or more wavelength channels (also referred to as "slots" herein).

As discussed below, Routing and Wavelength Assignment (RWA) and/or Routing and Spectrum Assignment (RSA) are well known and used in making routing decisions. In their application in the impending elastic optical networking paradigm, issues arising as a result of spectrum fragmentation will need to be taken into account. The quantitative approach described here makes use of the logarithmic Shannon information entropy approach which advantageously yields predictive results, is monotonic as a measure of entropy, and can sensitively distinguish between arbitrarily fine differences of fragmentation.

According to another aspect of the invention, there is provided a processor enabling routing and wavelength assignment decisions in an optical fibre system configured to determine an entropy value of the spectrum entropy level of the selected block based on a logarithm of the ratio of the number of wavelength channels in the block, to the total number of wavelength channels across the spectrum band of the optical path.

In a further aspect of the invention, there is provided a network management system enabling routing and wavelength assignment decisions configured to:

(i) in an optical path between a source node and a destination node, select a block of spectrum resource comprising one or more adjacent wavelength channels, the block being of a width sufficient accommodate a demand having a spectral width occupying one or more adjacent wavelength channels, and

(ii) determine an entropy value of the spectrum entropy level of the selected block based on a logarithm of the ratio of the number of wavelength channels in the block, to the total number of wavelength channels across the spectrum band of the optical path.

Systems, methods and apparatus embodying the present invention will now be described by way of example only, with reference to the following drawings, wherein: Figure 1 illustrates spectrum assignment in fixed grid and flexgrid scenarios,

Figure 2 depicts spectra of different entropy fragmentation levels,

Figure 3 depicts the optical frequency space of a fragmented optical spectrum,

Figure 4 depicts the results of the transformation of the optical frequency space shown in Figure 3 into a state distribution space,

Figures 5(a) and 5(b) respectively represent an unfragmented and ordered optical spectrum, and the associated transformed state space,

Figure 6 depicts the process flow for calculating a fragmentation entropy measure, Figures 7(a) and 7(b) together depict the process flow in a routing and spectrum assignment process of the invention,

Figures 8(a) and 8(b) respectively depict the link-based and path-based entropy assessment approaches,

Figure 9 is a graph depicting the evolution of a network in spectrum entropic terms, Figure 10 is a graph depicting the evolution of a network in terms of utilisation,

Figure 11 is a block diagram of network components involved in the routing and spectrum assignment process of the prior art, and

Figure 12 is a block diagram of network components involved in the routing and spectrum assignment process of invention.

As discussed above, Figure 1 illustrates the differences in the fixed grid and flexgrid paradigms, and a comparison of graphs (d) (depicting fixed grid spectrum use) and graph (e) (flexgrid) highlights the spectrum savings that can be achieved using flexgrid. The techniques and apparatus described herein are largely directed to use of the invention in flexg d networks given the inherent likelihood of fragmentation when slots are vacated in the manner described above. As will however be explained below, it is also possible to use them to advantage in fixed grid scenarios. Fragmentation Entropy Metric

The following describes how a usable entropy metric can be calculated for any system with suitable characteristics. A worked example is also discussed in connection with Figures 3 to 5, to demonstrate an application of an entropy metric in the specific context of fragmentable optical spectrum resources. The fragmentation entropy metric or measure obtained using the method described herein, takes into account distinguishable minimum-entropy states and degrees of freedom in multi- granular/wavelength link traffic demands. The method takes from the Shannon theory of information, and is based on the Boltzmann approach of countable and distinguishable states. The result is monotonic and scale-less, and discriminates varying spectrum topological complexity for efficient resource management.

Generally speaking, a system can exist in N different "states" (as will be described in more detail below), wherein the probability of existing in the ^h state is ,. As known from Shannon information theory, the entropy H of the system is generally given by H = = 1.

In measuring the fragmentation of a system's resource, two points of interest are:

(i) whether a particular resource of the system is being used; and

(ii) the relative location of that resource, in particular whether neighbouring resources to it are being used. Referring to Figure 2, the first (top-most) and last (bottom-most) spectra depict respectively the least and the most fragmented versions of the resource (in which empty boxes deoict unfilled slots and black boxes depicted filled slots). From a practical point of view, when a section of resource is being used contiguously (i.e. no gaps in between), how the different contiguous chunks of resource are arranged within the overall section of contiguous space is not relevant. The first-order requirement is that the resource be as unfragmented as possible. As may be expected, management of a complex topology comprising a puzzle of interstitially-inserted resources (such as the last-depicted spectrum in Figure 2) is more problematic than in a less fragmented system. As noted above in the case of an optical network, resource fragmentation is created when e.g. a demand or requirement previously in use is removed from the system, which is likely to worsen over time if left unchecked, particular in flexgrid implementations. An appropriate tool will be required to assess the relative topological complexity and/or fragmentation of the resource utilisation from time to time or on an ongoing basis. Maintenance in the form of e.g. spectrum re-ordering and re-allocation (akin to a "defragmentation" of a computer hard disk) will have to carried out to ensure efficient use of the available spectrum resource. This is potentially costly and disruptive as operations have to be temporarily halted while the spectrum defragmentation takes place. Having an entropic measure or metric will allow the network owner or operator to take informed decisions about when and how to undergo such maintenance, or test and diagnosis.

In calculating such a metric, a start point is the fill-factor f, which has a value lying between 0 and 1 . The fill-factor / indicates the fraction of the optical spectrum being consumed by the total number of data channels present. For a fully unfragmented spectrum, where the different permutations/combinations of the occupied spectrum are essentially indistinguishable from each other, the entropy of the system is given by

H _mi „(fl = -/ ln / - (l - /) ln(l - /) (1)

By definition, an unfragmented system comprises two sections, one of which is completely occupied and the other is completely unoccupied, so that a portion of optical spectrum exists in only N=2 different states. The probability that any resource quantum of the optical spectrum exists in the completely occupied or filled state is Pi = /, and the probability of existing in the completely unoccupied or unfilled state is p ₂ = 1 - /. The fragmentation entropy of this unfragmented spectral case is referred to as H _min , since it represents the lower bound of the system fragmentation entropy for any given fill-factor f. The shape of H _min as a function of f is the well-known symmetrical curve (even about f=0.5 where it reaches its maximum of 0.693) which is zero at r=0 and /=1.

A fragmented spectrum on the other hand, comprises separated chunks of used or unused optical resource each of which represents different states or blocks of the system. The states or blocks of the system are defined by considering the optical spectrum to be represented by a set of N different number of blocks of spectral resource. Each state or block consists of either a single wavelength or data channel, or a contiguous set of multiple data channels but each of the same kind and spectral width, e.g. all 10G, 40G, or 100G, etc. Referring briefly to the example shown in Figure 3, there are a total of Λ/=1 1 states or blocks, in which state 1 comprises one 10G channel, while state 5 comprises two 40G channels. The spectral width of the ^h block is given by C _h assumed to be an integer multiple of a minimum spectral resource quantum q, e.g. =6.25 GHz. A block of resource can therefore also represent either a single or a contiguous set of unused optical spectrum quanta (see states 2 and 4 respectively, in Figure 3); in which case the single (or repeated) channel is of unit spectral width q.

Because a block comprises either one single channel, or contiguous identical data channels, each block represents a maximally unfragmented (minimum entropy, MinEnt, or minimum topological complexity) sub-domain of the overall optical spectrum. This is because the position occupied by any one of the channels within a particular block is irrelevant to the entropy measure of the block, so that the entropy of the block itself can be thought of as a constant. Accordingly, swapping or shuffling channels within a block offers no advantage or difference in terms of optimal exploitation of the overall spectral resource.

The total (i.e. filled + unfilled) spectrum resource within an optical fibre is assumed to be Pq, i.e. there are a total of P spectral quanta of total resource.

In calculating the entropy measure, the next important abstraction is to consider each individual data channel (10G, 40G, 100G, etc.) as well as each unfilled quantum q, to represent a single "degree of freedom" (DoF) of the system, which is independent of the actual spectral width that each DoF represents. A DoF can be of varying spectral width, in the same way as a block or state. As noted above, a state or block can comprise more than one channel or unused section(s) of spectrum, so referring again to Figure 3, there are 2 DoFs in each of e.g., states 4 and 5. Characterising the data channels in DoF terms allows for appropriate account to be taken of the differing spectral widths of channels (and unused or free spectrum) especially in a flexgrid implementation. As would be expected, each DoF of the optical spectrum affects the overall entropy behaviour of the optical spectrum.

Each block contains Di DoFs, where D _f = Q /Q _{i t} and Qi represents the spectral width of the particular modulation format Q={q, 10G, 40G, 100G, etc.} present in that particular block. The total number of DoFs in the system is D, where D =∑i _=:1 D _i . The probability of a quantum of optical spectrum being in the ^h data block is ρ' _έ = Q/P. However, rather than using this probability quantity which essentially depends only on P (i.e. the total number of resource requirement quanta q in the optical spectrum), the fragmentation of the orderings of the D DoFs of the system holds greater interest, where in general D < P (equality is only achieved for a completely unused optical spectrum). The probability that any DoF lies within the / ^h block is p _t = D D.

The overall fragmentation entropy of the optical link is therefore given by: where D is the total number of DoFs present at that time across the spectrum band and Di is the number of DoFs in the ^h block of used or unused spectrum. This formula provides a highly versatile and usable measure of resource fragmentation entropy and topological complexity, which gives system or network owners or operators information allowing for e.g. defragmentation or other maintenance activity to be carried out on the system. In particular, the metric can be used to make routing and wavelength assignment decisions, as will be discussed below. The metric can be obtained on the unused section of the spectrum, the used section, or both.

As previously observed, the spectrum of an optical fibre is deemed to be completely unfragmented (i.e. with minimum topological complexity) where the spectrum is completely unfilled, or else completely filled to capacity with exactly the same sized channels (e.g. all 10G signals). When the optical frequency space of such an unfragmented spectrum is transformed into and viewed as a state distribution: in the first, completely unfilled, case there is just one state (i.e. unfilled) available with an associated probability ρ^ , such that the fragmentation entropy is _frag = 0. Likewise, where the spectrum is completely full (e.g. with only 10G channels), after transforming into the state distribution we yield a single state (i.e. 10G in this case) again with a probability p- _\ ^=J \ , and associated fragmentation entropy of _rrag = 0. Hence, the entropy-based fragmentation entropy metric can have an absolute minimum value of zero. A relative fragmentation entropy measure h _frag can also be defined, relative to the minimum possible entropy H _min for a given fill-factor f. This is given by:

H frag

nmin

and is zero for the maximally unfragmented case and with minimum topological complexity.

In summary, an entropy metric can be calculated by considering the optical spectrum in terms of a number of slots representing the individual flexgrid quanta. These can then be grouped into N contiguous blocks consisting of either used (coming from any number of individual signals) or unused spectrum. The Shannon entropy metric H _Frag of a spectrum can then be calculated using the formula (2) noted above. Large values of H _Frag indicate higher levels of fragmentation.

The entropy levels of the spectrum resources shown in Figure 2 have been calculated using this approach. The first (top-most) example shows a completely unused spectrum which has a H _Frag value of zero; as expected, a completely filled spectrum would also have a fragmentation metric of zero. The remaining examples show that for the same amount of free spectrum (represented by an equal number of filled and unfilled slots), the value of H _Frag varies (and increases monotonically) depending on the amount of fragmentation, which reaches a maximum in the last example comprising alternating filled and unfilled slots. One advantage of this approach is that longer contiguous blocks of used or unused spectrum, which are more useful for accommodating future demands, tend to result in lower overall fragmentation entropy levels,

An example to demonstrate and discuss how the entropy levels are calculated will now be discussed in connection with Figures 3 to 5 to illustrate the application of the above general principles Figure 3 depicts an example of an optical spectrum partly filled with data channels of differing bandwidths. The y-axis depicts channel power, while the x-axis is a representation of the spectrum resource. In this illustration, there are Λ/=1 1 available blocks (i.e. representing 1 1 states) for this particular spectrum configuration. Here, block or state 1 comprises a 10G data channel, block 2 is "unused" (U) spectrum, block 3 comprises a 100G data channel, block 4 comprises 2 unused quanta of spectrum, and so on. The resource is fragmented in that it comprises used sections as well as unfilled sections. Each channel (U (or q), 10G, 40G or 100G, etc.) is considered to be a single DoF and can be a state in itself (e.g. blocks 1 , 2, 3, 6, 7, 9, 10, 1 1 ). In Figure 3, each DoF is depicted along the linear x-axis, so 2 instances of repeating or adjoining 10G channels can be seen in state 8, for example. Multiple consecutive (contiguous) channels can collectively constitute a single state if they are of the same kind or bandwidth (blocks 4, 5, 8). In the example of Figure 3 therefore, there are a total of D=14 separate channels/quanta or DoFs.

Figure 4 depicts the optical frequency spectral distribution of the example shown in Figure 3 in terms of its state-space distribution. As previously noted, the number of states is N^ ' and the number of DoFs is D=14. Each state 1 to 1 1 is again depicted along the x-axis. In contrast with Figure 3 however, the channels within the states are represented in terms of its DoFs, so for example there is in state 2, 1 DoF (out of a total of 14 DoFs), while in state 4 there are 2 DoFs (out of a total of 14 DoFs). The height of each bar in each state graphically represents the number of DoFs in each state (so the bar in state 4 is twice the height of that in state 2), which is numerically confirmed by the figures appearing above each bar (1 /14 and 2/ 4 respectively in states 2 and 4). In "transforming" the view of the spectrum resource from the optical frequency view of Figure 3 to the state-space view in Figure 4, it is irrelevant whether the channels are arranged from left to right, or vice-versa.

By applying the above-mentioned Shannon formula H = -∑i ₌₁ p _i ln p _i , it can be calculated that the fragmentation metric of the spectrum shown in Figure 3 is _rrag = 2.342.

It may be expected that an unfragmented resource would yield a lower entropy value. This can be demonstrated by reference to the example shown in Figures 5(a) and (b), in which Figure 5(a) depicts the optical frequency of a set of data channels identical to those shown in Figure 3, which are arranged, however, so that the DoFs (which in total remains unchanged at D=14) are clustered together in a maximally unfragmented state. Referring to the 10G channels for example, while they previously occupied states 1 , 6, 8 and 11 (in Figure 3), they now are arranged contiguously to form a single block in state 2. The other channels are similarly treated, so that the 14 DoFs are now distributed across the N=4 states (i.e. Unfilled in state 1 , 10G in state 2, 40G in state 3, and 100G in state 4), which may be compared with the D=14 states in Figure 3. As a result, there are only N=4 states in Figure 5(a).

Figure 5(b) depicts the state-space distribution of that shown in Figure 5(a) with associated probability amplitudes, and the overall fragmentation entropy can be calculated to give an entropy value of H _frag = 1.291. The entropy-based fragmentation metric thus offers an objective means to ascertain the degree of fragmentation present in an optical link, and as the metric is monotonic, the higher H _frag = 2.342 entropy value from the example shown in Figure 3 clearly demonstrates that resource to be more fragmented than the more ordered configuration of Figure 5(a) which has a lower entropy measure Hf _rag = 1.291.

The fact that the fragmentation metric is monotonic makes it useful when considering the relative topological complexity of a fragmented spectrum. A monotonic metric offers a measure of the degree of fragmentation and topological complexity of a particular spectrum configuration. So for example, even for the case of Figure 3 and Figure 5(a) where the fill-factor (i.e. spectrum capacity utilisation) is the same for both cases, the lower fragmentation metric for Figure 5(a) confirms what we can visually ascertain as its lower fragmentation, and its relatively more simple topological- complexity. Although not shown as a figure, if the used spectrum channels of Figure 5(a) were actually all identical, e.g. only 10G channels, (i.e. still the same fill-factor as for Figure 3 and Figure 5(a)) then the associated fragmentation entropy can be calculated using H _min , and found to be /-/ _frag =0.389. This is again much lower than the fragmentation entropy of Figure 5(a) and confirms that it is essentially less fragmented than Figure 5(a), due to its lower topological complexity.

This finding is confirmed when viewing the spectrum resource entropy in terms of its fill-factor. Referring back to the example in Figure 5(a) and assuming for simplicity, quantum Q=10 GHz so that the channel bandwidths are given by 10G≡qf, 40G=4c/, 100G≡10qf, the total available spectrum is P=38g, following from the existence of 38 spectral resource quanta (5xunfilled + 5x10G + 2x40G + 2x100G). With 5 unfilled slots, the fill-factor in this case is =33/38=86.8%. The minimum fragmentation entropy for such a fill-factor is H _min = 0.389, but it can be noted that the total number of DoFs is now D=38. Likewise, for the case with the same fill-factor /=86.8% as is the case for the previous examples, but for an even more ordered, minimally low topological complexity situation of only one type of modulation format (e.g. 10G, so that there are 33x10G essentially indistinguishable contiguous channels), then the fragmentation entropy reduces further to H _frag = 0.389, equal to H _min .

The above approach to derive the entropy metric is essentially scale-less (due to its logarithmic nature) and therefore can be applied in sub-networks, national, and pan- national networks in a naturally additive fashion. Usefully, additionally because of its monotonic behaviour, it enables measurements of the fine differences between the performances of different defragmentation and RSA algorithms, and achieve optimum network resource exploitation. Furthermore, a link RSA having both local and global properties (e.g. modifying the spectral allocations of a local link will have global implications across the network), the entropic approach to measuring the state of disorder of links in a photonic network is analogous to measuring the entropy of a gas of particles, where each particle exhibits both local and global properties (i.e. a particle has a statistical distribution of short-range to long-range interactions/collisions with the other particles in the gas.) Hence, use of a local link fragmentation entropy also offers insights into a means to optimize the overall dynamic equilibrium performance of a photonic network.

Entropy-based Optical Routing

Apparatus and methods based on use of an entropy metric which may be obtained in the manner described above, will now be discussed in relation to making routing decisions within an optical network, as an example of how such metrics may be used. Making routing decisions based on, at least in part, spectrum entropy levels especially in flexgrid-based systems, will help address issues which will arise as traffic levels increase, and could help maintain or even reduce spectrum fragmentation over time. This could reduce the need for network operators to carry out spectrum defragmentation activities which would cause delay and disruption to traffic flows.

In particular, the entropy metrics derived from the Shannon-based approach described in the previous section can be applied to routing issues in at least two ways. The first method is based on a "link-based" minimised entropy measure (MinEnt) where the spectrum of each link along a particular path is considered in isolation. The second way it can be applied is a "path-based" MinEnt where the spectrum profiles along all the links in the path are combined together to form a single end-to-end profile. These two applications will be discussed in greater detail below.

Generally, optical system owners or operators currently use network Routing and Wavelength Assignment (RWA) and/cr Routing and Spectrum Assignment (RSA) algorithms in their network management systems (NMS) to find pathways through an optical network. In the classical RSA approach, the process comprises two separate steps: first, a route across the network is chosen following a shortest distance or minimum hops algorithm; and then a first fit algorithm is used to select the first free end-to-end wavelength block that can be found. The following pseudo code illustrates the approach:

ClassicalRouteAndAssignDemand(Network, Source, Dest, Width)

PathCandidates = MultiShortestPalhs (Network, Source, Dest) if no PathCandidates found then return and block Demand for each Path in PathCandidates in ascending order

Spectrum = GetPathSpectru.n(Network, Path) for each Slot in Spectrum

TxProfile = GetTxSpsctrum(Slot, Width) if SpectrumFree(Spectrum, TxProfile) then

return and allocate Demand <Path, Slot>

end if next Slot

next Path return and block Demand

Embodiments and implementations of the invention seek to include entropy measures into routing and spectrum assignment decisions to find a routing and spectrum assignment that minimises entropy. This is achieved by use of an algorithm which assigns routes and spectrum by allotting the smallest slot which can accommodate the particular demand, in order to keep larger (contiguous) blocks of spectrum free, and/or for larger demands that need it. In this way, demands of narrower spectral width are not assigned to a large slot where a narrower but suitable slot is available, thus reducing the creation of unused sections next to the newly-occupied slot and allowing for a higher number of demands of various sizes to be supported by making better use of the spectrum.

The entropy measure routine could be based on the approach described above using a version of e.g. formula (2) described above to obtain the H _Frag value. In this approach, the entropy value is calculated only for the unused spectral components of a given optical spectrum by finding each block of unused spectrum and calculating the following:

(number_of_unused_slots / total_slots_across_spectrum) *

LN (number_of_unused_slots / total_slots_across_spectrum) where LN is the ratural logarithm, and the "total_slots" is equal to the total number of spectrum slots (unused and used) across the whole spectrum band. This calculation is repeated for each block of unused spectrum and added together to get an overall calculation of the fragmentation entropy. It should be noted that a "slot" is deemed to be equivalent to a quantum q of spectrum, so that total slots across the spectrum is P as previously defined, and the number of unused slots is U where the total unused spectrum is therefore Uq. This can provide a usefully-quick (i.e. a less algorithmically complex) approximation to the fragmentation entropy of equation (2). The flow chart of Figure 6 depicts the steps involved in calculating the entropy measure of a given spectrum resource, in which the process begins by setting the entropy value ("frag_entropy") to zero (step S2). A block of unused spectrum is found (step S4) and the number of unused slots ("unused_slots") discovered up to that point is recorded (step S6). The fragmentation entropy level for the unused spectrum block under consideration is calculated using the above formula (step S8). If there are still unused blocks to be processed (step S10), the steps are iterated until all the blocks in the spectrum have been processed, at which point the entropy metric is obtained (step S12) by adding together all the entropy level values of each unused spectrum block to get an overall calculation of the unused-spectrum fragmentation entropy.

In pseudo code terms; the approach can be expressed thus:

CalcFrag Entropy( Spectru m )

TotalSlots = count of number of slots in Spectrum

. Entropy = 0 for each LnusedBlock in Spectrum

UnusedSlots = count of number of slots in UnusedBlock

Ent-opy = Entropy + UnusedSlots / TotalSlots ^* ln(UnusedSlots / TotalSlots) next block return -Entropy The obtained entropy measure sits at the heart of the routing process described in the flow chart set out in Figures 7(a) and 7(b). In brief, the approach involves the finding of a route though the optical network which has the lowest unused-spectrum fragmentation ent ^r opy, which is achieved by replacing the conventional lowest-cost path algorithm with a fragmentation entropy metric obtained by the exemplary method shown in Figure 5. This requires the used spectrum state of each branch of the network to be stored and the spectrum of each additional hop under consideration needs to be added to these states. The fragmentation entropy of the spectrum is calculated using the formula given above. Generally speaking, discovery of the path with the lowest unused-spectrum fragmentation entropy path allows the network owner or operator to select a spectrum allocation which reduces the level of end-to-end fragmentation entropy along the chosen path, or else causes the fragmentation entropy to increase by the smallest amount. To accomplish this, the spectrum required for the transceiver is placed in every position along the entire spectrum. If for a given position the spectrum is currently unused, then the fragmentation entropy is calculated as set out in the flow chart of Figure 6. The owest unused-spectrum fragmentation entropy state is stored. Once identified, the routing and spectrum assignment is output to the NMS to be configured on the network nodes. As the skilled person would appreciate, a number of variations can be made to the implementations discussed here: for example, the two steps (of identifying and apptying the entropy metric) can be more tightly coupled to form a single step, which might require additional computational time and resources. In any event, it can be expected that routing traffic in this manner would realise a network having an overall lower entropy measure, which will allow for bigger sections to remain available allowing for accommodation of future demands. Turning now to the exemplary process shown in the flow chart of Figure 6: at the start of the process, information about the topology and state of the network is used to find a number n of paths between source A and destination B (step S22) using conventional methods e.g. based on Dijkstra's algorithm. Where no path is found (step S24) (e.g. there is no discoverable path to the destination node S), the process is blocked (step S26) and terminates. If a path is available, the shortest path is selected and the used spectrum profile of that path determined (step S28). At this point, the profile of the transceiver device (which is carrying the traffic between source and destination) is applied to the path spectrum profile (step S30) at a first location along the spectrum profile. If a section of free spectrum is found (step S32, which includes the routine set out in Figure 6), the fragmentation entropy of the newly-identified path is calculated (step S34). A comparison is then made of the entropy levels of the latest-discovered path with any previously-discovered paths based on the spectrum configuration that would result from using the latest-discovered path (step S36), and the spectrum allocation in question is stored (step S38). The transceiver profile is now along by one slot to allow further placements along the spectrum to be checked (step S41 ). When the spectrum has been fully searched (step S40) and the lowest entropy allocation for this spectrum has been found for this new demand along the current considered path (step S42), a comparison with previously-discovered paths is made to determine if the latest-discovered path has the lowest entropy rating (step S44). If it does, the path and spectrum allocation is stored (step S46). The steps to find the path with the lowest entropy value are iterated until all n paths initially identified in step S22 have been checked (steps S48 and S49). If no spectrum and path combination to the destination node B has been identified, the process is blocked (step S52) and terminates. Otherwise, the path and spectrum allocation generating the least entropy in the spectrum resource is selected (steps S50 and S54), which seeks to guard against "waste" of the spectrum, e.g. avoid placing new demands within a large unused slot when narrower but suitable slots are also available. As noted above, the obtained entropy metric can be applied to RWAs and RSAs in at least two ways. The flow charts of Figures 6 and 7 allude to these methods, which will now be described in detail.

Entropy-based Optical Routing: (i) Link-based Measure

In the "link-based" approach based on minimised entropy level, the spectrum of each link or hop (i.e. between nodes) along a particular path is considered separately from each of the spectrums of the other links along the path. Specifically, the spectrum profile of each link is searched to find starting locations with enough free spectrum slots to support the transceiver's spectral width. For each available position, the change in the entropy levels of a spectrum in which the new signal is placed, is calculated. All other slots or positions which have insufficient spectral capacity, are ascribed an infinite delta (or change in entropy). This is repeated for the remaining links in the path and the sum of entropy deltas taken. The frequency slot with the lowest sum of entropy deltas is then selected.

The concept can be illustrated in pseudo code as follows:

MinEntLinkRouteAndAssignDemand(Network, Source, Dest, Width)

PathCandidates = MultiShortestPaths(Network, Source, Dest) if no PathCandidates found then return and block Demand

LowestEntropyDelta = infinity BestPath = null

BestSlot = null for each Path in PathCandidates in ascending order

EntropyDeltasrj = infinity for each Link in Path

Spectrum = GetSpectrum(Network, Link)

OrginalEntropy = CalcFragEntropy(Spectrum) for each Slot in Spectrum

TxProfile = GetTxSpectrum(Slot, Width) if SpectrumFree(Spectrum, TxProfile) then

TempSpectrum = Combine(Spectrum, TxProfile)

NewEntropy = CalcFragEntropy(TempSpectrum) EntropyDelta = NewEntropy - OriginalEntropy

EntropyDeltas[Slot] = EntropyDeltas[Slot] EntropyDelta

else

EntropyDeltasfSlot] = infinity

end if

next Slot

end if for each Slot in EntropyDeltas

if EntropyDeltasfSlot] < LowestEntropyDelta then

LowestEntropyDelta = EntropyDeltas[Slot]

BestPath = Path

BestSlot = Slot

end if

next Slot next Path if LowestEntropyDelta = infinity then

return and block Demand

else

return and allocate Demand <BestPath, BestSlot>

end if

In a worked example illustrating the above, Figure 8(a) depicts a path comprising 4 links (S1 to S4) with 16 frequency slots (f1 to f16) along the spectra of each link. Again, the empty boxes depict unfilled slots and black boxes depicted filled slots, so it can be seen that all the spectra are fragmented to varying degrees. End-to-end paths for single-slot demands are available through all 4 links in slots f5, f6, f8, f9, f11, f12 and f13. In this example however, a new demand requiring 2 slots needs to be placed on the path, for which there is end-to-end path availability in slots f5, f8, f11 and f12. The values in the boxes show the entropy change or delta (AHfrag) if the new demand is placed starting at that location. These are obtained using the entropy calculation method described above for each link. Referring to link S1 as an example, slot f4 has an available slot next to it (slot r ^~ 5) and so can accommodate a 2-slot demand. If it takes on the demand, the entropy delta value is -0.141. Slot f5 has identical characteristics, and so will also have an entropy delta value of -0.141 if it takes on the demand. Slot Π is occupied, and so slot f6 has no capacity to take on a 2-slot demand, and so its change value is infinity. The values in the line Sum(AHfrag) shows the overall delta values for the end-to-end path in each available slot across all 4 links. Slot f8 yields the lowest overall (summed) MinEnt value of -0.742, and so according to exemplary implementation of the invention, the best choice would be to place the new signal in slot f8.

In brief, the method can be summed up in the following steps: (i) identifying a plurality of paths between a source node and a destination node,

(ii) selecting one of the plurality of identified paths,

(iii) defining a new demand as occupying one or more adjacent wavelength channels. (iv) selecting one of the plurality of possible new demand placements from the free spectrum available in the spectrum band on the selected path between the source node and the destination node, temporarily allocating it within the spectrum band,

(v) defining within the spectrum band of the selected path one or more blocks of spectral resource, in which each block comprises of one or more adjacent used or unused wavelength channels,

(vi) obtaining an entropy value of the selected path defining the spectrum fragmentation across its spectrum band, based on a logarithm of the ratio of the number of wavelength channels in each of the one or more blocks, to the total number of wavelength channels across the spectrum band,

(vii) iterating steps (iv) to (vii) until the entropy value of each of the available demand placements has been determined,

(viii) iterating steps (ii) to (viii) until the entropy value of each of the plurality of identified paths has been determined, and

(ix) choosing from the plurality of identified paths and demand placements a path and wavelength assignment having the lowest entropy.

Entropy-based Optical Routing: (ii) Path-based Measure

The second way entropy measures can be applied in routing decisions is based on a path-based MinEnt value, where the spectrum profiles along all the links in the path are combined together to form a single end-to-end profile which is then searched to find the spectrum allocation that produces the smallest entropy delta.

The following pseudo code describes the idea behind the path-based approach:

MinEntPathRouteAndAssignDemand(Network, Source, Dest, Width)

PathCandidates = MultiShortestPaths(Network, Source, Dest) if no PathCandidates found then return and block Demand

LowestEntropyDelta = infinity

BestPath = null BestSlot = null for each Path in PathCandidates in ascending order

Spectrum = GetPathSpectrum(Network, Path)

OrginalEntropy = CalcFragEntropy(Spectrum) for each Slot in Spectrum

TxProfile = GetTxSpectrum(Slot, Width) if Spectrum Free(Spectrum, TxProfile) then

TempSpectrum = Combine(Spectrum, TxProfile)

NewEntropy = CalcFragEntropy(TempSpectrum)

EntropyDelta = NewEntropy - OriginalEntropy if EntropyDelta < LowestEntropyDelta then

LowestEntropyDelta = EntropyDelta

BestPath = Path

BestSlot = Slot

end if

end if next Slot

next Path if LowestEntropyDelta = infinity then

return and block Demand

else

return and allocate Demand < BestPath, BestSlot>

end if

Referring to Figure 8(b), which is based on the 4-link 16-frequency slot example initially discussed in connection with Figure 8(a), the approach in which the availability of an end-to-end path through all the links as a preliminary requirement is graphically depicted. Here, the two best slot choices are f5 and f8, both yielding the lowest MinEnt value of -0.31 1. This accords with the result of the link-based approach, with the added refinement to the result that of the two preferred slots, placement of the new signal into slot f8 would add less fragmentation to the system. Both schemes can be applied over a number of potential paths between the source and destination node to search for lower entropy routings and wavelength assignments.

Network Evolution Projections The applicants have performed simulations on a specific network configuration comprising source and destination nodes, to test the usefuless of the two entropy- based routing approaches described above. Figures 9 and 10 provide a view of network performance against increasing demand levels over time, and illustrate the different results using the three RSA schemes. The curves shown in Figure 9 show how the overall unused-spectrum network fragmentation entropy (the sum of the unused-spectrum fragmentation entropies across all network links) evolves as demands are added, in which the average network entropy on the y-axis is plotted against the level of demand on the x-axis. Figure 10 shows how the overall network utilisation (the amount of used spectrum across all links) evolves, and plots the average network fill factor (i.e capacity used, on the y-axis) against demand (on the x- axis). In both graphs, the projections based on the three RSA scheme are depicted for comparative purposes: the link-based MinEnt data is depicted by line (a), the path- based MinEnt data by line (b), and the classical shortest-path routing with first-fit spectrum assignment is indicated by line (c).

Referring now to Figure 9, which depicts a static based traffic generation simulation wherein traffic is only added and never changed or removed, it can be seen that using the link-based MinEnt scheme (line (a)) can yield a significant improvement in the capacity to absorb increased traffic increase levels, when compared to the classical approach (line (c)). Lower levels of entropy are maintained across the whole network indicating that fragmentation is being minimised. This is to be expected given the algorithm is attempting to obtain the lowest entropy delta across the links. As indicated by curve (b), use of the path-based MinEnt approach generated more entropy than the classical approach, yielding only a limited improvement over the conventional approach only at a high level of demand.

Figure 10 shows that network utilisation of link-based MinEnt schemes initially follows the classical result closely, indicating that similar length routes are being selected by both schemes. At higher levels of demand and traffic levels however, line (a) of the link-based MinEnt spectrum allocation approach gradually diverges from the result of the classical approach, indicating that demands can be allocated even as the classical routing and allocation approach starts to slow down on that front. In the case of the path-based MinEnt scheme, curve (b) shows that its use results in greater overall levels of entropy, but of greater critical importance, results in faster consumption of overall network resource.

Referring back to the worked example shown in Figures 8(a) and (b) provides an insight into the poor performance of the path-based MinEnt scheme. The path-based approach combining the spectrum profiles over the four links has hidden the complexity of the underlying spectrums and results in placements at f5 and f8 sharing the same lowest entropy delta. If f5 was selected, it reduces the potential to make use of the larger gaps present on links S1 to S3 in other paths, whereas placement at f8 results in filling in a common spectrum gap and reducing overall fragmentation. Longer paths are also likely to be selected as the number of links in the path is hidden by the spectrum profile combination resulting in higher levels of network utilisation. By making use of all the information available, the link-based MinEnt scheme correctly identifies f8 to be the optimum choice. Finally, Figure 1 1 is a schematic view of the process flows between the network components used in routing and allocation management in a conventional set up. The main functional block is the NMS (100) which sends and receives instructions. The process typically commences with a request for a connection to be made between source node A and destination node B. This request is sent to the NMS (step S61 ), which seeks and obtains data about the network state (steps S62 and S63). This information is then used to discover a suitable path based on the network state (step S64), which according to standard RSA methods (86) would include shortest-path calculations. The identified path is then returned (step S65) to the NMS. The WDM transceivers (102) are then configured (steps S66 and S67) to place the signal along the path identified for that purpose.

Figure 12 illustrates the relative ease with which an entropy-based metric can be deployed within the routing system by replacement of the conventional RSA method with a process or routine for calculating the entropy metric of the path or the link as the case may be. The skilled person would appreciate that as an alternative to replacement, the system may include entropy calculations in addition to standard routing decisions, and that both the link- and path-based methods could in suitable cases allow for deeper analyses of the network and identified routes to be carried out before a decision is taken.

The apparatus, methods and configurations described above and in the drawings are for ease of description only and not meant to restrict the scope of the invention to any particular embodiment. For example, it will be apparent to the skilled person that steps can be added or omitted from the methods and processes described herein. While the examples illustrating application of the invention are made in respect of an optical network and in particular in connection with flexgrid-based systems, it would be appreciated that other telecommunications systems as well as non-telecommunications systems can suffer from resource fragmentation as well during use, which could benefit from an analysis of entropy levels. In particular, entropy based fragmentation RSA techniques can also be applied in fixed grid scenarios to select a path and wavelength channel which reuses released channels in the network more optimally than starting to make use of a new wavelength that is currently unused in the network. It would also be appreciated that such entropy measures can be advantageously applied in a variety of situations, not being restricted to use only in respect of identifying network routes.