A66 J. OPT. COMMUN. NETW./VOL. 7, NO. 1/JANUARY 2015
Wright et al.
Minimum- and Maximum-Entropy Routing and Spectrum Assignment for Flexgrid Elastic Optical Networking [Invited] Paul Wright, Michael C. Parker, and Andrew Lord
Abstract—We present two complementary routing and spectrum assignment (RSA) algorithms that use a quantitative fragmentation metric using the concept of Shannon entropy in flexgrid networks. Applying the minimum-entropy (MinEnt) approach to the BT network, support for almost 10% more demands in a static growth scenario is shown. We also present results for a maximum-entropy (MaxEnt) RSA approach, implemented using a genetic algorithm, and operated on the same real BT network topology. The MaxEnt approach avoids fragmentation problems, and it is anticipated that this can increase network utilization. Index Terms—Entropy; Networking; Optical communications; Routing.
I. INTRODUCTION
F
lexgrid is emerging as an important new paradigm in the design of elastic optical networks (EONs) [1], as a means to more efficiently exploit spectrum resources, enable high-capacity superchannels, and allow greater flexibility in the choice of advanced modulation formats. As transmission speeds exceed 100 Gbit/s to support the ever-increasing rise in core network traffic, the use of flexgrid WDM systems is seen as a necessity to allow highly spectrally efficient modulation formats to be carried. Flexgrid removes the requirement to keep to strict 50 GHz ITU channel widths by allowing arbitrary sized channels (typically a multiple of either 6.25 or 12.5 GHz) to be created. While this flexibility allows support for transceivers requiring spectral widths above 50 GHz and allowing a network operator to potentially carry more traffic [2], the mixture of different demand sizes can bring about greater optical spectrum fragmentation, reducing the potential capacity gains in the overall network. Currently, there are two fundamental (but related) issues affecting flexgrid architectures, which could impede the deployment and successful operation of such networks: optical fragmentation and Manuscript received July 1, 2014; revised September 19, 2014; accepted September 24, 2014; published October 15, 2014 (Doc. ID 215078). P. Wright (e-mail:
[email protected]) and A. Lord are with British Telecom Laboratories, Adastral Park, Ipswich, Suffolk, IP5 3RE, UK. M. C. Parker is with Lexden Technologies, 67 Temple Road, Epsom, Surrey, KT19 8EY, UK. http://dx.doi.org/10.1364/JOCN.7.000A66
1943-0620/15/010A66-07$15.00/0
equipment complexity. The first is due to the arrival of dynamically varying and random demands of differing bandwidth requirements. The difficulty of predicting the size and source–destination pairing of such dynamic demands means that the optical spectrum tends to become fragmented into narrow bandwidth portions, which are potentially too small to be usefully utilized; i.e., only the incidence of a particular (and, hence, unlikely) set of demands would ever perfectly exploit the free-spectral gaps and lead to the most efficient utilization of the network. A related consequence of the fragmentation problem is that flexgrid photonic equipment has to work that much harder in order to efficiently deal with a fragmented spectrum. Devices such as sliceable bandwidth variable transceivers (S-BVTs), wavelength-selective switches (WSSs), and wavelength converters are being developed to deal with such spectral fragmentation; however, they end up being technically complex. This increases initial deployment costs (CapEx), since this complex technology is more expensive. Although experimental results for hitless defragmentation of the spectrum across a flexgrid optical network have already been reported [3,4], the innovations required at the data and control planes have yet to become a commercial reality, and therefore network operators may be wary of such approaches until the technology is proven. In the meantime, research on measuring and managing spectral fragmentation has continued, with recent work [5] presenting the concept of utilization entropy to measure spectrum fragmentation. But this qualitative approach does not use the underlying power that comes from the Shannon theory of information, in turn based on the thermodynamic theory of entropy [6]. Entropy offers a powerful universal insight into the intrinsic behavior of natural and engineered phenomena. Applying entropic concepts to networks is well known, e.g., [7], and the study of optical fiber spectrum as a statistically exploited yet finite resource is an obvious candidate for such analysis. That said, only as optical networks have scaled in recent years to ever-larger link and node numbers, with an ever-increasingly exploitable optical spectrum (atomized to an ever-smaller quantum, 100 GHz → 6.25 GHz, etc.) © 2015 Optical Society of America
Wright et al.
VOL. 7, NO. 1/JANUARY 2015/J. OPT. COMMUN. NETW. A67
have stochastically well-behaved studies of optical networking based on statistical mechanical principals become relevant. We have previously defined a quantitative Shannon entropy-based fragmentation metric [8], which accurately measures the fragmentation of a flexgrid optical spectrum. In this paper we present two complementary approaches based on applying a Shannon entropy-based fragmentation metric to the routing and spectrum assignment (RSA) problem. The first approach based on minimum-entropy (MinEnt) concepts [8] actively reduces spectrum fragmentation as traffic is increased, with the expectation that this will reduce the need for network operators to perform possibly disruptive spectrum defragmentation algorithms, but rather simply alter the existing RSA algorithms. The alternative, maximum-entropy (MaxEnt) approach [9] places RSA into a different paradigm context, and offers an innovative approach to the management of fragmented superchannels, with the expectation of lower complexity in network management and equipment requirements.
Fig. 1. Examples of calculating the Shannon entropy fragmentation of the unused spectrum for different levels of spectrum fragmentation. The red and green colors indicate whether a spectrum slot is currently used or unused, respectively.
II. SHANNON ENTROPY FRAGMENTATION METRIC A. Applying Shannon Entropy to MinEnt RSAs The Shannon entropyP fragmentation metric uses N the PN well-known H − i1 pi ln pi formula (where p 1) and can be calculated on the unused spectrum, i1 i the used spectrum, or both. To calculate the metric, the optical spectrum is first considered as a number of slots representing the individual flexgrid quanta (e.g., 6.25 GHz). These can then be grouped into blocks consisting of contiguous slots of either used (irrespective of the individual superchannel signals that make up that block) or unused spectrum. The Shannon entropy metric of a spectrum [8] can then be calculated as
H frag −
N X Di i1
D
ln
Di ; D
(1)
where D is the total number of slots (quanta) across the entire spectrum band and Di is the number of slots used in the current block of contiguous unused (and/or used) spectrum. Large values for H frag indicate higher levels of fragmentation. A number of simple examples of calculating this metric for unused spectrum fragmentation (the metric used in the remainder of this paper) are shown in Fig. 1. The first example shows that completely unused spectrum has a value of zero (as would a completely filled spectrum). The remaining examples show that for the same amount of free spectrum (represented as green squares), the value of H frag varies depending on the amount of fragmentation. The benefit of this approach is that it assigns lower entropies to spectra containing longer contiguous blocks of unused spectrum. These larger blocks are more useful for provisioning of future demands.
The Shannon entropy metric described in the previous section can be applied to MinEnt RSAs in one of two ways. The first method is link-based MinEnt, where the spectrum of each link along a particular path is considered in isolation. For each link in the path, the spectrum profile is searched to find the starting locations with enough free spectrum slots to support the transceiver’s spectral width. For each available position in the spectrum, the resulting difference (ΔH frag ) between the Shannon entropy of the spectrum with and without the new signal being placed there is then calculated. All other positions, which cannot provide enough free spectrum slots, are recorded as having an infinite difference. This process is repeated for the remaining links in the path and the sum of the entropy differences calculated. The frequency slot that minimizes the cumulative entropy difference across all the links in the path is then selected. The second way a Shannon entropy-based RSA can be applied is path-based MinEnt, where the spectrum profiles (represented by a bit sequence in which a 1 indicates a used spectrum slot and a 0 an unused spectrum slot) along all the links in the path are bitwise-OR’ed together to form a single end-to-end profile, which is then searched (as described previously for a single link) to find the spectrum allocation that minimizes the entropy difference. An example of calculating these metrics for a four-link path with 16 frequency slots and a new demand requiring two slots is shown in Fig. 2. The numbers given in the boxes show the resulting entropy difference (ΔH frag ) of placing the new demand starting at that location, and the line of numbers below the link-based example shows the overall differences for the path. Boxes containing “∞” indicate starting locations without enough free spectral slots to
A68 J. OPT. COMMUN. NETW./VOL. 7, NO. 1/JANUARY 2015
Wright et al.
Fig. 2. Examples of calculating link-based and path-based minimized entropy, where red indicates that a spectrum slot is already in use, and green means that a spectrum slot is currently free. The numbers give the entropy difference (at either a link or a path level) of adding a new two-slot demand starting at that slot, and which extends toward the right (higher f number). The infinity symbol indicates that for the two-slot demand request, those particular slots are unavailable—either due to them being already used or because the required neighboring slot to the right is already in use.
support this new two-slot demand. For the link-based approach, placing the new signal at f 8 has the lowest resulting entropy, whereas for the path-based approach placement at either f 5 or f 8 results in the lowest entropy difference. Both MinEnt schemes should be applied over the k-shortest paths between the source and destination nodes to search for possible lower entropy routings and assignments.
B. Simulation Details Simulations have been performed on the 22-node BT core reference network shown in Fig. 3, using a purpose
built modeling tool. Each link in the network can carry up to 5000 GHz of C-band spectrum, and no wavelength conversion of signals occurs in the network. Traffic is randomly generated between source and destination nodes consisting of a mixture of different demand sizes in the proportions and with the parameters as shown in Table I (the modulation format and spectral width for each traffic type are theoretical values [10], assuming an additional 12.5 GHz guard band is added). The blocking probability is measured, and when this reaches 5% for a particular traffic matrix, the number of demands reached is recorded and the next traffic matrix is simulated. The simulation is repeated using three different RSA schemes using the same set of 400 traffic matrices for each. The first two schemes are the link- and path-based MinEnt schemes described previously. The third uses a classical shortest-path routing with first-fit spectrum assignment to provide a comparison with the MinEnt Shannon entropy approach, where shortest paths are checked in increasing length order until a route with available spectrum is found. In all three cases, eight shortest paths (based on fiber length) between the source and destination nodes are chosen as potential routes. The difference in the number of demands reached before 5% blocking is encountered when compared against the classic routing scheme is then obtained.
TABLE I TRAFFIC PARAMETERS AND PROPORTIONS
Fig. 3. BT 22-node core reference network: the 20 blue nodes are nodes that produce and receive traffic, and the two yellow nodes only provide network connectivity without adding or dropping traffic.
Type
Modulation
Width
Proportion of Traffic
100G 200G 400G 1000G
DP-QPSK DP-16QAM DP-16QAM DP-16QAM
37.5 GHz 50.0 GHz 87.5 GHz 200 GHz
25% 25% 25% 25%
Wright et al.
VOL. 7, NO. 1/JANUARY 2015/J. OPT. COMMUN. NETW. A69
C. Simulation Results Given the static-based traffic generation scheme of these simulations, where traffic is only added and never changed or removed, link-based MinEnt shows on average a 9.7% increase in the number of demands that can be supported before a 5% blocking probability is reached when compared to the classical approach. This is a significant increase given that in this static growth scenario the levels of fragmentation are fairly low. The performance of path-based MinEnt is less impressive, only showing a 0.7% increase when compared to the classical approach.
D. Network Analysis A more detailed analysis of the network performance is shown in Figs. 4 and 5 to help illustrate the differences between the three RSA schemes. The curves shown in Fig. 4 show how the overall network fragmentation entropy (the sum of the fragmentation entropies across all network links) evolves as demands are added, while Fig. 5 shows how the overall network utilization (the amount of used spectrum across all links) evolves. Focusing first on the link-based MinEnt scheme, it is clear from Fig. 4 that it maintains low levels of fragmentation entropy across the whole network, indicating that fragmentation is being minimized. This is to be expected given that the algorithm is attempting to obtain the lowest entropy difference across the links. Figure 5 shows that the initial network utilization of link-based MinEnt closely follows the classical result, indicating that similar length routes are being selected. The spectrum allocation of link-based MinEnt means that it can continue to allocate demands as the classical approach starts to experience blocking, causing the utilization curves to diverge. Moving on to the path-based MinEnt scheme, Fig. 4 shows that it results in greater overall levels of entropy (i.e., greater levels of overall spectrum fragmentation across the whole network), but, of greater critical importance,
Fig. 4. Network fragmentation entropy evolution.
Fig. 5. Network utilization evolution.
which therefore results in faster consumption of overall network resources. The reason for this poor performance can be seen in the example shown in Fig. 2, where the combination of the link spectrum has hidden the complexity of the underlying link spectra. Although f 5 and f 8 share the same lowest entropy delta, using f 5 reduces the potential to make use of the larger gaps on links 1 to 3 for other demands, whereas using f 8 results in filling in a common spectrum gap. 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 utilization (defined as the number of occupied spectrum slots divided by the total number of spectrum slots across all links in the network), as indicated in Fig. 5.
III. MAXENT OPERATION
OF A
FLEXGRID EON
To date, the conventional approach when managing a flexgrid network has been to keep the optical fiber spectrum as defragmented as possible, so as to operate in the MinEnt regime, thus allowing easier introduction of new demands into the optical spectrum. For example, in the context of dynamic RSA, introducing a new source– destination pair demand requires the introduction of a new channel within the optical spectrum, while tearing down the demand necessitates the removal of a particular channel bandwidth from the optical spectrum. In optimizing the management of such a MinEnt photonic network, minimized fragmentation is required, with established channels being repacked together as former demands are removed or reduced, so as to simplify the setting up of future new demands by ensuring that large, contiguous spectrum blocks are available. Whereas the typical assumption for maximum efficiency means that demands are packed together as closely as possible (MinEnt), we now describe the opposite operational mode, based on the MaxEnt approach. Such a MaxEnt approach to flexgrid network management automatically solves many of the problems outlined above.
A70 J. OPT. COMMUN. NETW./VOL. 7, NO. 1/JANUARY 2015
Wright et al.
In this case, source–destination pair bandwidth demands are located as far apart from one another as possible across the optical spectrum. Rather than establishing, shifting, and tearing down demands as bandwidth varies between different network nodes, we instead consider having a set of perpetual optical demands between all network node pairs, whose bandwidths narrow or widen according to the load between any pair of nodes. In having perpetual optical channels that can increase or decrease in bandwidth as required over time, it is clear that having all channels maximally spread out (fragmented) across the optical spectrum is the most desirable state to have, so that the unfilled intervening spectral resource between channels can be easily exploited at any time (with minimal negotiation) should additional bandwidth capacity be required between a node pair. Apart from not requiring the ongoing establishment and tearing down of data channels, operating the fiber’s spectrum in a maximally fragmented mode also means that optical guard bands play a less critical role—with the overall spectral efficiency of operation therefore potentially increasing.
indicating that the network configuration shows greater signal separation, i.e., nearer to the MaxEnt ideal. 3) Once the NSE is calculated for each valid configuration within a generation, configurations with a NSE that exceeds a certain value (such as being above the average entropy for the generation) are bred together by means of randomly choosing whether to splice or mutate the values within the network configuration vector to generate a new generation of several thousand candidates with the better traits of the current generation. 4) Repeat steps 1 to 3 on the new generation until the average NSE of successive generations has stabilized. 5) The network configuration with the largest NSE in the final generation is then selected as the solution to the problem and used to configure the network nodes and for allocating future demands.
A. Determining a MaxEnt RSA Using Genetic Algorithm Optimization We employ a genetic algorithm (GA) approach to generate the basic MaxEnt EON design. In order to use a GA algorithm, a way of encoding the network state (the network configuration vector) needs to be defined that represents the RSA for all demands in the network. This process starts by precomputing k-shortest paths between all pairs of nodes, where k is fairly small (e.g., 8) and there is a limit to the end-to-end path length (e.g., 1400 km). The selected route for each node pair can now be recorded by an integer between 1 and k. Assuming 5000 GHz of C-band spectrum on each fiber with no spectrum conversion, and an underlying flexgrid with 1 GHz granularity (as supported by the latest WSS devices), the wavelength allocation for each node-pair demand can be identified by an integer between 1 and 5000. The route/spectrum assignment for each node pair can therefore be defined by two integers and, subsequently, for an N node network, and assuming symmetrical reverse routing for each node pair, the configuration can be defined by a series of N N − 1 integers. The GA algorithm is initialized by creating the first generation of network configurations by randomly filling several thousand network configuration vectors. To generate each successive generation, the following process is followed: 1) validate each network configuration vector to check that there is no signal overlap on any link in the network, ignoring any configuration in which an overlap is present. 2) For validated configurations, the network Shannon entropy (NSE) is calculated, by summing the individual entropies of all links in the network using Eq. (1), where for N blocks of available spectrum, Di is the number of slots in the current block and D is the total number of flexgrid slots in the entire spectrum band. The NSE calculation is used to quantify how good a solution is (i.e., how separated the signals are from one another) with larger values
Additional fiber links, i.e., space-division multiplexing (SDM), added between network nodes to ease hot spots and bottlenecks are straightforwardly dealt with by modeling the increased capacity simply as additional (contiguous) spectrum, but with a periodic property analogous to the free-spectral range characteristic of arrayed-waveguide gratings (AWGs), such that continuity with the spectral paths with other non-SDM links is maintained.
B. Results of Applying the MaxEnt RSA Approach to the BT Network Topology The approach described in the previous section has been applied to the 22-node BT reference network shown in Fig. 6(a), where we assume now that all 22 nodes can generate/receive traffic, meaning there are 231 unique node pairings. Two scenarios were considered: the first, in which each link is served by a single-fiber (SF) pair, and the second, in which each link is served by a dual-fiber (DF) pair. Each run of the MaxEnt GA algorithm on a network of this size took around 8 h to compute on a standard single-core PC. The final SF MaxEnt solution showing the usage and spread of demands across all links is shown in Fig. 6. The final solution for both scenarios showed a range of different link utilizations, with some links being heavily used (e.g., 114 node-pair channels for SF, 103 for DF) and others being lightly used (five for both SF and DF) in terms of the number of demands passing through them. The link figures and coloring in Fig. 6(a) illustrate how this usage varies across the network in the SF scenario. However, both scenarios show very similar link usage, as can be seen in Fig. 7(a), and with an almost identical average number of node-pair channels (46.9 for SF, 46.6 for DF). The reason for this is attributed to the fact that as extra fibers are modeled for additional spectrum capacity, maximal entropy is still achieved by spreading almost the same set of demands further apart across the doubled contiguous DF spectrum. The minor differences are attributed to different GA optimization pathways for the two scenarios. The high utilization of some links in the network shows that the SF solution may not be able to deliver the required bandwidth flexibility/growth for some demands, and indicates that a multiple fiber solution may be required.
Wright et al.
VOL. 7, NO. 1/JANUARY 2015/J. OPT. COMMUN. NETW. A71
Fig. 6. (a) 22-node BT reference network showing MaxEnt link utilization in the single-fiber (SF) pair scenario, where the numbers in nodes show node identifiers and the numbers on links indicate the number of demands being routed over each link. (b) Link usage diagram showing the center-frequency placements of node-pair demands across all links in the network in the SF pair scenario.
(a)
(b)
Fig. 7. Results from MaxEnt GA on the BT reference network. (a) Link usage of solutions. (b) Distribution of allocated spectrum per demand.
Figure 7(b) shows the distribution of the widths of the allocated spectrum slots after the MaxEnt GA algorithm has been run, with the red and green columns showing the results of the SF and DF scenarios, respectively. In the SF scenario the minimum allocation is 16 GHz with the mode at 60–70 GHz, and for the DF scenario the minimum is 25 GHz with the mode at 110–120 GHz. We note for any given demand, its allocated spectrum is shared with the demand’s neighboring allocations, assuming a superchannel approach in which demands can dynamically flex with traffic demands, although ultimately, spectrum is limited by the most utilized link that a demand passes through. Although the modal allocation of both of these scenarios provides useful spectrum bands, especially where shorter paths can make use of modulation formats with higher spectral efficiency [11], the rather low spectrum allocation for a few of the demands does indicate that specific links (hot spots) in the network could benefit from additional fiber pairs. The MaxEnt GA solution of Fig. 6(b) has only been optimized in terms of maximizing the NSE without
considering the size of the spectrum allocations this produces. Further MaxEnt optimization can take place to take account of actual and predicted node-pair traffic levels, so that larger demands are given wider spectrum allocations.
IV. CONCLUSIONS In this paper, we have presented a Shannon entropybased metric to RSA in flexgrid networking, employing either a MinEnt or MaxEnt approach. On the one hand, we have shown that a MinEnt-based RSA approach enables a significant improvement of almost 10% in the number of demands that can be served before reaching critical blocking levels when applied to a realistic network topology when compared to classical routing algorithms. Given that much greater levels of spectrum fragmentation are expected in a dynamic traffic scenario, where demands can be altered or removed, this initial result in a static traffic scenario indicates that the link-based MinEnt approach
A72 J. OPT. COMMUN. NETW./VOL. 7, NO. 1/JANUARY 2015
Wright et al.
shows potential. We have also presented a new methodology for designing an EON using a MaxEnt approach, where each unique node pair within a network is preallocated a route and spectrum assignment, and we use BVTs to dynamically serve traffic between the demands. We have shown how the MaxEnt problem can be solved using a GA-based optimization, and analyzed the results of applying it to a real (BT) network topology. The results show that the MaxEnt approach produces useful insights into network design, and indicate that a multiple fiber solution may be required at certain key links in the network in order to achieve the required bandwidths between network nodes.
[10] J. Zyskind and A. Srivastava, Optically Amplified WDM Networks. Academic, 2010. [11] D. J. Ives and S. J. Savory, “Fixed versus flex grid with route optimised modulation formats and channel data rates of 400 Gbits and above,” in ECOC, London, UK, Sept. 2013, paper P.5.11.
ACKNOWLEDGMENTS The research leading to these results has received funding from the European Community’s Seventh Framework Programme FP7/2007-2013 under grant agreement no. 317999 IDEALIST project.
Paul Wright graduated from the University of York in 2006 with a first-class M.Eng. degree in computer systems and software engineering. After graduating, he joined BT working on optical core and access network research projects in areas such as multilayer networks, fixed and flexible grid WDM networks, routing and spectrum assignment algorithms, SDN control, and PON networks. He has been an author of more than 25 papers at international conferences and in journals and has contributed to the STRONGEST and IDEALIST EU collaborative research projects. He is a member of the BCS and IEEE.
REFERENCES [1] O. Gerstel, M. Jinno, A. Lord, and S. J. B. Yoo, “Elastic optical networking: A new dawn for the optical layer?” IEEE Commun. Mag., vol. 50, no. 2, pp. s12–s20, Feb. 2012. [2] P. Wright, A. Lord, and S. Nicholas, “Comparison of optical spectrum utilization between flexgrid and fixed grid on a real network topology,” in Optical Fiber Communications Conf. (OFC), Mar. 2012, paper OTh3B.5. [3] F. Paolucci, A. Castro, F. Fresi, M. Imran, A. Giorgetti, B. Bhownik, G. Berrettini, G. Meloni, F. Cugini, L. Velasco, L. Poti, and P. Castoldi, “Active PCE demonstration performing elastic operations and hitless defragmentation in flexible grid optical networks,” Photon. Netw. Commun., to be published. [4] L. Gifre, F. Paolucci, A. Aguado, R. Casellas, A. Castro, F. Cugini, P. Castoldi, L. Velasco, and V. Lopez, “Experimental assessment of in-operation spectrum defragmentation,” Photon. Netw. Commun., vol. 27, pp. 128–140, 2014. [5] X. Wang, Q. Zhang, I. Kim, P. Palacharla, and M. Sekiya, “Utilization entropy for assessing resource fragmentation in optical networks,” in Optical Fiber Communications Conf. (OFC), Mar. 2012, paper OTh1A.2.
Michael C. Parker (M’93) received his B.A. (first-class) degree in electrical and information sciences from Cambridge University, UK, in 1992, and his Ph.D. in optical communications also from Cambridge in 1996. In 1997, he joined Fujitsu Telecommunications Europe, Ltd., Colchester, UK, and from 2000 to 2003 he was also with the Photonics Networking Laboratory, Fujitsu Network Communications, Richardson, Texas, USA. From 2003 to 2007, he was with Fujitsu Laboratories of Europe, Ltd. He was appointed Visiting Professor at the University of Essex, UK, in 2004. Since 2008 he has been associated with both Lexden Technologies, Ltd., and the University of Essex. Dr. Parker has extensive experience in WDM optical core/metro/access networking, and both photonics and wireless technologies. He has filed more than 20 patents and has authored more than 150 papers in international journals and at conferences. He has been involved with numerous European research projects, including the e-photon/ONe and BONE European FP7 Networks of Excellence, as well as the OASE, STRONGEST, FIVER, and the current SODALES and IDEALIST research projects.
[6] W. H. Zurek, Complexity, Entropy and Physics of Information. Addison-Wesley, 1990. [7] J. Li, B.-H. Wang, W.-X. Wang, and T. Zhou, “Network entropy based on topology configuration and its computation to random networks,” Chin. Phys. Lett., vol. 25, no. 11, pp. 4177–4180, 2008. [8] P. Wright, M. C. Parker, and A. Lord, “Simulation results of Shannon entropy based flexgrid routing and spectrum assignment on a real network topology,” in ECOC, London, UK, Sept. 2013, paper We.2.E.4. [9] P. Wright and M. C. Parker, “Maximum entropy (MaxEnt) routing and spectrum assignment for flexgrid-based elastic optical networking,” in Optical Fiber Communications Conf. (OFC), San Francisco, CA, Mar. 2014, paper Th4E.7.
Andrew Lord joined BT in 1985 after receiving a degree in physics from Oxford University. He has worked on a wide range of optical network systems and technologies, including long haul subsea and terrestrial DWDM networks. He currently heads BT’s optical core and access research. He has had many years of European project coordination and currently helps lead the IDEALIST FP7 project. He regularly speaks at conferences, sits on several organizing committees, and is one of the Technical Program Chairs for OFC 2015. He is an Associate Editor of the Journal of Optical Communications and Networking (JOCN) and is a Visiting Professor at Essex University.
