Motivation Access Metro Core Capacity Increase Ciscos Visual

Motivation Access Metro Core Capacity Increase (Cisco’s Visual Networking Index) Improve efficiency of current systems through better resource allocation Algorithms for next generation systems (higher rate WDM, MLR WDM, flexgrid) OFC 2013 3

Meta-heuristics Iteratively try to improve a candidate solution with regards to a given metric Do not guarantee to find an optimal, as opposed to exact methods (like ILP) A meta-heuristic typically defines: The representation or encoding of a solution The cost function Iterative procedure Meta-heuristic types Local search: iteratively make small changes to a single solution Constructive: construct solutions from their constituting parts Population-based: iteratively combine solutions into new ones However, these classes are not mutually exclusive and many algos combine them Popular meta-heuristics: Genetic/evolutionary algorithms, ant colony optimization, tabu search, simulated annealing OFC 2013 15

Heuristics Heuristic: simple, fast, and can find good enough solutions Depending on the problem, a heuristic can be optimal (but not for the majority of problems that we face) Greedy : at each step make a choice that seems good (towards a local optimum), with the hope of finding a global optimum Combinatorial problems can be solved by allocating resources one-by-one to demands Routing problems: shortest-path, k-shortest paths (weight= #hops, or distance) Wavelength assignment: random, first-fit, least used, most used wavelength Slot assignment: similar to wavelength assignment, but can take into account the size of voids created OFC 2013 16

Single and Multi-objective optimization Most problems are formulated as single-objective optimization problems e. g. minimize #transponders, or # wavelengths, or energy consumption, etc. What if we want to optimize more than one metric Scalarizing: use a single-objective defined as a weighted combination of the multi-objectives minimize: (w. #transponders) + [(1 - w). # wavelengths)] weighing coefficient w controls the dependence on each metric Use multi-objective methods OFC 2013 Objective 2 (#wavelengths) e. g. minimize both the #transponders and #wavelengths No single solution simultaneously accomplishes the two Non-dominated or Pareto front: the set of solutions that cannot be improved in one objective without deteriorating their performance in at Use single least oneobjective of the restmethods Objective 1 (#transponders) 22

RWA algorithms Joint RWA or decomposed R+WA Joint RWA ILP formulations: path and link formulations Path formulation Pre-calculate all or a set of paths for each demand Variable: xp, w is 1 if the specific path p and wavelength w is selected Constraints: flow constraints only at source node, discrete wavelength assignment constraints, no need for wavelength continuity constraints Link formulation Variables: xdlw is 1 if demand d is served by link l and wavelength w Constraints: flow constraints at source & intermediate & destination nodes, (including wavelength continuity), discrete wavelength assignment constraints, Although path formulation seems more efficient, extensions of the RWA problem (e. g. regeneration placement) might need link-related variables Large number of meta-heuristics and heuristics in the literature 30 OFC 2013

RWA + physical layer IP Router Input: Optical X-Connect WDM IP Router Optical X-Connect Network topology, traffic matrix, Physical layer models and parameters (link and OXC model) Optical X-Connect WDM IP Router Output: routes and wavelengths Network layer - RWA: Satisfy traffic and minimize the number of used wavelengths Physical layer - IA: use lightpaths with acceptable quality of transmission Optical X-Connect WDM Switch model Based on WSS IA-RWA cross-layer optimization Link model OFC 2013 36

IA-RWA algos classification Based on where IA is applied RWA + (separate) PLI verification module IA in either R or WA Joint IA-RWA (IA in RWA formulation) Indirect Based on how PLIs are accounted for Worst-case assumption e. g. constraint the path length, # of hops Direct e. g. use analytical models for ASE OFC 2013 calculate PLIs as if all wavelengths are utilized Actual case calculate PLIs based on the lightpaths that are (or will be) established 37

Worst Case and Actual Interference Worst case interference algo: Consider PLIs that do not depend on interference (1 st class PLIs) Assume all wavelengths active (2 nd class PLIs) Prune candidate lightpaths that do not have acceptable Qo. T Actual interference: cross layer optimization algo: n Consider PLIs that do not depend on interference (1 st class PLIs) n Prune candidate lightpaths that do not have acceptable Qo. T n Formulate the interference among lightpaths into the RWA Illustrative example: DTnet topology - single connection request between all (s, d) pairs The reduction in the solution space can deteriorate wavelength performance

IA-RWA algorithm performance (optimality) Problem instances solved using The proposed LP-relaxation algo ILP 100 random traffic instances Zero blocking solutions Using ILP we were able to solve all instances within 5 hours up to load ρ=0. 7 LP-relaxation: the optimality is lost in 2 -3 cases but the execution time is maintained low OFC 2013 39

Indirect (Parametric) IA-RWA algo Number of active adjacent channels (Affected PLIs: Intra-XT, XPM and FWM) Number of intra-channel XT sources (Soft) constrain the number of adjacent channel interfering sources on lightpath (p, w) (Soft) constrain the number of intra-XT interfering sources on lightpath (p, w) B is a large constant used to activate/deactivate the constraint Similarly we constrain the second-adjacent channel interfering sources Cαrry the surplus variables in the minimization objective

WDM network evolution As the network evolves, established connections are teared-down and new are established Operational phase Establish new connection one-by-one (or a small set) Penalize re-routing of established lightpaths Re-plan (re-optimize) the network Periodically or On-demand Network planning Network evolution Establish new Reoptimize 50

Mixed-line-rate (MLR)networks Network with more than one rate (various types of Tx. Rx) Higher rate Tx. Rx, more expensive, less reach Exploit the heterogeneity Serve distant connections with inexpensive, low-rate/long-reach Tx. Rx, and short-distance high-rate connections with more expensive but fewer, high-rate Tx. Rx Use advanced RWA algos to account for the different types of Tx. Rx with different capabilities and costs More complicated PLIs: cross-rate interference effects OFC 2013 51

Transmission Reach & Effective Length of link l : Maximum transmission reach at r : Effective length of fiber t for a transmission of rate r Effective length of the path p at rate r Satisfy: n Rates: r={10, 40, 100}Gbps n Fiber types: t={{10}, {40}, {10, 100}, {40, 100}, {10, 40, 100}} n Interference between different modulation format/rates n mr, t≥ 1: the increase of the length of the link for a connection of rate r, due to interference effects generated by the other modulation formats/rates concurrently transmitted over the fiber t n mr, t≡{r}=1: a fiber (t≡{r}) is used only by connections of a certain rate r (its effective length is equal to its real length) 52

Transmission Reach & Effective Length n Length of link l : n Maximum transmission reach at r : n Effective length of fiber t for a transmission of rate r n Effective length of the path p at rate r n Satisfy: n AC and CB are t={10} n AB is t={10, 40} n Effective length of the path p. ACB at rate 10 n Effective length of the path p. ABD at rate 10 n m 10, {10}=1 n m 10, {10, 40} ≥ 1 53

ILP Formulation utilization of different rates of a link Enable/disable the use of a certain path for a certain rate based on the effective lengths of the links that it comprise it type of fiber used on link E. Varvarigos ONDM, Bologna, 2011 Prohibit the utilization of lightpaths over paths that cannot be used for a transmission at a certain rate 54

WDM cost evolution - MLR migration Plan WDM network considering the evolution of traffic (e. g. yearly for a 5 year window) Minimize the total cost of transponders used throughout the period Given: traffic for each year and cost of transponders for each year Plan for each year from scratch, or Plan using the traffic expected at the end of the period (after 5 years), or Gradually deploy higher rate transponders Clearly 3 rd option can yield lower cost, but the question is when to introduce higher rate transponders, where is the sweet-spot for price vs. traffic growth Need: ILP and heuristic algorithms that take into account the intermediate phases of the network OFC 2013 55

Requirements for Flexible Optical Networking 10 Gb/s n Continuous growth of consumers IP traffic n Emerging high-rate applications, sometimes bursty (video on demand, HDTV, cloud and grid applications, DC interconnection) Se op mi-s tic ta al tic cir cu it C B WDM network 50 G b/s 20 G b/s D A it 10 0 G b/s E ic cu m cir a n al Dyptic o WDM: advanced modulation formats and electronic digital equalization 40 and 100 Gbps channel bandwidths ý WDM has rigid and coarse granularity. A problem that becomes even more severe at higher channel rates þ Requirements: cost and energy scalable, flexible, and with fine granularity network þ Flexgrid optical networks • 6. 25 or 12. 5 GHz slots 56

Flexgrid optical network Spectrum variable (nonconstant) connections, in contrast to standard WDM Prototypes reported Spectrum flexible OXCs Spectrum flexible transponders 2 flexibility degrees: modulation level and spectrum used Benefits þ Finer granularity, spectrum savings, higher spectral efficiency þ Enable dynamic spectrum sharing: statistical multiplexing gains OFC 2013 57

Planning flexgrid networks Input: Network topology, traffic matrix, physical layer models Proposed approach: describe Tx. Rx feasible configurations with (reach-rate-spectrum-guardband-cost) tuples Output: Routes and spectrum allocation RSA (and also the modulation-level used - RMLSA) Minimize utilized spectrum and/or number of transponders, and/or… Satisfy physical layer constraints OFC 2013 58

Physical Layer in Flexgrid Assume tunable transponders (Tx. Rx) Physical feasibility function of a Tx. Rx of type (cost) c reach=fc (rate, spectrum, guardband) reach at which transmission is feasible in terms of BER/Qo. T Alternatively, f can have the modulation format or baudrate as From f we calculate the feasible transmission tuples parameters (reach, rate, spectrum, guardband, cost) describing the feasible transmission configurations OFC 2013 59

Fixed vs. flexgrid Fixed 50 GHz grid Channel spacing 50 GHz Slot width 50 GHz n= -2 occu pied (a) 50 GHz Frequency (THz) Channel occupancy -1 occu pied 0 free 1 occu pied 2 free Channel/slot index Channel/slot occupancy

RSA vs. RWA Flexgrid networks have more flexibility degrees Modulation level # or allocated spectrum slots New formulations are required Link & path formulations (as in RWA) Spectrum slot allocation 1. 2. 3. Slot-related variables: need constraints to allocate contiguous slots + discrete slot-assignment constraints (similar to RWA) Super-slot (set of contiguous slots) variables: need discrete superslot assignment constraints Starting slot variables: need spectrum-ordering of demands to avoid slot overlapping #spectrum slots > # wavelengths (could be >>) Formulations 1 and 2 that depend on the #slots might scale badly OFC 2013 61

RSA algorithm example Places regenerators Decides how to break in more than one connections (if capacity demand at required distance> Tx. Rx capabilities) Multi-objective optimization: minimize Tx. Rx cost and spectrum (a weighted combination of the two) w. TR_cost + (1 - w). spectrum_slots 0≤w ≤ 1 ILP formulation Path formulation, based on starting slot variables K. Christodoulopoulos, P. Soumplis, E. Varvarigos, “Planning Flexgrid Optical Networks under Physical Layer Constraints”, submitted to JOCN OFC 2013 62

RSA ILP algorithm Pre-processing phase 1. 2. Given: Network graph, feasible (reach-rate-spectrum-guardbandcost) transmission configurations (tuples) of the Tx. Rx Calculate for each demand, a set of k-shortest paths Identify the configurations (tuples) that can be used by the transponders over a path define (path-tuple) pairs and calculate the #Tx. Rx, #Reg, #spectrum slots required by each (path-tuple) pair A (path-tuple) pair is a candidate solution to serve a demand RSA ILP algorithm selects the (path-tuple) pair to serve each demand allocates spectrum slots Also developed a heuristic that serves demands one-by-one in some particular ordering (highest demand first), and uses simulated annealing to search among different orderings OFC 2013 63

ILP formulation OFC 2013 64

Fixed vs. flexgrid Flexgrid Slot 12. 5 GHz width Channel occupancy 31. 25 GHz 43. 75 GHz Channel spacing Frequency (THz) Start slot End slot n = -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 Slot index Occupied. Occu Free Occupied Free Slot occupancy pied

Spectrum flexible optical network Spectrum variable (non-constant) connections Spectrum flexible OXCs Spectrum flexible transponders Gains: Spectrum savings, higher spectral efficiency Dynamic spectrum sharing : statistical multiplexing gains similar to those observed in time sharing systems (e. g. OBS, OPS nets) Traditional RWA algos are not directly applicable in OFDM networks

Optical OFDM Transmission Bandwidth-variable OFDM transponders 2 degrees of flexibility Frequency domain: elastic allocation of spectrum, in terms of subcarriers Modulation format: control the modulation format of the subcarriers through DSP: single bit per symbol binary phase-shift keying (BPSK), QPSK (2 bits per symbol), 8 QAM (3 bits per symbol), etc. M. Jinno, et. al. , “Distance-adaptive spectrum resource allocation in spectrum-sliced elastic optical path network” IEEE Commun. Mag. , 2010. 68

Spectral density: Reach vs. capacity After 20% SD-FEC Note: typical commercial reach at 100 Gb/s Zhou et al. , IEEE Comm. Mag. , 2013. (PDM-QPSK, 12% FEC) is 2500 km.

ILP formulation OFC 2013 70

RSA vs. MLR Tx. Rx capabilities according to (*) Reach vs rate capabilities of the flexgrid Tx. Rx Flexgrid vs. MLR network (assuming similar reach-rate capabilities) 2 optimization options: optimize spectrum (w=1) or cost (w=0. 01) * A. Klekamp, R. Dischler, R. Buchali, “Limits of Spectral Efficiency and Transmission Reach of Optical-OFDM Superchannels for Adaptive Networks”, IEEE Photonics Technology Letters, 23(20), 2011. OFC 2013 71

Flexgrid network evolution Flexgrid: finer granularity and more flexibility (when compared to WDM that have wavelength-level granularity, nontunable transmissions) Flexgrid network evolution differs from WDM Traffic variation can be accommodated at different levels new connection requests traffic variation of established connections, served by tuning the Tx. Rx Re-optimization: spectrum fragmentation (more severe in flexgrid) Hard disc defragmentation OFC 2013 72

Benefit: Energy Nowadays, 7 -8% of the world energy consumption is due to ICT Internet represents ~25% of this amount Power Consumption • Average power consumption in current network -25% What mean power consumption should be C. Labovitz, What Europeans do at Night, http: //asert. arbornetworks. com Today, the network is designed and works all the time for the peak requested traffic. If the power consumption of optical transport was proportional to the requested capacity, in average core networks would consume 25% less than current networks.

Flexgrid network evolution Traffic variations can be accommodated at different levels 1 st level: New connection request RSA algo serves the request (assign path and reference frequency) 2 nd level: traffic variation of existing connection Rate variation Spectrum Expansion/Contraction (SEC) If the SEC fails (cannot find free additional slots) trigger RSA to setup an additional connection or reroute the existing OFC 2013 74

SEC policies and dynamic RSA algorithm SEC policy examples CSA policy (no sharing) DHL policy (dynamic sharing) Expansion: use higher spectrum slots until blocked, then use lower spectrum slots Analytical models to calculate network blocking Connection exclusively uses a set of slots Based on multi-dimension Markov-models Algorithm for serving time-varying traffic Transform dynamic to semi-static problem Performs offline allocation using estimations for slot requests and then take into account the analytical model to calculate the network blocking probability K. Christodoulopoulos, I. Tomkos, E. Varvarigos, “Time-Varying Spectrum Allocation Policies in Flexible Optical Networks”, IEEE JSAC, 2013 OFC 2013 75

Performance results Traffic: Single connection between every pair of nodes Each connection generates slots according to a birth-death process Network supports T slots, Guardband G=1 slot Compare Spectrum Flexible network to a WDM system with T/2 wavelengths T=250 slots 1000 Erlangs Spectrum flexible network exhibits superior performance (DHL is up to 2 orders of magnitude better than WDM-RWA case) Dynamic spectrum sharing (DHL policy) reduces the blocking compared to constant spectrum allocation (CSA policy) The proposed analytical models are in close agreement with the corresponding simulations OFC 2013 76

Network Planning and Operation Tool Consolidate planning and operation algorithms in a software tool: Network Planning and Operation Tool (NET-POT) Useful for network operators, equipment vendors and researchers To investigate several issues: optical technology to be used topology design placement of optical equipment (e. g. , Tx. Rx, regenerators, etc) Resource allocation algorithms (RWA, RSA) Physical-layer impairments OFC 2013 77

MANTIS – Upatras NET-POT MANTIS developed at University of Patras Service (cloud) Desktop application Current MANTIS state Web-page UI Desktop application engine Core application engine Offline RSA algorithm Heuristic and ILP (using CPLEX) Goal: Mantis to be a reference to compare network architectures and algorithms OFC 2013 78

MANTIS OFC 2013 79

The “whole” picture

Control

Summary General methods to solve optimization problems in networks WDM networks Goal of planning: satisfy traffic and optimize resource usage Physical layer impairments (cross-layer optimization) Network evolution: establish new connections and re-optimize Flexgrid networks Added complexity due to more flexibility degrees Interdependence among reach-rate-spectrum-guardband parameters Traffic variation can be accommodated at different levels Develop novel formulations Network Planning and Operation Tools - Mantis OFC 2013 82