COMBINING NETWORK ECONOMICS AND ENGINEERING OVER SEVERAL SCALES* Debasis Mitra Mathematical Sciences Research Center Bell Labs, Lucent Technologies Murray Hill, NJ 07974 mitra@lucent.com *JOINT WORK WITH QIONG WANG, STEVEN LANNING, RAM RAMAKRISHNAN and MARGARET WRIGHT BELL LABS MATH CENTER COMPOSITION FUNDAMENTAL MATH Non-linear Analysis - Special focus on Wave Propagation Combinatorics, Probability and Theory of Computing - Applications to Algorithms and Optimization Algebra and Number Theory - Applications to Coding Theory and Cryptography STATISTICS Statistical Computing Environments -Analysis of RDBMS Data traffic measurements, models and analysis -Packet header capture and analysis Online analysis of data streams -Fraud detection Data visualization Statistics in Manufacturing MATH OF NETWORKS & SYSTEMS Networking Fundamentals - Scheduling, statistical multiplexing, resource allocation - Asymptotics and Limit Laws - Large Deviations, Diffusions, Fluid Limits Data Networking - Traffic Engineering, IETF Optical Networking - Design, Optimization, Tools Wireless Networking: - Air-interface Scheduling, Traffic Engineering,Tools Supply Chain Networks - Modeling, Optimization MATH OF COMMUNICATIONS Information Theory Wireless: Multiple Antenna Communications Coding: Fundamental Theory Applications - Optical, Data, Wireless Signal Processing: Source Coding Spectral Estimation BUSINESS PLANNING & ECONOMICS RESEARCH Economics/Business Planning Fundamentals - Models for Competition, Game Theory, Price-Demand Relationships Network Economics - Optimization of Investments, Technology Selections, Net Present Value Strategic Bidding 23 My roots in networking Network Modeling model scale time scale stochastic fluid, diffusion, large deviation models circuit-switched, packet (ATM, IP) spatial scale core & access, wireless & wireline ( optical, data) Network (and QoS) Control closed loop congestion control, designing for delay-bandwidth product open loop leaky-bucket regulation, traffic shaping, priorities effective bandwidth burstiness measure, admission control Network Resource Management scheduling generalized processor sharing + statistical multiplexing resource sharing trunk reservation, virtual partitioning service level agreements structure & management Network Design & Optimization multi-service loss network framework connection-oriented network design traffic engineering deterministic, stochastic, nonlinear software packages PANACEA, TALISMAN, D’ARTAGNAN, VPN DESIGNER Network Economics and Externalities new services diffusion, pricing and investment strategies A Modelling Approach Combining Economics, Business Planning and Network Engineering strategic, long term CAPACITY PLANNING given price-demand relationships and unit cost trends, determine optimal capacity growth path network capacity fixed COMBINED ECONOMICS & TRAFFIC ENGINEERING - joint optimization of multiservice pricing and provisioning - services have characteristic price elasticity of demand and routing constraints Objective: Maximize Revenue wrt prices and routing prices and expected demand fixed RISK-AWARE NETWORK REVENUE MANAGEMENT - revenue from carrying traffic and bandwidth wholesale/acquisition tactical, short term - uncertainty in traffic demand implies risk in revenue generation Objective: Maximize risk-adjusted revenue 4 AGENDA CAPACITY PLANNING - long time scale, strategic COMBINED ECONOMICS & TRAFFIC ENGINEERING - intermediate time scale, strategic/tactical RISK-AWARE NETWORK REVENUE MANAGEMENT -short time scale, tactical 5 Elasticity of Electricity Demand -1.40 1100 1200 1300 1400 1500 -1.60 -1.80 -2.00 Elasticity = 2.2 1926-1970 = 2.2 1962-1970 with very close fit -2.20 -2.40 -2.60 -2.80 -3.00 In (Electricity Generated (M k Wh)) Source: Shawn O’Donnell from Historical Statistics of the Electric Utility Industry: Through 1970, New York: Edison Electric Institute, 1973, Tables 7 and 33. Functional Form Is Constant Elasticity Demand Estimated Price Elasticity is 1.3 to 1.7 for Data Bandwidth, and 1.05 for Voice Bandwidth 6 PRICE vs. DEMAND (log scales) (a) DRAM (b) Electricity 7 DEMAND FUNCTION D p D p p p DEMAND ELASTICITY, E D E D In the limit, REVENUE, R = pD p R ( E 1) R p if E 1 then (reduction in price revenue increases) D CONSTANT ELASTICITY D A A pE A is “demand potential” p 1 8 A FRAMEWORK FOR CAPACITY PLANNING Economic Model Max NPV Technology Roadmap Network Design Economic Model: High price elasticity of demand for bandwidth Technology Roadmap: High rate of innovations in optical networking Exponential decrease in time of unit cost Network Design Algorithms to optimize network design for various technologies 9 OVERVIEW OF CAPACITY PLANNING TECHNOLOGY CONTINUOUS EMERGENCE OF NEW OPTICAL SYSTEMS OPTIMIZATION innovations & cost compression OPTIMAL PLANNING optimize NPV ECONOMICS ELASTIC DEMAND FUNCTIONS decision variables: price, investment, equipment deployment BUSINESS/MARKET DECISIONS DEPLOYMENT OF NEW SYSTEMS PRICING STRATEGIES nonlinear, mixed-integer optimization price-demand relations 10 A SPECIFIC MODEL (Phil. Trans. Royal Soc. 2000) OPTIMIZE NET PRESENT VALUE (NPV) OVER TIME carrier’s long-haul transport network PARAMETRIC MODEL OF PROJECTED INNOVATIONS IN DWDM capacity growth & cost compression exponentiality MODEL PRICE-DEMAND RELATIONSHIP constant elasticity model JOINT OPTIMIZATION OF PRICES & INVESTMENTS multiple time periods nonlinear objective function, nonlinear constraints, integer variables EXAMPLE: 5 CITY, SINGLE RING sensitivity analysis CONCLUSION CARRIER WILL MAXIMIZE NPV BY DROPPING PRICES AND GROWING NETWORK CAPACITY FREQUENTLY 11 PRICES OVER TIME LARGER ELASTICITY PRICES UNIFORMLY LOWER FOR ALL TIME PERIODS LARGER DISRUPTIVENESS HIGHER INITIAL PRICE, LOWER PRICE IN LATER PERIODS 12 CAPACITY (ON A LOG SCALE) OVER TIME EXPONENTIAL GROWTH IN CAPACITY LARGER ELASTICITY LARGER CAPACITY IN ALL PERIODS LARGER DISRUPTIVENESS LOWER INITIAL CAPACITY, GROWS MORE RAPIDLY IN LATER PERIODS 13 “OPTIMAL PLANNING FOR OPTICAL TRANSPORT NETWORKS” S. LANNING, D. MITRA, Q. WANG, M.H. WRIGHT in Phil. Trans. R. Soc. Lond. A Vol. 358, pp. 2183-2196, 2000 BOON Business Optimized Optical Networks Pricing Strategy Business/economic assumptions Financials Network Architecture Capacity Expansion BOON Technology Roadmap Technology Adoption 15 AGENDA CAPACITY PLANNING - long time scale, strategic COMBINED ECONOMICS & TRAFFIC ENGINEERING - intermediate time scale, strategic/tactical RISK-AWARE NETWORK REVENUE MANAGEMENT -short time scale, tactical 16 JOINT OPTIMIZATION OF PRICING & ROUTING IN MULTI-SERVICE NETWORKS • Intermediate time scale i.e. network link capacities are fixed, prices for services are decision variables • Voice & Data are examples of services • Services have distinct demand elasticity to price • Services have distinct traffic engineering/routing requirements e.g. voice needs to be routed over fewer hops than data SERVICE PROVIDER’S PROBLEM: Set prices, which generate demands, and route demands over network to maximize network revenue. 17 OVERVIEW OF THE PROBLEM price Network Pricing network resources price-demand relationship demand generated routing carried demand revenue Traffic Engineering Fixed network capacity, Price is adjustable Traffic Engineering: Mapping generated demand to network resources Dual role of price: (a) determines demand (b) determines revenue SERVICE PROVIDER’S JOINT OPTIMIZATION PROBLEM: Set prices, which generate demands, and route demands over network to maximize network revenue. 18 MORE ON PROBLEM GIVEN: (a) Network and C , capacity on link , (b) s, , set of admissible routes for s, , i.e., s, = r Route r is admissible for service s and (origin, destination) = 1, 2 (c) Constant demand elasticity to price D s D s 1 A s Pss 1 P D.. is demand, P.. is price, A.. is demand potential Assume: elasticity X sr , r ( s, ), s 1 is carried bandwidth (flow) of service type s on route r NETWORK REVENUE, W Ps X sr s , r s , Note:Dual role of price P in determining (a) demand and (b) revenue REVENUE MAXIMIZATION PROBLEM max W Ps ,X sr st r s , s. Ps s, X sr r ( s , X sr ) Ds ( s, ) r s , :r Ps 0 X sr C X sr 0 : demand constraint :link constraint :nonnegativity OBERVATIONS (a) Note Ps Ps Ds (b) Justified in replacing by = in demand and link constraints. TRANSFORMED JOINT PRICING + ROUTING PROBLEM max W Ds , X sr st s , r ( s , s, 1 s As s 1 s Ds X sr Ds s, ) r s , :r X sr C :demand satisfaction :link constraints Ds 0, X sr 0 CONCAVE OBJECTIVE FUNCTION, LINEAR CONSTRAINTS EFFECTIVE ALGORITHMS EXIST FOR CONCAVE PROGRAMMING. NOTE PATH BASED FORMULATION LAGRANGE’S METHOD, SHADOW COSTS Lagrangian, s 1) L( D, X , , ) As1 s Ds( s s , s X sr Ds s , r s , C X sr s , r s , :r Lagrange multipliers, shadow costs: s end-to-end demand matching link capacity constraint RESULTS FROM LAGRANGE’S METHOD OPTIMAL PRICES As Ps D s 1 s s s 1 s OPTIMAL ROUTING either or X sr 0 and X sr 0 and r r s s If is “link cost”, and for any route r, “route cost” then s is “minimum route cost for ( s, )” That is, concave programming r , “minimum cost routing” policy is optimal NOTE UNIFICATION OF OPTIMAL PRICING & ROUTING MECHANISMS AN ILLUSTRATIVE EXAMPLE A 7 1 D B 1 3 C Consider traffic source A, destination B – – – – – Link costs ( l from optimization) shown in figure Min-hop route cost = 7 Least cost of route = 5 Voice required to take min-hop route(s) Data allowed to take up to 5 hops In example, voice route is ( A B) data route is ( A D C B) 24 ASYMPTOTIC PROPERTIES OF OPTIMAL SOLUTION “UNIFORM CAPACITY EXPANSION”: capacities on all links scaled up uniformly i.e. C mC,0, m OPTIMAL PRICES Ps m 1 O 1 max Ps 1 m Optimal prices decrease, but at a lower rate than capacity increase. OPTIMAL DEMANDS Ds m O m s D s 1 max Demand for most elastic service grows linearly with capacity. Demands for all other services grow at sub-linear rates. ASYMPTOTIC PROPERTIES OF OPTIMAL ROUTING Uniform Capacity Expansion m 1. does not necessarily result in minimum-hop routing, 2. provided capacities are sufficiently high, i.e. s 1 s ) As ] , Cmin [( s ( s , ) high price elasticity of one service minimum-hop routing for all services SAMPLE NETWORK 5 2 3 4 6 1 8 Service: voice: 1=1.05, A1,=2000 data: 2=1.5, A2,=200 7 for all Capacity: Cl=400 for all l 27 CHANGE OF TRAFFIC MIX WITH UNIFORM CAPACITY EXPANSION Traffic Mix 80% data 60% 40% voice Capacity 20% 100 10000 28 MINIMUM-HOP ROUTING IS IMPLIED BY HIGH PRICE ELASTICITY r_A 5 3 2 4 FIXED VOICE ELASTICITY 6 ROUTING OF DATA DEMAND WITH CHANGING DATA ELASTICITY r_B R_B 1 8 FIXED LINK CAPACITIES 7 R_A =1.1 =1.2 =1.3 =1.4 =1.5 r_A (4 hops) 100% 100% 62.7% 9.5% 0% R_A (3 hops) 0% 0% 37.3% 90.5% 100% r_B (3 hops) 100% 28% 0% 0% 0% R_B (2 hops) 0% 72% 100% 100% 100% 29 References D.Mitra, K.G.Ramakrishnan, Q.Wang, “Combined Economic Modeling and Traffic Engineering: Joint Optimization of Pricing and Routing in Multi-Service Networks”, Proc, 17th International Teletraffic Congress, 2001 D.Mitra, Q.Wang, “Generalized Network Engineering: Optimal Pricing and Routing for Multi=Service Networks”, Proc. SPIE, 2002 (on my website: http://cm.bell-labs.com/~mitra) AGENDA CAPACITY PLANNING - long time scale, strategic COMBINED ECONOMICS & TRAFFIC ENGINEERING - intermediate time scale, strategic/tactical RISK-AWARE NETWORK REVENUE MANAGEMENT -short time scale, tactical 31 Risk-Aware Network Revenue Management: Overview supply wholesale • • • commodity deterministic demand routing policy constraints • wholesale revenue from selling capacity • installed capacity • opportunity to buy capacity to serve retail and wholesale demands model • quantify revenue reward and risk; • optimize the weighted combination retail revenue management decisions • provisioning • routing • buying • • • differentiated services random demand routing policy constraints • revenue from retail, associated with risk risk tolerance short-term tactical decisions on provisioning, routing and buying capacity - prices and installed capacity stay fixed 32 Objectives •Understand the implications of (uncertain) demand variability on network management, i.e., on provisioning, routing, resource utilization, revenue and risk • Understand the implications of service provider-specific risk averseness •Make the value proposition for resource-sharing between carriers •Create tool for service providers to use for risk-aware network revenue management Problem Formulation network model (L: set of links) cl : installed capacity on link l , pl : unit price for short-term capacity increment bl (decision variable): amount of capacity to buy on link l cl + bl: total capacity on link l Note: we allow cl =0, in which case l is considered a virtual link retail (service) market (V1: set of node pairs) v ( v V1 ) : unit retail price for node pair v, Fv(x) : CDF of retail demand dv (decision variable): bandwidth provisioned between node pair v for serving retail demand, which is random Wr (d ) wv (d v ): retail revenue (random variable) vV1 E[Wr ( d )] mv ( d v ) v 0dv Fv ( x )dx vV1 vV1 Var[Wr ( d )] v2 v2 ( dv ) v2[ 0dv 2 xFv ( x )dx mv2 ( dv )] vV1 vV1 wholesale (commodity) market (V2: set of node pairs) ev (v V2 ) : wholesale price for unit bandwidth between node pair v yv (decision variable): bandwidth provisioned between node pair v for wholesale e vV2 v yv : wholesale revenue 34 The Optimization Model max ( d , y, b , , ) where ( d , y, b , , ) E (W ) (W ) i.e. v mv ( d v ) ev yv pl bl vV1 vV2 retail (mean) 0 dv wholesale lL buying : W is total network revenue (random variable) v v 2 2 vV1 risk ( v V1 ) : provision capacity on route r d is minimum bandwidth rR1 ( v ) v y ( v V ) r required to satisfy GoS v 2 rR2 ( v ) r r cl bl ( l L) : link capacity constraint rR1 ( v ):lr 0 r 0 r 0 bl r dv rR2 ( v ):lr ( r R1 ( v ) : v V1 ) : non-negativity condition ( r R2 ( v ) : v V2 ) for traffic and bandwidth variables ( l L) yv 0 for certain v, bl 0 for certain l 35 : markets in selected links only Example Illustrating Efficient Frontier of Revenue and the Influence of Risk Parameter () standard de viation 3500 3300 inefficient 3100 2900 infeasible 2700 e xpe cte d value 2500 75000 76000 77000 78000 79000 % increase in provisioned bandwidth for wholesale 100% 75% % of total capacity to serve retail demand 50% % decrease in expense of buying bandwidth 25% 0% 0 0.5 1 1.5 2 2.5 36 References D.Mitra, Q.Wang, “Stochastic Traffic Engineering, with Applications to Network Revenue Management”, to appear in Proc. INFOCOM 2003. BACK-UP MULTI-SERVICE NETWORKS • • • Voice & Data are examples of services Demand formulated at aggregated level: total bandwidth for each (s,)=(s, (1, 2)) Service characterization: – distinct QoS routing restrictions (e.g.. voice needs to be routed over fewer hops than data) set of admissible routes for (s,) s, = – r Route r is admissible for service s and (origin, destination) distinct price-demand relationship, as reflected in different values of price elasticity s dDs / Ds dPs / Ps Ds As Pss D A 1 s 1 P 39 Bandwidth Economics: Impact of Rapidly Descending Prices Revenue $300K $258K Revenue Elasticity* $200K 1,000 Units @ $100/each $100K Revenue $39K Revenue 1.5 $100K 1.0 0.5 $100 $15 10% Price Decline / 18 Periods Estimated Price Elasticity for * Elasticity is actually expressed as a Negative 40 Elasticity of Electricity Demand -1.40 1100 -1.60 1200 1300 1400 1500 -1.80 Elasticity = 2.2 1926-1970 = 2.2 1962-1970 with very close fit -2.00 -2.20 -2.40 -2.60 -2.80 -3.00 In (Electricity Generated (M k Wh)) Source: Shawn O’Donnell from Historical Statistics of the Electric Utility Industry: Through 1970, New York: Edison Electric Institute, 1973, Tables 7 and 33. Functional Form Is Constant Elasticity Demand 41 Bandwidth • Bandwidth market is characterized by: – High elasticity---our updated estimate is 1.3-1.7 – rapidly decreasing unit capital costs WDM Capacity doubling every generation (2 years) Elasticity = 2.2 1926 -1970 = 2.2 1962 -1970 with very close fit 3 Tb/s 1 Tb/s 300 Gb/s 100 Gb/s 30 Gb/s 10 Gb/s Functional form is constant elasticity,i.e.,linearity 42 ELASTICITY THERE IS EMPIRICAL SUPPORT FOR THE CONSTANTELASTICITY DEMAND FUNCTIONS D A/ pE E log( D2 / D1 ) /log( p1 / p2 ) Memory (DRAM) 1965 – 1992 Electricity 1926 – 1970 Services voice traffic residential voice traffic (France Telecom, 1999) Equipment digital circuit switch WAN ATM core switch ATM edge switch 1.05 1.337 1.28 2.84 2.11 Optical Systems (source: Lucent Tech.) capacity doubling for same cost every 2 years traffic demand 1.5 every year E 1.6 43 MODEL FOR TECHNOLOGY K = set of WDM technologies k = time period that tech. k is introduced k = max capacity (in OC1) of tech. k CAPACITY GROWTH k k 1 exponentiality ( 1) COST Ikt = acquisition cost of a WDM system of tech. k at time period t exponentiality in per-unit investment costs I k k k (1 d ) I k 1, k 1 k 1 d = “disruptiveness” COST COMPRESSION I k , t 1 I k , t t k 44 PROBLEM FORMULATION: REVENUE, COST single UPSR ring length L I = set of city pairs REVENUE time periods 1, 2, . . . , T E Dijt Aijt /pijt Rt N cities i, jI (i, j ) I pijt Dijt COST conduits, laying fiber are sunk costs, not modelled investment cost for OTU, terminals, regen. & amplifiers: (Ikt) maintenance cost per fiber per mile: mkt bkt = # (WDM systems of tech. k bought in period t) ukt = # (WDM systems of tech. k used in period t) Expense t N I kt bkt 2L mkt ukt k k 45 TECHNOLOGY CONSIDERATION SET MODELED Period Transmission Speed Wavelengths 1 2 3 4 5 6 ... OC48 OC192 OC192 OC192 OC768 OC768 ... 40 20 40 80 40 80 ... Define q, technology disruptiveness, I k k k (1 q) I k 1 k 1 k 1 where I k k is the investment expense of a new system in period k, and k is the capacity of the new system in period k 46 PROBLEM FORMULATION: NPV, CONSTRAINTS CASH FLOW, Ct Rt Expense t DISCOUNT RATE, TERMINALVALUE, 1 f TV CT 1 NPV T t Ct TV t 1 CONSTRAINTS (i) (i, j )I (ii) (iii) Dijt ukt k t k ukt bkt uk,t 1 bkt 0 t k k ,t k PROBLEM max NPV { pi, jt }, {ukt }, {bkt } st constraints nonnegativ ity 47 RESULTS: PARAMETERS 5 city 20 city pair L = 2500 mile T= 10 CAPACITY GROWTH =2 INVESTMENT COST per system cost for tech. 1 in period 1, I11 4.8 10 6 $ 2.5 103 $ per OC - 1 d = 0.2, 0.3, 0.4 e.g. d = 0.3 30% reduction per-unit cost with each new technology = 0.9 per-period reduction in investment cost of already introduced tech. is 10% I kt I11, , d, 48 TECHNOLOGY ACQUISITIONS OVER TIME LARGER ELASTICITY NEW TECHNOLOGIES ACQUIRED SOONER, IN LARGER NUMBERS, MORE FREQUENTLY LARGER DISRUPTIVENESS LESS ACQUISITIONS IN EARLY TIME PERIODS, MORE IN LATER PERIODS 49