Flow Control KAIST CS644 Advanced Topics in Networking Jeonghoon Mo <[email protected]> School of Engineering Information and Communications University 1 Acknowledgements Part of slides is from - tutorial of R. Gibbens and P. Key at SIGCOMM 2000 - S. Low’s OFC presentation Jeonghoon Mo October 2004 2 Overview Problem Objectives Kelly’s Framework - Wired Data Networks Extensions - Quality of Service - Wireless Network - High Speed Network: Aggregated Flow Control Jeonghoon Mo October 2004 3 Problem Flows share links: How to share the links bandwidth? Jeonghoon Mo October 2004 4 Problem How to control the network to share the bandwidth efficiently and fairly? Jeonghoon Mo October 2004 5 Link Model Set of resources, J; set of routes, R A route r is a subset r J. Let A jr 1 jr 0 otherwise Capacity of resource j is Cj. A’x c x0 Jeonghoon Mo October 2004 6 A Few System Objectives Max Throughput Max-min Fairness (Most Common) Proportional Fairness (Kelly) -Fairness (Mo, Walrand) Jeonghoon Mo October 2004 7 Max System Throughput 6 6 x2 Maximize: x(1) + x(2) + x(3) x*= (0,6,6) maximizes the total system throughput. However, user 1 does not get anything. => unfair x1 x 3 Two links with capacity 6 Three users: 1,2,3 x(i) : bandwidth to user i x(1)+x(2) <= 6 x(1)+x(3) <= 6 Jeonghoon Mo October 2004 8 Max-Min Fairness 6 6 x2 x1 x 3 Most commonly used definition of fairness. Maximize Minimum of the x(i), recursively. x*= (3,3,3) is the max-min allocation. However, user 1 uses more resources. Jeonghoon Mo October 2004 9 Proportional Fairness 6 6 x2 x1 x 3 Proposed by Frank Kelly Social Welfare: Sum of Utilities of Users log( xi ) i Maximize the Social Welfare x*= (2,4,4) is the Proportional Fair Allocation. Can be generalized into “Utility Fairness”. Jeonghoon Mo October 2004 10 -Fairness Generalized Fairness Definition (1-) x System Objective: Max i pi 1- includes proportional-fair, max-min-fair, max throughput - = 0 : Maximum allocation (p=1) - 1 : Proportional fair allocation - = 2 : TCP-fair allocation - : Max-min fair allocation (p=1) Jeonghoon Mo October 2004 11 -Fairness Trade-off between Fairness and Efficiency - Bigger favors Fairness - Smaller favors Efficiency (source: Is Fair Allocation Inefficient, INFOCOM 04) Jeonghoon Mo October 2004 12 Fairness and Efficiency (Infocom 04) Counter-Example Jeonghoon Mo October 2004 13 Algorithms How to achieve those system objectives? 14 Players source - controls its rate or window based on (implicit or explicit) network feedback router (link) - Generate (implicit) feedback or controls packets Jeonghoon Mo October 2004 15 Source Algorithm TCP Vegas, RENO, ECN XCP Jeonghoon Mo October 2004 16 Active Queue Management (AQM) Priority Queue WFQ RED REM XCP Router Jeonghoon Mo October 2004 17 Kelly’s Model and Algorithm 18 User: rate and utility Each route has a user: if xr is the rate on route r, then the utility to user r is Ur(xr). Ur() --- increasing, strictly concave, continuously differentiable on xr [0 , ) --elastic traffic Let C=(Cj, j J), x=(xr, r R) then Ax C. Jeonghoon Mo October 2004 19 System problem Maximize aggregate utility, subject to capacity constraints max U x r rR Jeonghoon Mo October 2004 subject to Ax C over x0 r 20 User problem User r chooses an amount to pay per unit time wr, and receives in return a flow xr = wr/r wr max U r wr r over wr 0 Jeonghoon Mo October 2004 21 Network problem As if the network maximizes a logarithmic utility function, but with constants (wr, rR) chosen by the users max w log x rR Jeonghoon Mo October 2004 r subject to Ax C over x0 r 22 Three optimization problems SYSTEM(U,A,C) USERr(Ur;r) NETWORK(A,C;w) Jeonghoon Mo October 2004 23 Decomposition theorem There exist vectors , w and x such that 1. wr = rxr for r R 2. wr solves USERr(Ur; r) 3. x solves NETWORK(A, C; w) The vector x then also solves SYSTEM(U, A, C). Jeonghoon Mo October 2004 24 Thus the system problem may be solved by solving simultaneously the network and user problems Jeonghoon Mo October 2004 25 Result A vector x solves NETWORK(A, C; w) if and only if it is proportionally fair per unit charge Jeonghoon Mo October 2004 26 Solution of network problem Strategy: design algorithms to implement proportional fairness Several algorithms possible: try to mimic design choices made in existing standards Jeonghoon Mo October 2004 27 Primal algorithm d xr t wr xr t j t dt rR j t p j xs t s: js Jeonghoon Mo October 2004 28 Interpretation of primal algorithm Resource j generates feedback signals at rate j(t) signals sent to each user r whose route passes through resource j multiplicative decrease in flow xr at rate proportional to stream of feedback signals received linear increase in flow xr at rate proportional to wr Jeonghoon Mo October 2004 29 Related Work Optimization Flow Control (S. Low) Window based Model (Mo, Walrand) Jeonghoon Mo October 2004 30 Optimization Flow Control Distributed algorithm to share network resources Link algorithm: what to feed back - RED Source algorithm: how to react - TCP Tahoe, TCP Reno, TCP Vegas Source alg Link alg Jeonghoon Mo October 2004 31 Welfare maximization Primal problem: max ms x s M s subject to U ( x ) s s s x sS ( l ) s cl , l L Capacity cl can be less than real link capacity Primal problem hard to solve & does not adapt Jeonghoon Mo October 2004 32 Model Network: Links l each of capacity cl Sources s: (L(s), Us(xs), ms, Ms) L(s) - links used by source s Us(xs) - utility if source rate = xs x1 ms xs M s x1 x2 c1 x1 x3 c2 c1 c2 x2 x3 Jeonghoon Mo October 2004 33 Distributed Solution D(p) Dual problem: min p 0 D( p) Bs ( p s ) pl cl s l Bs ( p ) max U s ( xs ) xs p s s ms x s M s ps p l L ( s ) BW price along path of s l s Given p sources can max own benefit individually s s xs ( p ) indeed primal optimal if p is dual optimal Solve dual problem! Jeonghoon Mo October 2004 34 Distributed Solution (cont…) min D(p) Dual problem: Grad projection alg: Update rule: p 0 p(t 1) [ p(t ) D( p(t ))] pl (t 1) [ pl (t ) ( xl (t ) cl )] xs (t 1) U s'1 ( p s (t )) A distributed computation system to solve the dual problem by gradient projection algorithm Jeonghoon Mo October 2004 35 Source Algorithm xs ( p s ) = U s '-1 ( p s ) ps s Decentralized: Source s needs only U s ' ( xs ) and p Jeonghoon Mo October 2004 36 Router (Link) Algorithm pl (t 1) [ pl (t ) ( x l (t ) cl )] aggregate source rate Decentralized Rule of supply and demand Any work-conserving service discipline Simple Jeonghoon Mo October 2004 37 Random Exponential Marking (REM) Source algorithm - Identical but does not communicate source rate Link algorithm - At update time t, sets price to a fraction of buffer occupancy: Theorem: Synchronous convergence Under same conditions (with possibly smaller - Price update maintains descent direction - Gradient estimate converge to true gradient - Limit point is primal-dual optimal Jeonghoon Mo October 2004 ): 38 RED - Idea: early warning of congestion - Algorithm Link: Source (Reno): marking window 1 queue B Jeonghoon Mo October 2004 time 39 RED - Idea: marks for estimation of shadow price - Algorithm Link Source marking rate 1 queue Q 1 fraction of marks Global behavior of network of REM: stochastic gradient algorithm to solve dual problem Jeonghoon Mo October 2004 40 Window-based Model [Mo,Walrand] d1 w1 q11 c1 x1 q21 c 2 x2 q 12 x3 q 23 d2 d3 xi 0, i = 1, 2, 3 qi 0, i = 1, 2, x1 + x2 c1 q1(c1 - x1 - x2) = 0 Q = diag{qi }; X = diag{xi }. A’x c Q(c - A’x) = 0 w = X(d + qA) w1 = x1 d1 + x1 q1 + x1 q2 Jeonghoon Mo October 2004 41 Window-based Algorithm Theorem: [Mowlr98] Let dwi ds =-k i i dt ti wi si := wi - xi di - pi ti := end-to-end delay Then x(t) -> unique weighted -fair point x* Proof: The function i (si /wi ) 2 is a Lyapunov function Jeonghoon Mo October 2004 42 Extensions - Aggregated Flow Control Quality of Service Wireless Network Maxnet and Sumnet Jeonghoon Mo October 2004 43 Aggregate Flow Control Motivations: - High Capacity of Optical Fiber Idea: - player are core routers and access routers. access router: regulates the rate of aggregated flow core router: provide feedbacks to access routers Jeonghoon Mo October 2004 44 Quality of Service Only bandwidth is modeled. QoS is affected by - loss and delay also How to incorporate other parameters? Jeonghoon Mo October 2004 45 Non-Convex Utility Function (Lee04) Considered sigmoidal utility function Non-convex optimization problem =>duality gap (source: J. Lee et. al. Non-convexity Issues, INFOCOM 04) Jeonghoon Mo October 2004 46 Non-Convex Utility Functions Dual Algorithm with Self-Regulating Property Without Self-Regulation With Self-Regulation (source: J. Lee et. al. Non-convexity Issues, INFOCOM 04) Jeonghoon Mo October 2004 47 Wireless Ad-Hoc Network [RAD04] Physical Model:Rate r is an increasing function of SINR. MAC : Each time slot determines power pn,which determines rate xn Routing matrix R and flow to path matrix F are given. Jeonghoon Mo October 2004 48 Random Topology Results 100m x100m grid 12 random node, with 6 pairs of transmissions Jeonghoon Mo October 2004 49 In the wireless Ad-hoc Networks The max-min fair rate allocation of any network has all rates equal to the worst node. The capacity maximization objective leads to starving users. Proportional Fair Allocation give reasonable trade-off between fairness and efficiency. - The worst node does not starve. Jeonghoon Mo October 2004 50 MaxNet and SumNet Source takes max(d1,d2,…, dN) in the maxnet architecture Source takes sum(d1,d2,…,dN) in the sumnet architecture. Max Si Source Jeonghoon Mo October 2004 d1 Link 1 d2 Link 2 dN Link N Destination 51
© Copyright 2025 Paperzz