optimal network flow allocation
play

Optimal Network Flow Allocation EE 384Y Almir Mutapcic and Primoz - PowerPoint PPT Presentation

Optimal Network Flow Allocation EE 384Y Almir Mutapcic and Primoz Skraba 27/05/2004 Problem Statement Optimal network flow allocation Find flow allocation which minimizes certain performance criterion Lowest average delay through the


  1. Optimal Network Flow Allocation EE 384Y Almir Mutapcic and Primoz Skraba 27/05/2004

  2. Problem Statement � Optimal network flow allocation � Find flow allocation which minimizes certain performance criterion Lowest average delay through the network � Minimize maximum link utilization � Fair bandwidth allocation and QoS agreements � � Trade-off between optimality and simplicity � Devise practical schemes with low computational complexity and guaranteed performance bounds

  3. Motivation � Internet backbone and PoPs are over-engineered Overcome link failures � Underutilized (multiple routes exist) � � Current Protocols Typically find shortest path(s) � Do not directly minimize � delay through PoPs

  4. Background � IS-IS & OSPF Link-state routing protocols � Limited load balancing � Manually tuned to a few routes (traffic engineering) � � MPLS Re-labels packets in the internal network � � Previous work Resource Allocation (minimize max link utilization) � Routing Heuristics �

  5. Informal Formulation � Optimal network flow allocation Decide how to distribute packets from a particular � flow across the network links ( x variables) Satisfy conservation laws � � Definition of a flow Aggregate flow to each destination (sink) node � Every other node can be a source to the sink node � Source-sink vector �

  6. Mathematical Formulation Convex cost function for each link i � Total link i traffic ( x is flow’s traffic) � Node-link incidence matrix – A � Link capacity vector – C �

  7. Piece-wise Linear Approximation � Goal: Convert problem into LP Approximate convex function by K (PWL) segments � Epigraph minimization ( p variables) �

  8. Cost Function Minimize delay over all links � Delay for one link � 10 M/M/1 queue delay 9 � 8 7 6 5 4 A convex function � 3 Problem � 2 1 Complex algorithms � 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Slow convergence �

  9. PWL Approximation MSE s plit 10 � How to approximate? 9 8 7 Uniform � 6 5 MSE � 4 3 Min-Max � 2 1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Min-Max s plit Uniform split 10 9 8 8 7 7 6 6 5 5 4 4 3 3 2 2 1 1 0 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

  10. PWL Optimization Algorithm � Centralized “one-shot” algorithm ( K > 100 ) Computational intensive � Very accurate results for underutilized networks � � Centralized iterative algorithm ( K < 10 ) Solve LP for the given K (start) � Identify link segment k = 1,…,K * that contains traffic flow � Split marked link segment into K more segments � Update LP constraints (slopes and intercepts) � Repeat until stopping criteria satisfied �

  11. Algorithm Simulation � Experimental Setup MATLAB linprog() � Sprint IP backbone network topology � Traffic matrix � Uniform traffic � Sparsity pattern �

  12. Results � Uniform traffic, unit capacities, ( K = 2, 3, 5 ) objective value optimality gap 74 2 10 72 1 10 70 68 0 10 66 64 -1 10 62 60 -2 10 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8

  13. Results � Iterations (how many?) Stopping criteria � total cumulative iterations 250 � Feasability 10E-6 � 200 � Convergence 150 Always finds � feasible solution 100 (if one exists) 50 0 1 2 3 4 5 6 7 8

  14. More Results � Gaussian traffic with objective value 29.6 29.5 sparsity pattern, unit C 29.4 29.3 optimality gap 29.2 0 10 29.1 29 28.9 -1 10 28.8 28.7 1 2 3 4 5 6 7 8 total cumulative iterations 200 180 -2 10 160 140 120 100 -3 10 80 60 40 20 -4 10 0 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8

  15. Even More Results objective value � Heavy Gaussian traffic 85 with sparsity, random C 80 75 optimality gap 2 10 70 65 1 60 10 1 2 3 4 5 6 7 total cumulative iterations 200 180 160 140 120 0 10 100 80 60 40 20 -1 10 0 1 2 3 4 5 6 7 1 2 3 4 5 6 7

  16. Traffic Distribution Total flow allocation 3 13 2 1 9 14 8 1 2 12 15 3 10 0 16 17 4 -1 7 5 -2 18 11 6 -3 -3 -2 -1 0 1 2 3

  17. Computational Complexity � LP interior-point algorithms number of arithmetic operations � number of iterations � � M is number of variables + inequality constraints � For our problem: M = LF + LK + L � For i iterations = iL(K+F+i/2+3/2) �

  18. Storage Complexity � Memory storage requirements � F – Flows � L – Links � N – Nodes � K – number of segments � � (KLFN + N + KLF) α ≈ KLF(N + 1) α � i th iteration: (K + i)LF(N + 1) α �

  19. Distributed algorithm � Centralized implementation � Easily distributed (especially PWL approximation) � Dual methods Subgradient ascent � Lagrangian relaxation � � Path augmentation approach

  20. Protocol Implementation � Routing protocols MPLS-like labeling � At most M flows � M – Number of edge routers or PoPs � DEST determines p i � DEST IP PACKET p 1 p 2 p 3

  21. Edge Routers � Full look-up DEST – ID of edge router where packet leaves PoP � Identifies flows within PoP � � Congestion Control All congestion control can be done at the edges � Detect when traffic not admissible – not feasible � Drop packets at edge � Estimate flows – recalculation at substantial change � Should not occur often – large aggregation of flows �

  22. Conclusion � Most links underutilized (IP backbone) � But still cannot guarantee performance ( e.g. delay) � Optimal network flow allocation could help � We suggest a practical algorithm Converges to near optimal solutions � Few iterations � Standard LP � Can make it distributed � � Special thanks to Yashar Ganjali for all his help!!

  23. Questions?

  24. PWL Approximation (1) � Convert PWL problem into LP

Recommend


More recommend