1. Virtual Network Embedding with Coordinated Node and Link Mapping N. M. Mosharaf Kabir Chowdhury Muntasir Raihan Rahman and Raouf Boutaba University of Waterloo

2. Motivation Network Virtualization – Coexistence of multiple virtual networks (VNs) over a shared substrate – Applications Research experiments Future Internet architecture (e.g., Clean-slate) Major challenge – Efficient assignment of substrate network resources to virtual networks requirements June 4, 20142INFOCOM 2009 @ Rio de Janeiro, Brazil

3. VN Embedding Problem Given: – Single substrate network: G S = (N S, E S ) – Online VN requests: G V = (N V, E V ) – Requirements and Constraints of virtual nodes and virtual links Task: – Assign virtual nodes and links to substrate nodes and links – Allocate resources CPU, bandwidth June 4, 20143INFOCOM 2009 @ Rio de Janeiro, Brazil

4. VN Embedding Objectives Maximize – Acceptance ratio Percentage of request accepted – Revenue Based on resources requested for a VN Minimize – Cost Based on substrate network resources allocated for embedding VN requests June 4, 20144INFOCOM 2009 @ Rio de Janeiro, Brazil

5. Virtual Network Embedding 5 C AB DEF GH 60 8055 50 7065 85 90 a bc 10 e fd 20 a bc d e f 10 12 55 22 15 12 10 15 27 17 20 25 June 4, 2014 NP-hard INFOCOM 2009 @ Rio de Janeiro, Brazil

6. Observations Existing heuristics: 1.Disjoint node and link mapping 2.Ignore (completely) location constraints on virtual nodes Our approach – Node mapping influences link mapping – Location constraints take the front seat Meta-VN request June 4, 20146INFOCOM 2009 @ Rio de Janeiro, Brazil

7. Substrate Graph Augmentation 7 C AB DEF GH 60 8055 50 7065 85 90 22 15 12 10 15 27 17 20 25 a bc 10 a b c 12 June 4, 2014INFOCOM 2009 @ Rio de Janeiro, Brazil

8. Objective Flow variables: – Multi-commodity flow constraints Binary variables: – Exactly one substrate node is selected for each meta-node – At most one meta-node is mapped onto a substrate node Mixed Integer Program June 4, 20148INFOCOM 2009 @ Rio de Janeiro, Brazil α, β: Tuning parameters R: Remaining capacity

9. Mixed Integer Program Solution 9 C AB DEF GH 60 8055 50 7065 85 90 22 15 12 10 15 27 17 20 25 a bc 10 a b c 12 June 4, 2014 NP-hard INFOCOM 2009 @ Rio de Janeiro, Brazil

10. LP Relaxation Relax the binary constraints on the x variable Problem 1: x values do not select single substrate nodes for every meta-node – Use rounding techniques (e.g., deterministic, randomized) Problem 2: x values are inconsistent – Use the product of x and f while rounding June 4, 201410INFOCOM 2009 @ Rio de Janeiro, Brazil

11. FINALIZATION LINK MAPPING NODE MAPPING INITIALIZATION ViNEYard (D-ViNE & R-ViNE) 11 For each VN request: – Augment the substrate graph – Solve the resulting LP – For each virtual node: Calculate the probability for each meta-node to be selected for the corresponding virtual node Selection: – D-ViNE: Select the meta-link with the highest probability – R-ViNE: Select a meta-link randomly with the calculated probability – Use MCF to map virtual edges – If the VN request is accepted Update residual capacities of the substrate resources June 4, 2014INFOCOM 2009 @ Rio de Janeiro, Brazil

12. Performance Evaluation Compared Algorithms Description D-ViNEDeterministic Node Mapping with Splittable Link Mapping using MCF R-ViNERandomized Node Mapping with Splittable Link Mapping using MCF G-SPGreedy Node Mapping with Shortest Path Based Link Mapping G-MCFGreedy Node Mapping with Splittable Link Mapping using MCF D-ViNE-SPDeterministic Node Mapping with Shortest Path Based Link Mapping D-ViNE-LBDeterministic Node Mapping with Splittable Link Mapping using MCF, where α uv = β w = 1, for all u, v, w Є N S June 4, 201412INFOCOM 2009 @ Rio de Janeiro, Brazil

13. Simulation Setup Substrate network – 50 nodes in a 25x25 grid with 0.5 link probability – CPU/BW uniformly distributed in the range: 50-100 units VN requests – Poisson arrival rates from 4 VN requests 100 time units – Exponentially distributed lifetime of 1000 time units – 2-10 nodes with 0.5 link probability Tools: GT-ITM (Georgia Tech Internet Topology Models), GLPK (GNU Linear Programming Toolkit) June 4, 2014INFOCOM 2009 @ Rio de Janeiro, Brazil13

14. Acceptance Ratio June 4, 201414INFOCOM 2009 @ Rio de Janeiro, Brazil

15. Revenue Vs Cost RevenueCost June 4, 201415INFOCOM 2009 @ Rio de Janeiro, Brazil

16. Link Utilization June 4, 201416INFOCOM 2009 @ Rio de Janeiro, Brazil

17. Conclusions ViNEYard Algorithms – Improved correlation between the node mapping and the link mapping phases – Increased acceptance ratio and revenue with decreased cost Ongoing and Future Work – Window-based extension to D-ViNE and R-ViNE a.k.a. WiNE – Extension to inter-domain scenario a.k.a. PolyViNE – Approximation factors for D-ViNE and R-ViNE June 4, 201417INFOCOM 2009 @ Rio de Janeiro, Brazil

18. June 4, 201418 Thank You! Questions? INFOCOM 2009 @ Rio de Janeiro, Brazil http://www.mosharaf.com/

19. BACKUP SLIDES June 4, 201419INFOCOM 2009 @ Rio de Janeiro, Brazil

20. June 4, 201420INFOCOM 2009 @ Rio de Janeiro, Brazil

21. June 4, 201421INFOCOM 2009 @ Rio de Janeiro, Brazil

22. D-ViNE June 4, 201422INFOCOM 2009 @ Rio de Janeiro, Brazil

23. R-ViNE June 4, 201423INFOCOM 2009 @ Rio de Janeiro, Brazil

24. Resource Utilization Node UtilizationLink Utilization June 4, 201424INFOCOM 2009 @ Rio de Janeiro, Brazil

25. Effect of Increasing Load 25June 4, 2014INFOCOM 2009 @ Rio de Janeiro, Brazil