1. VirtualKnotter: Online Virtual Machine Shuffling for Congestion Resolving in Virtualized Datacenter Xitao Wen, Kai Chen, Yan Chen, Yongqiang Liu, Yong Xia, Chengchen Hu 1

2. Datacenter as Infrastructure 2

3. Congestion in Datacenter 10:1~100:1 2:1~10:1 Packet loss! Queuing delay! Degrading Throughput! 3

4. Congestion in the Wild 4 General Approaches Problem Formulation Main Design Evaluation

5. Spatial Pattern Unbalanced utilization – Hotspot: Hot links account for <10% core links [IMC10] – Spatially unbalanced utilization 5 Sender Receiver

6. Temporal Pattern Long congestion event – lasts for 10s of minutes – Individual event has clear spatial pattern 6 Core Link Index

7. Traffic Stability Bursty at a fine granularity – Not predictable at 10s or 100s or milliseconds [IMC10][SIGCOMM09] Predictable at timescale of 10s of minutes – 40% to 70% pairwise traffic can be expected stable – 90%+ predictable traffic aggregated at core links 7

8. 8 General Approaches Problem Formulation Main Design Evaluation Congestion in the Wild

9. General Approaches Network Layer – Increase network bandwidth Fat-tree, BCube, OSA… – Optimize flow routing Hedera, MicroTE Application Layer – Optimize VM placement Expensive Requires to upgrade entire DC network Expensive Requires to upgrade entire DC network Not scalable Requires hardware support Depends on rich path diversity Not scalable Requires hardware support Depends on rich path diversity Scalable Lightweight deployment Suitable for existing over- subscribed network Scalable Lightweight deployment Suitable for existing over- subscribed network 9

10. Virtualization Layer VM Live Migration – Keep continuous service while migrating – 1.1x – 1.4x VM memory transfer Server VM Server DC Network VM Major Cost! 10 Background on Virtualized DC

11. Optimize VM Placement Offload traffic from congested link active VM idle VM 11

12. Congestion in the Wild General Approaches Problem Formulation 12 Main Design Evaluation

13. Design Goal Mitigate congestion – Maximum link utilization (MLU) Controllable migration traffic (i.e. moving VM) – Less than reduced traffic Reasonable runtime overhead – Far less than target timescale (10s of mins) Objective Constraint 13

14. Problem Statement Input – Topology and routing of physical servers – Traffic matrix among VMs – Current Placement Variable & Output – Optimized Placement NP-hardness – Proof: reduced from Quadratic Bottleneck Assignment Problem 14

15. Related Work Optimize VM placement – Server consolidation [SOSP07] – Fault tolerance [ICS07] – Network scalability [INFOCOM10] 15

16. Main Design 16 Evaluation Congestion in the Wild General Approaches Problem Formulation

17. Inspiration Stretch the tie violently, making it loose and less tangled. Solve each tie gently, by carefully reeving the end out of the tie. 17

18. Two-step Algorithm Fast and greedy Search for localizing overall traffic May stuck in local minimum Fast and greedy Search for localizing overall traffic May stuck in local minimum Fine-grained and randomized Search for mitigating traffic on the most congested links Help avoid local minimum Fine-grained and randomized Search for mitigating traffic on the most congested links Help avoid local minimum 18

19. Multiway Θ-Kernighan-Lin (KL) Top-down graph cut improvement Introduce Θ to limit # of moves O(n 2 log(n)) Top-down graph cut improvement Introduce Θ to limit # of moves O(n 2 log(n)) 19

20. Multiway Θ-Kernighan-Lin (KL) Top-down graph cut improvement Introduce Θ to limit # of moves O(n 2 log(n)) Top-down graph cut improvement Introduce Θ to limit # of moves O(n 2 log(n)) 20

21. Multiway Θ-Kernighan-Lin (KL) Top-down graph cut improvement Introduce Θ to limit # of moves O(n 2 log(n)) Top-down graph cut improvement Introduce Θ to limit # of moves O(n 2 log(n)) 21

22. MLU=.60 MLU=.53 Simulated Annealing Searching (SA) Randomized global searching Terminate when obtains satisfied solution, or predefined max depth is reached Randomized global searching Terminate when obtains satisfied solution, or predefined max depth is reached 22

23. Evaluation 23 Congestion in the Wild General Approaches Problem Formulation Main Design

24. Methodology Baseline Algorithm – Clustering-based algorithm – Pro: best-known static optimality – Con: high runtime and migration overhead Metrics – MLU reduction without migration overhead – Overhead Migration traffic Runtime overhead – Simulation results 24

25. MLU Reduction without Overhead 25 VirtualKnotter demonstrates similar static performance as that of Clustering.

26. Migration Traffic 26 VirtualKnotter shows significantly less migration traffic than that of Clustering.

27. Runtime Overhead 27 VirtualKnotter demonstrates reasonable runtime overhead.

28. Simulation Results 53% less congestion 28 Altogether, VirtualKnotter obtains significant gain on congestion resolving.

29. Conclusions Collaborative VM migration can substantially resolve long-term congestion in DC Trade-off between optimality and migration traffic is essential to harvest the benefit DC networking projects of Northwestern LIST: http://list.cs.northwestern.edu/dcn 29

30. Thank you! 30

31. Backup 31

32. General Approaches Cost Hardware Support Scalability Other Dependency Increase Bandwidth HighYesVaries Optimize Routing LowYesLow Rich path diversity Optimize VM Placement LowNoHigh VM deployment 32

33. Problem Statement Objective – Minimize Maximum Link Utilization (MLU) – Cool down the hottest spot Constraints – Migration traffic – Server hardware capacity – Inseparable VM NP-hardness – Proof: reduced from Quadratic Bottleneck Assignment Problem 33

34. Observation Summary Unbalanced jam (spatial) Long-term congestion (temporal) Predictable at 10s of minutes scale (stability) 34

35. Two-step Algorithm Multiway Θ-Kernighan-Lin Algorithm (KL) Fast search for approximation Simulated Annealing Searching (SA) Fine search for better solution 35