1. DECOR: A Distributed Coordinated Resource Monitoring System Shan-Hsiang Shen Aditya Akella

2. Outline Background and movtivation Challenge How DECOR works Implementation Evaluation results

3. Why resource coordination? Optimize network traffic Resources is limited in different nodes Resource coordination

4. Why resource coordination? Redundancy elimination: EndRE [Aggarwal’10] – Node capacity (decoding): 5 Unit 4 Encode decode Traffic: 10 Unit 1.Can only remove half of total redundancy 2.Still have resource available along the paths Motivation 1: Need coordination to use resources more efficiently Motivation 1: Need coordination to use resources more efficiently

5. Centralized Resource Coordination Encode Decode 5 Traffic: 10 Unit SmartRE [Anand’08] – Node capacity (decoding): 5 Unit Decode: 5 Unit A central control become bottleneck and a single point of failure Motivation 2: A distributed framework Motivation 2: A distributed framework

6. Challenge Lack of a central controller to gather global view and make decision 6 Local Optimization Global Optimization

7. DECOR Overview 7 Goal: assign job to each node to increase global benefit without using out resources

8. DECOR Overview 8 Optimize local benefit Rearrange resource assignment

9. Control Packets Two kinds of control packets sent periodically: – HELLO packets: collect accepted resources and traffic features along a path. – ACK packets: deliver job assignment to each node. HELLO ACK Ingress node egress node Calculate optimized job assignments Accepted resources 9 Traffic feature Resource arrangement

10. Resource Arrangement Interior nodes assign resources to different paths. Calculate for each path. Assign more resources to the path that can contribute more benefit. benefit Unit resource Ingress node Interior node Egress node

11. Job assignment optimization Optimize job assignment by linear programming – Maximize: total benefit along the path. – Subject to: resources needed cannot surpass accepted resources – Variable to be determined: the fraction of traffic each node needs to process 11 Ingress node Interior node Egress node

12. Implementation We implement distributed SmartRE (SmartRE + DÉCOR) in Click software router as modules. The Click modules is running in a desktop with Intel® Core™ 2 Quad CPU Q6700 and 3GB of memory. 12

13. The implementation issue of distributed SmartRE An ingress node synchronizes cache with interior nodes by using buckets. Ingress node Interior node 1 Interior node 2 egress node Bucket 1  interior node 1 Bucket 2  interior node 2 Bucket 3  egress node 13

14. Recovery from route failure Stop using invalid buckets. Ingress node Interior node 1 Interior node 2 egress node Bucket 1  interior node 1 Bucket 2  interior node 2 Bucket 3  egress node Interior node 3 Bucket 1  interior node 1 Bucket 2  interior node 2 Bucket 3  egress node Bucket 4  interior node 3 14

15. Evaluation setup Compare distributed SmartRE with the following approaches. – Hop-by-hop RE. – Centralized SmartRE. – Edge-based RE. – Ideal case. Use a real trace with 2GB traffic. 15

16. Evaluation result 16

17. Convergence time evaluation Topology (AS#)PoPs#Test Flows#Iterations NTT (2914)70346 Level 3 (3356)63305 Sprint (1239)52264 GEANT22103 Internet21143 17

18. Other Applications DECOR can apply to other path-based applications C S AMP [Sekar’08] – Coordinate resources to sample traffic – DECOR can provide a distributed solution

19. Conclusion Resource coordination to use resource more efficiently Distributed solution to avoid bottleneck and single point of failure problems DECOR can apply to multiple applications The performance can be as good as centralized solution in SmartRE case

20. THANK YOU QUESTION?

21. Backup slides

22. CSAMP setup NTTLevel3SprintGEANTInternet2 Flows(x 10 6 )514637168

23. CSAMP setup DECOR-based CSAMP: apply DECOR to coordinate resources Flow sampling: each node picks up one packet per 100 packets in each flow Packet sampling: each node picks up one packet per 100 packets of all traffic Edge packet sampling: edge nodes pick up one packet per 50 packets of all traffic Max sampling: each node samples as much as traffic as possible

24. Evaluation results

25. Job assignment optimization B p,r is the maxima benefit path p can provide in node r –Distributed SmartRE: B p,r = distance p,r × match p,q,r × matchlen p,q,r Constraint: – –Distributed SmartRE: – Maximize 25

26. Quota distribution Limited resource quota in each node. Paths go through the node request resource quota with the node. The node arrange its resource according to the benefit each path can provide. 26

27. Job assignment Multiple iterations are needed to converge. Total: 5 HELLO Total: 0 HELLO 5 5 5 ACK Assumption: Both of paths need 5 unit resources in total Path 2 can create more benefit for the network Path 1 Path 2 27

28. Job assignment Multiple iterations are needed to converge. Total: 5 HELLO Total: 0 HELLO 5 5 ACK Assumption: Both of paths need 5 unit resources in total Path 2 can create more benefit for the network Path 1 Path 2 28

29. Job assignment Multiple iterations are needed to converge. HELLO Total: 0Total: 5 ACK Assumption: Both of paths need 5 unit resources in total Path 2 can create more benefit for the network Path 1 Path 2 29

30. Job assignment Multiple iterations are needed to converge. Total: 5 HELLO Total: 0 HELLO 5 ACK HELLO 5 5 Total: 5 Assumption: Both of paths need 5 unit resources in total Path 2 can create more benefit for the network Path 1 Path 2 30

31. Job assignment Multiple iterations are needed to converge. Total: 5 HELLO Total: 0 HELLO 5 ACK Assumption: Both of paths need 5 unit resources in total Path 2 can create more benefit for the network Path 1 Path 2 31

32. Redundancy-Aware Routing Ingress nodes direct traffic to get more benefit from inter-path redundancy. Node 1 Node 2 Node 3 Route A Route B Route C Route D