1. 1/25 Generic and Automatic Address Configuration for Data Center Networks 1 Kai Chen, 2 Chuanxiong Guo, 2 Haitao Wu, 3 Jing Yuan, 4 Zhenqian Feng, 1 Yan Chen, 5 Songwu Lu, 6 Wenfei Wu 1 Northwestern University, 2 Micrsoft Research Asia, 3 Tsinghua, 4 1 Northwestern University, 2 Micrsoft Research Asia, 3 Tsinghua, 4 NUDT, 5 UCLA, 6 BUAA SIGCOMM 2010, New Delhi, India

2. 2/25 Motivation Address autoconfiguration is desirable in networked systems Manual configuration is error-prone 50%-80% network outages are due to manual configuration DHCP for layer-2 Ethernet autoconfiguration Address autoconfiguration in data centers (DC) has become a problem Applications need locality information for computation New DC designs encode topology information for routing DHCP is not enough - no such locality/topology information

3. 3/25 Research Problem Given a new/generic DC, how to autoconfigure the addresses for all the devices in the network? DAC: data center address autoconfiguration

4. 4/25 Outline Motivation Research Problem DAC Implementation and Experiments Simulations Conclusion

5. 5/25 DAC Input Blueprint Graph (G b ) A DC graph with logical IDs Logical ID can be any format Available earlier and can be automatically generated Physical Topology Graph (G p ) A DC graph with device IDs Device ID can be MAC address Not available until the DC is built and topology is collected 10.0.0.3 00:19:B9:FA:88:E2

6. 6/25 DAC System Framework Physical Topology Collection Device-to-logical ID Mapping Logical ID Dissemination Malfunction Detection

7. 7/25 Two Main Challenges Challenge 1: Device-to-logical ID Mapping Assign a logical ID to a device, preserving the topological relationship between devices Challenge 2: Malfunction Detection Detect the malfunctioning devices if the physical topology is not the same as blueprint (NP-complete and even APX-hard)

8. 8/25 Roadmap Physical Topology Collection Device-to-logical ID Mapping Logical ID Dissemination Malfunction Detection

9. 9/25 Device-to-logical ID Mapping How to preserve the topological relationship? Abstract DAC mapping into the Graph Isomorphism (GI) problem The GI problem is hard: complexity (P or NPC) is unknown Introduce O 2 : a one-to-one mapping for DAC O 2 Base Algorithm and O 2 Optimization Algorithm Adopt and improve techniques from graph theory

10. 10/25 O 2 Base Algorithm G b : {l1 l2 l3 l4 l5 l6 l7 l8} G p : {d1 d2 d3 d4 d5 d6 d7 d8} G b : {l1} {l2 l3 l4 l5 l6 l7 l8} G p : {d1} {d2 d3 d4 d5 d6 d7 d8} G b : {l1} {l5} {l2 l3 l4 l6 l7 l8} G p : {d1} {d2 d3 d5 d7} {d4 d6 d8} Decomposition Refinement

11. 11/25 O 2 Base Algorithm G b : {l1 l2 l3 l4 l5 l6 l7 l8} G p : {d1 d2 d3 d4 d5 d6 d7 d8} G b : {l5} {l1 l2 l3 l4 l6 l7 l8} G p : {d1} {d2 d3 d4 d5 d6 d7 d8} G b : {l5} {l1 l2 l7 l8} {l3 l4 l6 } G p : {d1} {d2 d3 d5 d7} {d4 d6 d8} G b : {l5} {l1 l2 l7 l8} {l6} {l3 l4} G p : {d1} {d2 d3 d5 d7} {d6} {d4 d8} Decomposition Refinement

12. 12/25 O 2 Base Algorithm G b : {l5} {l6} {l1 l2} {l7 l8} {l3 l4} G p : {d1} {d6} {d2 d7} {d3 d5} {d4 d8} G b : {l5} {l6} {l1} {l2} {l7 l8} {l3 l4} G p : {d1} {d6} {d2} {d7} {d3 d5} {d4 d8} G b : {l5} {l6} {l1} {l2} {l7} {l8} {l3} {l4} G p : {d1} {d6} {d2} {d7} {d3} {d5} {d4} {d8} Refinement Decomposition Decomposition & Refinement

13. 13/25 O 2 Base Algorithm O2 base algorithm is very slow for 3 problems: P1: Iterative splitting in Refinement: it tries to use each cell to split every other cell iteratively G p : π 1 π 2 π 3 …… π n-1 π n P2: Iterative mapping in Decomposition: when the current mapping is failed, it iteratively selects the next node as a candidate for mapping P3: Random selection of mapping candidate: no explicit hint for how to select a candidate for mapping

14. 14/25 O 2 Optimization Algorithm Heuristics based on DC topology features Sparse => Selective Splitting (for Problem 1) Symmetric => Candidate Filtering via Orbit (for Problem 2) Asymmetric => Candidate Selection via SPLD (Shortest Path Length Distribution) (for Problem3) We propose the last one and adopt the first two from graph theory R1: A cell cannot split another cell that is disjoint with itself. R2: If u in G b cannot be mapped to v in G p, then all nodes in the same orbit with u cannot be mapped to v either. R3: Two nodes u, v in G b, G p cannot be mapped to each other if have different SPLDs.

15. 15/25 Speed of O 2 Mapping 12.4 hours 8.9 seconds

16. 16/25 Roadmap Physical Topology Collection Device-to-logical ID Mapping Logical ID Dissemination Malfunction Detection

17. 17/25 Malfunction Detection Types of Malfunctions Node failure, Link failure, Miswiring Effects of Malfunctions O 2 cannot find device-to-logical ID mapping Our Goal Detect malfunctioning devices Problem Complexity An ideal solution 1.Find Maximum Common Subgraph (MCS) between G b and G p say G mcs 2.Remove G mcs from G p => the rest are malfunctions MCS is NP-complete and even APX-hard

18. 18/25 Practical Solution Observations Most node/link failures, miswirings cause node degree change Special, rare miswirings happen without degree change Our Idea Degree change case: exploit the degree regularity in DC Devices in DC have regular degrees (common sense) No degree change case: probe sub-graphs derived from anchor points, and correlate the miswired devices using majority voting Select anchor point pairs from 2 graphs probe sub-graphs iteratively, stop when k-hop subgraphs are isomorphic but (k+1)-hop are not, increase the counters for k- and (k+1)- hop nodes Output node counter list: high counter => high possible to be miswired Isomorphic 1 2 k+1 k 1 2 k Isomorphic …… Non-Isomorphic

19. 19/25 Simulations on Miswiring Detection Over data centers with tens of thousands of devices with 1.5% nodes as anchor points to identify all hardest-to-detect miswirings 1.5%

20. 20/25 Roadmap Physical Topology Collection Device-to-logical ID Mapping Logical ID Dissemination Malfunction Detection

21. 21/25 Basic DAC Protocols CBP: Communication Channel Building Protocol Top-Down, from root to leaves PCP: Physical Topology Collection Protocol Bottom-Up, from leaves to root LDP: Logical ID Dissemination Protocol Top-Down, from root to leaves DAC manager: 1.handle all the intelligences 2.can be any server in the network

22. 22/25 Implementation and Experiments Over a BCube(8,1) network with 64 servers 1.Communication Channel Building ( CCB ) 2.Transition time 3.Physical Topology Collection (TC) 4.Device-to-logical ID Mapping 5.Logical IDs Dissemination (LD) The total time used: 275 milliseconds

23. 23/25 Simulations Over large-scale data centers (in milliseconds) 46 seconds for the DCell(6, 3) with 3.8+ million devices

24. 24/25 Summary DAC: address autoconfiguration for generic data center networks, especially when the address is topology-aware Graph isomorphism for address configuration 275ms for a 64-sever BCube, and 46s for a DCell with 3.8+ million devices Anchor point probing for malfunction detection with 1.5% nodes as anchor points to identify all hardest-to- detect miswirings DAC is a small step towards the more ambitious goal of automanagement of the whole data centers

25. 25/25 Q & A? Thanks!