PortLand : A Scalable Fault-Tolerant Layer 2 Data Center Network Fabric

1. PortLand: A Scalable Fault-Tolerant Layer 2Data Center Network Fabric B97703099財金三婁瀚升

2. Outline • Introduction • Background • Design • Implementation • Conclusion

3. Introduction • LAN insufficiency • Requirement Network in datacenter : • VM Migration without IP change • Needn’t config switch before deployment • Host in datacenter efficiently communicate • No forwarding loop • Fault-tolerant and recovery

4. Introduction • Layer 2 kind fabric • Layer 3 not workable • VM migration with IP change • Config switches of subnet and DHCP synchronite • TTL method: prevent forwarding loop • Prevent routing broadcast

5. Background:DatacenterNetwork • Topology • Forwarding • Layer 3 : IP assigned hierachically • Broadcast (Failure avoidance, overhead ) • Config switch subnet & DHCP syncro. Fault • VM migration unable • Layer 2 : MAC • Single spanning tree problem (not shortest way) • Broadcast entire fabric • VLAN : • Source pre-assignment (decrease flex. & scal.) • Switch need to maintain VLAN’s state

6. Background:DatacenterNetwork • End host Virtualization • Layer 3 setting : not work • ARP to solute ?

7. Fat Tree Network • Multi-rooted • Stage : edge, aggregation, core • K-port swithes: • k3/4 end hosts • 5k2/4 individual k-port switches • k individual pods • each pod : k2/4 hosts

9. Design : Fabric Manager • Centralized Manager • Fuction : • Mantain soft state network config (ex.topology) • Responsible for • ARP resolution • Fault tolerance • Multi-cast • Only soft state (no hard state, ex.#of switch)

10. Design : Pseudo MAC • Assigned to end hosts • Information include: • Location : same pod, same prefix • Pod number→ position numer • End host with their own AMAC • LDP (location discovery protocol) • employed to assign • Pod.position.port.vmid

12. Design : Proxy-based ARP • Ethernet : B-cast to all host(same layer2) • Used for communication in datacenter • If FM not available for IP-PMAC mapping →bcat to core(O(k) state ) • VM migration supporting • FM sending invalidation message to old pos. • If contact with this message → new PMAC address in the host’s cache

14. Design : Distributed Location Discovery • Location Discovery Protocol (LDP) • No administative config (no manul set) • Location Discovery Message: • Sent by switches • Several information • Edge got from aggregation, learned →Aggregation learned →Core learned

15. Design : Distributed Location Discovery • Location Discovery Message: • Position number acquisition: • Randomly chosen number • Verified by aggregation • Pod number acquisition: • FM assignment to aggregation • Exception : non-existence • LDM not correct • Disable suspicious port

16. Design:Loop free Forwarding • Prevent using Spanning tree • Downward seperate from Upward

17. Design : Fault Tolerance • Unicast Fault Detection and Action • Multicast Fault Detection and Action

21. Deisign : Comparison

22. Implementation : System

23. Implementation :Evaluation

24. Implementation :Evaluation

25. Implementation :Evaluation

26. Implementation :Evaluation

27. Implementation :Evaluation

28. Implementation :Evaluation

29. Conclusion • Commercial Use • Datacenter Network Protocol