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

Summary

This paper discussed layer 2 protocols that would be optimal for data centers. The goals they try to achieve are i) VM's should be easily migratable without changing IP addresses, ii) no admin configuration required, iii) every node should be able to communicate with every other node, iv) no forwarding loops and v) recover from faults quickly.

PortLand uses a centralized fabric manager that maintains soft state about the network. Nodes uses a psuedo MAC address instead of their actual MAC address. The fabric manager is also able to reduce the overhead that results from broadcasting ARPs.

The authors compare PortLand to 2 other architectures with similar goals: TRILL and SEATTLE. Finally, they run experiments on their testbed to evaluate PortLand using the following criteria: convergence time with increasing faults, tcp convergence, multicast convergence, scalability and VM Migration.

Criticism & Questions

The authors do several experiments to evaluate the criteria given above and present those results quantitatively. However, it would be helpful to know how those numbers compare to what currently happens (ie. they say that after vm migration, communication resumes after 32 seconds - how does this compare to current vm migration?)