DESIGN: Avoid an exponential number of ETGs

[Aaron Gember-Jacobson]

How can we avoid the need to create a separate ETG for every traffic class, which results in s^2 (s = the number of subnets) ETGs?

[Aaron Gember-Jacobson]

Aditya suggested we might leverage network symmetry: compare parts of ETGs that are supposed to be symmetrical, and note critical differences (e.g., some edge is present due to a missing ACL). How do we determine which parts of the network should be symmetrical?

[Aaron Gember-Jacobson]

We should generate control plane ETGs and data plane ETGs. Control plane ETGs are created for each destination, because forwarding table lookups are based on the packet's destination. Data plane ETGs are created for source/destination pairs which have different policies (e.g., ACLs) than traffic from other sources to the same destination.

[Aaron Gember-Jacobson]

Finished in commit fd3da15