Ethan J. Jackson

Flow caching for high entropy packet fields

Packet classification on general purpose CPUs remains expensive regardless of advances in classification algorithms. Unless the packet forwarding pipeline is both simple and static in function, fine-tuning the system for optimal forwarding is a time-consuming and brittle process. Network virtualization and network function virtualization value general purpose CPUs exactly for their flexibility: in such systems, a single x86 forwarding element does not implement a single, static classification step but a sequence of dynamically reconfigurable and potentially complex forwarding operations. This leaves a software developer looking for maximal packet forwarding throughput with few options besides flow caching. In this paper, we consider the problem of flow caching and more specifically, how to cache forwarding decisions that depend on packet fields with high entropy (and therefore, change often); to this end, we arrive at algorithms that allow us to efficiently compute near optimal flow cache entries spanning several transport connections, even if forwarding decisions depend on transport protocol headers.

Taking an AXE to L2 Spanning Trees

I think that I shall never see a structure more wasteful than a tree. Most links remain idle and unused while others are overloaded and abused. And with each failure comes disruption caused by the ensuing tree construction. Thus, L2 must discard its spanner, requiring flooding in a different manner. For the trees fragile waste to be abated, trim no branches and detect packets duplicated. (With apologies to Radia Perlman and Joyce Kilmer.)