Better router tech: Mind the flows, not the packets

In a recent article for the IEEE’s Spectrum magazine, Dr. Lawrence Roberts explains how the large routers that power the core of today’s Internet are doing it all wrong. They spend too much time processing each packet individually, then storing packets in a queue for during peak loads. This buffering makes VoIP calls and video streams stutter, and these routers use lots of hot-running and expensive memory, and they’re stuffed with specially-created chips. 

Roberts’ company Anagran has a different approach: only do the expensive work for the first packet in a flow, then treat subsequent packets in the same flow just like that first packet, doing away with queues and specially-designed chips in lieu of simple processors and cheap DRAM. (In the late 1960s, Roberts led the team that created the ARPANET, which would morph into the Internet that we know today in the following two decades.) As an added benefit, slowing down flows that go too fast can now be done with precision, rather than bluntly as in today’s routers. 

Upon reading Roberts’ article, the claims look rather extraordinary. Could it be that Cisco, Juniper and the other router vendors have been barking up the wrong tree for decades? In a word, no. Cisco has used numerous different router designs, including several that were fully or partially flow-based. Other vendors have built flow-based routers, too, such as Riverstone. For a while, there were also two-part multilayer switches, where a router would look at the first packet in a flow and then instruct a separate switch on how to switch packets belonging to the same flow.

Click here to read the rest of this article