Wednesday, October 21, 2009

RouteBricks: Exploiting Parallelism to Scale Software Routers

RouteBricks: Exploiting Parallelism to Scale Software Routers, Mihai Doberscu, Norbert Egi, Katerina Argyrako, Byung-Gon Chun, Kevin Fall, Gianluca Iannaccone, Allan Knies, Maziar Manesh, Sylvia Ratnasamy

This paper got Best Paper award at SOSP, and I got pointed to it by a blog post here. Actually, I believe that Katerina and/or Sylvia discussed the concept of RouteBricks during some lunch during SIGCOMM in Barcelona, so I kinda had heard of it. Congrats for the award!

Paper is deserving of it, for sure. It presents an architecture to create a router using commodity hardware (well, sophisticated commodity hardware, prototype Nehalem servers, you can't build this in your garage yet) to achieve the performance of a Cisco router at a fraction of the cost. SOSP has the slides and the recorded audio of the presentation on the website, which is pretty neat.

Most of the details are in how to avoid the bottlenecks in the architecture to spread the traffic over several servers rather than into a choking point. They use valiant load balancing, and a multi-hop butterfly topology to connect them. They also need some optimization in the processing to get to faster rates (multi-queue NIC, batch processing).

My thoughts reading this was: what a good idea, and what a nicely written paper! Also, what kind of reliability do you get from this platform (I would assume Cisco routers have some up-time requirements, fault tolerance, hot swappable line cards, whatever, to keep the ISP happy); I was thinking about reliability from the cellular network operator point of view (since I was thinking if this router design could translate into, say, a commodity base station. Of course, on 2nd thought, what un-commoditize the BS is the air interface, not the switching part of it). I also would like to know what Cisco margins are, since they can only compare the cost of their router to a list price.

No comments: