Sunday, September 21, 2008

Blah blah blah, optics, blah blah blah, router.

I nominate Scaling Internet Routers Using Optics as the least fun paper to read so far this semester. They describe in excruciating detail two designs of multi-rack routers (1-hybrid electro-optical, 2-fully optical switch) which achieve 100Tb/s load balanced routing. Their router gives 100% throughput without a scheduler. They address four obstacles to the load balanced switching approach:
  1. Rapidly configuring switch fabric is slow or expensive.  Solution: replace switch fabric with fixed mesh of optical channels
  2. Packets can be mis-sequenced.  Solution: (Full Ordered Frames First) bound the length of queues at middle stage and introduce a re-sequencing buffer at a later stage
  3. Periodic traffic can cause bad throughput.  Solution: same as for problem #1 (replace crossbar switches by fixed optical mesh)
  4. Missing/failed linecards are not easily handled.  Solution: presented in sections 6.2 and 6.3, they present two solutions that allow the switch to be reconfigured to scatter traffic uniformly over all present linecards in the face of some linecard failures.
The are designing for ~3 years out (in 2003), I wonder if the future aligned with their predictions. 

The paper was a bit dense on technical/hardware details so I ended up being lost for most of it :-/

1 comment:

Randy H. Katz said...

Interesting reaction ... but it indicates that most students who study networking don't have much insights into what goes on inside the router. In retrospect, I think the paper might have made more sense if it had been read after the more tutorial paper that described a crossbar interconnect on iSlip scheduling scheme.