Routers with two switched backplanes and optics

Isaac Keslassy, Shang-Tse Chuang, Kyoungsik Yu, David Miller, Mark Horowitz, Olav Solgaard, Nick McKeown, "Scaling Internet Routers Using Optics"

This paper blew my mind. 100Terrabit/s.

• When I read the other paper on routers, I noticed that there are two things about a router we might want to increase: the number of ports N or the per port/total throughput.The other paper still has a centralized structure: the scheduler. Luckily, because whole cells are transfered through the crossbar before reconfiguration, the scheduler is given a relative high amount of time to come up with a scheduling decision, but I don't see how the scheduler can be implemented in a way which has run time less than O(N^2), so the router would not really scale in N.
• This paper actually analyzes how the components scale with N and overcomes the centralized scheduler by introducing a "load balancer" before the VOQ (see previous post): it has two crossbars (though they can be replaced by one at twice the speed) in total, each crossbar follows a fixed scheme which cyclically permutes every input with every output, as shown in Figure 2. They show that this will bring 100% throughput. I think memory scales with N^2, but not sure.
• The paper next analyzes how to use an optical backplane. The suggested switched backplane is a two-tier hierarchy of optical switches. I lost track when what optical switch is doing what to which frequency in WDM.
  
• The paper discusses how to use this scheme so that some line cards can be removed (the VOQ are distributed across line cards, so if a line card is missing or failed, traffic needs to rerouted around it)