1. 20
  1.  

  2. 5

    I agree with the general point about the ubiquity of networks, but I strongly disagree with the author’s “law” on the convergence to tiered-star topologies.

    Looking at early switched networks, key innovations were the non-blocking Clos topology in the 50s and then the Benes network in the 60s; these are still used today in many telephone switches.

    Looking at processor interconnect, we went from simple busses to 2-d mesh and torus networks (for ease of implementation). n-cube / hypercube networks came next because they’re the obvious next step, and minizming the diameter is an “obvious” optimization target. A couple papers came out that showed that low-dim topologies were better given the constraints of the time, which motivated a move back towards 2-d and 3-d mesh and torus networks, though the degree has snuck back up as physical constraints have changed. And lately, fabrics with high node-degree have gone to butterfly and Clos networks.

    This comment is long enough as-is, but suffice to say that you could make similar comments about other networked applications. It’s complicated, there are a lot of tradeoffs, and it’s not always (or even usually) the right decision to go with a star topology.

    I suppose this all depends on what you mean by tiered-star. IMO, this doesn’t look much like this, let alone something like this.