Distributed Networks and Grad Students — Week 1 (ish)

A post: I’ve read the first 6 chapters of Where Wizards Stay up Late by Hafner and Lyon as well as a dozen or so wikipedia pages. This post concerns my thoughts on these.

What stuck me most about the early development of the ARPANET were the parallel discoveries of packet-switching (or distributed adaptive message block switching) by Donald Davies and Paul Baran. The idea was that instead of using circuit-switching like the telephone network where an uninterrupted circuit between two participants was held open for the entire length of the call, the next generation network (ARPANET, internet, etc.) should divide data into discrete packets and send them along their way, bouncing from node to node until they arrived at their destination, where they would be reassembled. This method called for a distributed network as Baran showed in 1964.

Baran came up with this idea because a distributed network would be more resistant to nuclear attacks than a centralized network. I think this somewhat silly because having a network that doesn’t fail after one node fails is nice even it’s not the Russians breaking it. The distributed model is necessary for a functioning network that doesn’t need constant maintenance at central hubs. This was Davies’s motivation. He wanted an efficient solution to the networking issue.

To me the fact that both came up with the same idea separately indicates that the distributed net wasn’t just an engineer’s solution to a given problem, but a scientific discovery of an optimal model.  Just like the wheel or arches, the distributed net is a concept beyond its physical applications (let’s not get any further into this as the reader may have majored in philosophy and actually have read Plato). The inventor of the arch may have wanted to add extra features, but the arch remains just a wheel.

Of course the above argument fails to hold after Baran and Davies as those were the last to discover the internet (a distributed network based on packet-switching). Next came the actual real internet aka ARPANET.

ARPA (as in DARPA before they put “defense” into everything) contracted out the work of designing and building their net to several different companies. BBN (Bolt, Beranek and Newman) was, at least in Hafner and Lyon’s account, the central designer of the internet’s physical layer. The wires themselves were from AT&T but those weren’t very special, rather the special parts were the IMPs (Interface Message Processors) put together by BBN. The IMPs were the computers that divided up the data into packets and sent them along to the next IMP until they arrived at the destination IMP which would hand over the data to the computer(s) it was attached to. BBN made several important design decisions concerning their IMPs and how they worked which I list in order of interest to me.

  1. Their threat model was curious grad students who wanted to play around with the new computer in the office.
  2. The IMPs were designed to be as invisible as a wall outlet. (toddlers with forks vs grad students with screwdrivers?)
  3. The IMPs were controlled centrally; BBN could update their software over the air (wires) and nobody (grad students) was supposed to tinker with their own versions of IMP software.

The first line of defense against 1. was a military specification case for their Honeywell 516 minicomputers. (This also meant that all IMPs had identical software and hardware; talk about a large attack surface. Not that it mattered back then). The second line of defense was 2.; who tries to reprogram outlets anyways? By keeping the network working BBN kept the grad students occupied and out of the precious IMPs. Of course this had far far greater implications as it allowed the internet to scale (see Where’s Waldo’s last post) and eventually allowed for a nation internet users without a damn clue as to how the thing works or what it’s made of (it’s probably a series of tubes but don’t quote me on that one).

Lastly; the extremely centralized control of the early ARPANET. BBN’s model of control is the complete opposite to the philosophy of the distributed or even a decentralized network. Yes, the nodes are distributed and packets flow through them whichever way is quickest, but the puppet master’s at BBN still controlled everything except for ARPA which controlled them. But what if ARPA had ceded control to BBN, and AT&T had bought BBN and inevitably ruined everything? If the internet is going to be controlled centrally, then it must be treated as a public good or we risk some telecom coming in and adding “features” (charging per byte etc.). The other solution is to make governance so confusing, opaque and most importantly boring sounding that only a few good(ish) souls bother to take part. As far as I can remember, that’s what ended up happening, along with RFCs.

I think I will write about RFCs next week and maybe by then I’ll have figured out exactly what happened after ARPA. My intuition from seeing Wikipedia’s uses of RFCs is that they are key to more distributed governance.

Leave a Comment

Log in