Chapter 6 - Networking: Arpanet 1969-1972
6.7 Network Measurement Center - 1969-1970
In December 1969, after all four IMPs had been installed; the Network Measurement Center (NMC) under the direction of Kleinrock began generating weird configurations of data traffic with artificial traffic generators in both the IMPs and Hosts in order to stress the Subnet to bring it to its knees; to stop it from functioning. This role put the NMC on a collision course with BBN since BBN had to correct the problems. Kahn of BBN had concerns of his own, especially regarding potential “deadlocks,” when the network simply froze and would not function. In December, he paid a visit to NMC and remembers:
The very first test I ran on the net was a deadlock test. We deadlocked it. Dave Walden worked with me on the experiment and got to see it happen. I think it was six months later before everybody else sort of gradually came along to really understand the implications. It was a very frustrating time, because I couldn’t convey that notion to anybody else despite the results.
At this time, a long lasting professional and personal friendship began between Kahn and Cerf began. Cerf remembers:
The whole system was largely Bob’s architecture, and he came out to UCLA to go and find out, by kicking the tires, how this thing would really perform. He had some theories about places where it would break that not everyone agreed with him about, but he said he was going to go and prove that it would break in certain ways by generating traffic at UCLA and forcing it through this four node network in various and sundry ways. So, he and I worked together, me doing the traffic generation and measurement software and he was figuring out what experiments to perform for three or four weeks at UCLA. That’s a collaboration that has gone on ever since!
In many respects, BBN became the new AT&T - seemingly indifferent to the users of the network. But from BBN’s perspective, their objective remained to satisfy their client, Roberts, and to assure a working network. Not to be concerned with every little problem arising from perverse network testing. Kleinrock recalls telling BBN:
In the early days, when they came out with their proposal, we looked in their routing algorithm, which we were very interested in, and said: ‘This thing has loops. A packet can go back and forth, or in a circle.’ And they said: ‘No, there are no loops.’ But there are loops! They said: ‘No.’ So we simulated them and said: ‘Look, there’s loops.’ They said: ‘There are no loops.’ Then this four node network was implemented and we showed them loops, and the answer was; ‘Yes, there are loops.’
This early work led to a new understanding of the phenomenological behavior of networks. Kleinrock explains:
You give me a network and I can analyze it’s performance in a very simple way. We developed exact mathematical formulas as well as approximations to understand networks. We also proved that measurement was very important. We were able to demonstrate deadlocks and degradation’s and explain what was causing them. We found that the whole issue of flow control was a key issue. For example, problems of sequencing, the keeping of things in order, causes deadlocks. The function of flow control is to make things move smoothly. If you want to do that, you put in controls. Once you put in controls, that’s called a constraint. If the network can’t meet the constraint, it crashes. So the thing you put in to help out is the thing that kills you, or degrades performance. We had a whole catalog of degradation’s and deadlocks which BBN eventually fixed. Those things are still present in every network today.
Despite lingering problems and differences of opinion, the network worked. By June 1970, the network included SDC and RAND on the west coast, and BBN, MIT and Harvard on the east coast. All the while, network performance testing continued uncovering problems, most of which would be corrected, including some thought intractable.2 By mid-year, a rudimentary version of the Network Control Center was established at BBN.3 Four more nodes were added by the end of the year: Stanford, Lincoln Labs, Carnegie Mellon and Case Western.
- [2]:
“In late 1971, it became apparent that serious problems were being encountered with the flow control system. BBN set about to design a new flow control system. By mid-1972 the revised system was ready for installation after undergoing extensive testing in the laboratory on a small scale network (3-4 nodes).” ARPANET Management Study by Cabledata Associates, Inc., Jan. 1974, pg. 9
- [3]:
ARPANET Completion Report, pg. III-55