When the installer has started, just install as usual. When it asks you to install the kernel and drivers, you don't select it, but set up the network first, then select it. You will see the nfs option enabled after you have set up the network. Then, you can install via nfs. I had the nfs server at 192.168.1.201, so it was like this : 192.168.1.201:/root/shared/ and when it asked for the directory with a default of /debian, I answered / and I just hit enter several times.
The installer asks once more on installing the base system, and once again use the nfs method there. I did not want anything, so I skipped the rest, saying no to everything, and just quitting dselect. One problem I had was that the system was not globally ip-reachable, I could not get anything.
For ease of maintenance, I installed telnetd. It was not IP-reachable, so I got it off somewhere else and put it on the server, and did a ftp from there on each of the machine.[8] This is rather important, since clusters are usually hard-to-access physically. I towered it up and it was quite hard to reach the machines at the bottom because all the think wires (power supply cables) have accumulated at the bottom. It was quite important that it was maintainable from a remote machine [9] .
Doing ping's and telnet to test connection is quite important. I have never had an experience of linking 18 machines before, and I never had to connect more than one hubs. I did not know that hubs had a special method of connecting when it was connected to another hub. It took a long time to figure it out. It seems like the method depends on each hub. In my case, the 8-port hub I used had connectors 1-8 and another one labeled 8x. Apparently, another hub had to be connected to the 8x port or it was not recognized.