diskless 72 node

[Stephan Mertens]

Hi Beowulfers,

we are planning a 72 node Beowulf cluster with diskless nodes.
We managed to get the grant for it, but the corresponding
referee doubts that a diskless cluster of this size can work.
He wants us to find a working "reference installation"
before we actually can spend the money.

Here is our setup:

72 PIII double boards, each with 512 MByte, floppy and 2 NICs
(100Mbps). One NIC is for interprocess communication,
the other for NFS.

1 dedicated NFS-Server (Linux, PIII, 19 GB RAID)

Everything is connected via 100Mbps switches.

Do you know of anything comparable that is running smooth?
Or any serious pitfalls?

We could afford local disks, actually, but we don't like extra heat, 
power consumption, noise, sources of failure etc.

Thanks for your help,

Stephan