> support as many users as the processor and RAM will allow. All the servers
> should be the same as this minimises support costs. Also, you can remove the
> HDDs from the servers, which reduces the cost of the hardware and is less
> components to break. Also if all servers use a central file store, then they
> can all be kept consistant. When adding a new server, you just plug it in
> and it's ready to go.
I agree with you, there is plenty advantages of using diskless nodes.
> Has anyone done this before? Are there any implementations? I am currently
> in the process of knocking this together but have a few questions.
2 years ago in an internship, I did some scripts for thin-oscar. This is
a project you could be interested in. http://www.sf.net/projects/thin-oscar
The principle of operation is that the root is in ram. At this time, we
had to make the root as a raid device to support big ram disk. We had
lot of fun.
I didn't work on it since then, but it could be cool to reopen this
> What should one do with /var? I am thinking of making it a RAM disk
We let it in RAM.
> What should one do with /dev? I understand each server needs there own /dev?
> I've heard it can be either done with a disk or devfs?
You can have it in the ramdisk, but devfs is a must. There are plenty of
devices and they consumes inodes in the ramdisk.
> Finally how do I get DHCPcd working?
In the user space after pre-boot process, we used all executables and
scripts available with the distro.
Hope that it can help you,