I'm attempting to set up SGE 6.1 to do load sharing on my network. I am becoming quite confused though with the installation and configuration.

I've installed the Master and Exception host on the main server and all services are started. I'm attempting to install a second Exception server now which is causing some of the confusion.

Do I do this even on the computers that aren't masters? I'll be setting up Grid on up to 10 computers, does this mean I have to go through the installation process 20 times(1 install of master and execution per computer)?

Classic or Berkeley Spooling? By the looks of it Berkeley is the way to go as it is faster than Classic. This will require me to install more software, correct?

If there is a full, non-terribly written guide somewhere I would love to see one. I've been using the ones from Sun but they don't go into detail on most things that could use it.

Be clear they are "Execution" hosts not "Exception" hosts.

No. There must be one master. From that master you can then install the execution hosts.

What you need to understand is that the installation directory (e.g. /opt/sge) should be accessible from ALL hosts. Each host does not have a separate installation, there is a shared installation across them all. This allows all hosts to have the same information about the current state of the system.

For example, you should share /opt/sge via NFS and mount it on all execution hosts as /opt/sge. Then using the installer on the Master server, perform the execution host installation on all hosts.

There is a lot of documentation available as to which to use and I cannot comment without knowing a lot more about the setup.

Start with classic.

The Sun documentation is sufficient.
http://docs.sun.com/app/docs/coll/1017.3

There are many many guides on the Internet detailing what individuals have done, mainly blog posts, that will also help. Google is your friend.

I managed to get it running correctly. The first guides I used had no mention about having to share the installation directory. Another thing I had to do in order to set up the execution hosts at the same time as the master was to set up password-less SSH access, otherwise all other hosts would show as "Resolvable".

Another Q:

I want to limit the number of jobs for each execution host in the cluster. I see a spot to limit jobs per user/group but nothing for host. Is there a way to do this or does it have to be done through resource management only(free mem, cpu, etc)?

My cluster consists of 5 computers and some of the regressions that get compiled use up to 6GB of memory but I have been asked to have it set up so that only one job can be run on each execution host at a time instead of doing it the simple way with the free memory variable.