🚀 go-pugleaf

I'm not sure if Sun wants to introduce a somewhat hybrid sort of
diskless/dataless clients into suninstall(8).  See, the problem with
diskless clients is that it makes the entire configuration hard to scale
up.  I mean, for people want to run relatively large (CPU & memory) jobs
on their desktops, you'll need to allocate 25~60meg of swap per client and
if you've 10 of these, you'll already have consumed a large chunk of space
on the fileserver.  And this means that you'll need to get more disk space
for the fileserver -- I'm not sure this is the most economical route to go
though since SMD/IPI disks tends to be more expensive.  Besides, you
generate a lot of network traffic when everyone is cranking up their load.

In this situation, making every node to be a dataless client is the way to
go.  But it is a hassle to setup and more difficult to manage since each
dataless client has its root partition on local disk.

What would be nice is to make diskless clients swap locally and have
temporaries like (/tmp, /var/spool/mqueue, /var/tmp, ...) reside on local
disks.  This way, you don't have to worry about allocating large chunk of
space on the fileservers for client swapping and it should reduce net
traffic quite a bit.  And you can manage these nodes as conveniently as
before (since each client's root only takes up to 2meg, it can scale up
easily and besides, you can rehack each client's root so that they all
*share* a common root except for some files that are really
machine-specific (e.g. /etc/rc.boot, ..etc...) and you'll basically get
everything for free).

I've just turned a dataless 3/75 into a pseudo-diskless 3/75 and it works
pretty well so far (it has a 71meg shoebox attached).  I just thought that
Sun should provide this option in suninstall(8).

Eric Ho  
Princeton University
eho@clarity.princeton.edu