Re: First Beowulf cluster. Some pointers?
Quoting stramiello <stramiello@whiterabbit.resnet.gatech.edu>:
> > I was thinking of making a minimal installation (I managed to have
> > a installation of ~ 40Mb on a firewall system I made a while ago) on
> > the master node, and then export /usr and /tmp. What exactly to do
> > with /var I'm not quite sure of yet, but parts of it could be NFS'ed
> > I suppose...
>
> Just remember that if you're running NFS for the root (diskless), you
> either need a memory galore to dump it in on the nodes, or else separeate
> copies of the OS for each client on the masternode.... The only things we
> export via nfs (from the master) are /home, /usr/local, and /usr/shared
The disk less setup I did a couple of years ago only had a special boot floppy,
which had a DHCP/BOOTP kernel dd'd on it ('rdev /dev/nfs' or something...)
which mounted / from a NFS server... It was slow, but worked find.
If I remember correctly, I only had separate /etc's for each client, the rest
was the same...
> > stramiello> specifically written in MPI to run across the nodes.
> >
> > Darn. And how 'difficult' is this? What -dev packages do I need for
> > this? Any FAQ/HOWTO's about this?
>
> MPI isn't difficult. It's a lot like pthreads. I'm not sure of any FAQ's
> or HOWTO's. Where I work bought me some books on it and I read those.
PThreads I've used before (a couple of years ago though) so don't sound to
daunting. I'll see if I can find some books to then...
> > In theory (I'm reasonably good at perl/C/C++), could bash/gcc be
> > modified to use MPI?
>
> In theory, yes. In practice? Not really, as far as I can tell. You have to
> realize that not everything can/should be paralellized, and not everything
> can/will gain from being paralellized.
>
> Although I qualify that... I could be thinking you're wanting to do
> something different. GCC paralellized... do you mean the compile itself
> occurs in parallel across the nodes? Or it compiles a program to
> automagically use MPI?
The first.
What I would want is one HUGE (oki, so in this example, with the SPARC, it's
not HUGE, but more like 'just bearly enough' :) machine, but which is in
reality many computers where people login to the cluster, not to any one
of the machines.
Say (for simplicity) that each separate node have a capacity of 1 486/33
(which I've been told it really have). I don't remember how many bogomips
a machine like that would have, but for simplicity I'll invent the number
'50'...
Clustering 10 machines (the reasonable amount of SPARC stations I could
get working I think) would equal a machine with 500 bogomips...
THAT is what I would like. I put the master node as the front, and have the
slaves behind that (somehow) and you login to a machine that is just a
simple 486/33, but with the processing power of, what? A PII/200 (or whatever
it would result in)?
I might have misunderstood exactly what a cluster (or Beowulf cluster perhaps),
but a couple of years ago a friend (same friend as in previous mails) had a
HP 300/something... There where three nodes, one disk. I don't think I ever
saw this running, but I was told "it's one machine, with the load spread to
three CPU's"...
My experience with Linux/UNIX wasn't so high back then, so... :)
--
spy Khaddafi Noriega 767 radar fissionable nitrate arrangements
smuggle PLO plutonium Cocaine quiche munitions North Korea
[See http://www.aclu.org/echelonwatch/index.html for more about this]
Reply to: