[Beowulf] Beginner advice
Robert G. Brown
rgb at phy.duke.edu
Tue Dec 13 16:14:24 PST 2005
On Wed, 7 Dec 2005, Sam Odio - OdioWorks, LLC wrote:
> First of all, I would like to than beowulf.org for hosting this mailing list
> - it's a great resource.
> I am looking to build a budget beowulf machine using commodity hardware. I
> expect it to use a Linux OS such as Debian or Gentoo. I have never built a
> clustered computer before... and was hoping to get some hardware / software
At the risk of starting a religious war, I'd avoid Gentoo -- you'll have
enough work building the cluster without rebuilding the OS N times
(every upgrade). I'd look into PXE based solutions. Debian should be
fine. So should Fedora Core 4. So should Centos 4. I haven't heard
anything negative about Mandriva or RHEL or SuSE (well, I have heard
some negative things about SuSE but don't have any personal experience
there and people certainly use it and do fine with it). I personally
prefer an RPM-based, kickstart installable, yum maintainable solution.
I'd stronly suggestion taking a look at warewulf. It lets you boot your
nodes completely diskless, which saves you a lot of money on a home
cluster. You'd need a single "powerful" server with a lot of disk, but
the nodes can be a simple stack of whatever is cheapest/fastest that you
can find, with a GOOD QUALITY PXE NIC and a "lot" of memory (but no
disks at all). Look at the "value cluster" on www.clustermonkey.net
(featured in the now defunct Cluster World Magazine). $2K for 8 nodes.
> I don't have a purpose for the machine yet - I am really just doing it to get
> some clustered computing experience. Therefore I'm just looking to build a
> solid general-purpose clustered machine.
> These are the specs I have come up with so far, all suggestions welcome:
The second simplest specs are:
a) Pile of vanilla boxes with PXE NICs (3coms, Intels, something
decent) and at least 512 MB of memory. No need for floppy, CD, hard
disk. A cheap or builtin video interface would be useful. No need for
dedicated KB or mouse.
If you are serious about performance (especially numerical performance)
go with an AMD 64 based cluster. This is the cheapest/fastest out
b) One box configured as a "server". This one might do well to have a
GB of RAM, 3-4 hard disks configured as an MD RAID, a couple of network
interfaces, a snazzy (linux supported!) video card and nice monitor, as
this is where you will "work". Install the linux distro of your choice
on this box, rememebering that you can change your mind if it doesn't
work out for you (all the suggestions above are free but RHEL and SuSE).
Build your RAID, create a nice NFS export partition or two on it, put
e.g. a FC4 mirror on it, learn about yum. Go get wulfware, and give it
c) A KVM switch with as many ports as you have nodes would be useful
but isn't essential.
d) A decent ethernet switch IS essential. If you want to get
semiserious, get a GB ethernet switch (and GB NICs). If you just are
doing it for fun, get any old 100 Mbps ethernet switch and any old PXE
Following the warewulf directions, you should be able to build a cluster
in a day or so of gentle play. You buy what you can afford here,
starting with a server and at least one node, then scaling up as you
The CHEAPEST and EASIEST cluster is even simpler. Take N ordinary
desktop computers. Install linux on them (with a shared account space
and an NFS-served home directory). Install (with yum) e.g. PVM and/or
MPI. Intone in a deep voice "Thou are a cluster! So mote it be!" while
stabbing a white-handled athame into a can of Jolt cola.
Works for me...
> Thanks for your help,
> Sam Odio, Owner
> OdioWorks, LLC Business Services
> Website: www.OdioWorks.com
> Main: 1-866-544-4132 (toll free)
> Cell: 1-888-282-2846 (toll free)
> Fax: 434.979.3397 (local)
Robert G. Brown http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567 Fax: 919-660-2525 email:rgb at phy.duke.edu
More information about the Beowulf