[Beowulf] Begginer questions

[Jim Lux]

At 11:08 AM 8/6/2005, Nicholas Papadakos wrote:
>Hello,
>
>I have some machines laying around and I was thinking for the past 3-4 
>months to make a small Beowulf cluster.
>I have tried something similar in the past, but I had used the openmosix 
>approach.
>
>The machines are:
>
>2x <mailto:p4 at 2.4GHZ>p4 at 2.4GHZ
>1x <mailto:p4 at 2.0GHZ>p4 at 2.0GHZ
>1x p3 at 900MHZ
>

If you're just fooling around, arrange them any way you want.  It's pretty 
easy to rearrange things, and the information you gain from the exercise is 
the most useful thing, anyway.


>I was thinking of putting the p3 to do the master node and the rest the 
>slave nodes( under the assumption that the master node doesn't do anything 
>else than coordinating the slave nodes).
>Reading a bit more about beowulf  I found that the master node needs to be 
>the most powerfull machine in the cluster? ( this is not very clear to me 
>if its true or not)

Depends on what the nodes are doing, and how much they hit the disk (or 
require other services from the headnode). I will mention that you usually 
wind up doing things like compiles and editing on the head node (because 
it's most convenient), so having a reasonably fast processor there is nice. 
I have  a fairly funky cluster (battery powered, and 802.11a wireless 
interconnects, for instance), and virtually all the work is done by the 
nodes (which are 533 MHz Via EPIA-Ms).  Originally, my head node was a 
120MHz P1, figuring that all it had to do was respond to DHCP and PXE stuff 
and serve out the boot image, which it did perfectly well.   What a royal 
pain that was, though, since I was building that cramfs boot image on the 
head node, and it was creepingly slow.

Currently, we're hassling through getting NTP to work reliably.



>Should I use the p3 as the master or one of the p4?
>
>Also , since I have several spare 100mb Ethernet cards arounds, to put 2 
>in each node and bonding them together. Is it worth it?

Try it and see how it works.  Again, the experience from tinkering is the 
best result from an experimental cluster.  In your case, trying to quantify 
the difference with bonding or not might be the more interesting thing.



James Lux, P.E.
Spacecraft Radio Frequency Subsystems Group
Flight Communications Systems Section
Jet Propulsion Laboratory, Mail Stop 161-213
4800 Oak Grove Drive
Pasadena CA 91109
tel: (818)354-2075
fax: (818)393-6875
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.beowulf.org/pipermail/beowulf/attachments/20050806/c4d752ee/attachment.html>