[Beowulf] Please help to setup Beowulf

[John Hearns]

2009/2/17 Robert G. Brown <rgb at phy.duke.edu>:
>>
> Right now the short answer for a vanilla starter cluster is:  Connect
> your systems via a switched network on a private IP subnet, setup a
> shared common account space and remote mounted (e.g. NFS) working
> directories from a shared server.

Just to amplify on what Bob says.
To build a vanilla Beowulf, you need to connect all eight of these
boxes (for want of a better term) to an ethernet switch via patch
cables. This is your own switch, not one controlled and run by your
campus network people (but see below).
This is the 'private network'.
Allocate one box as the 'cluster head node' or 'master node'.
This should have TWO network interfaces - if it is a modern box, it
will have two network sockets for those RJ45 patch cables already on
the back of it. If not, you will have to get an additional network
card and slot it in.


Actually, you could perfectly well run a computational cluster over
the campus mamaged network.
Just install them with campus 'public' IP addresses, and install the
appropriate MPI libraries and scheduling systems.

However, the advantage of a private network/subnet is that you can
(a) control how your boxes boot up - ie reinstallation and network booting
(b) you have the full banmdwidth of your private switch for the
communications between the nodes - both the parallel messages the
programs send and the storage traffic to the NFS server on the head
node.


ps. that reminds me - if any one box is more powerful, or has more
disk space then that's the one to choose for the head node.


Now, once you have neatly piled up those boxes, and have done the
mains and patch cables neatly (think of the future - one day it is you
that has to fix a failed box! - get those cables tied neatly in
bunches),
email the list and we can help from there.