hahn at physics.mcmaster.ca
Fri Jun 18 05:59:02 PDT 2004
> And if we want to upgrade the software (like MPI libs, c libs, etc.), we want
> to do it real fast.
> What's the best bet?
don't install anything on the nodes. booting over the net (dhcp/tftp/pxe)
works extremely well, and means that upgrading a package for all nodes
takes exactly one serial rpm/yum invocation on the fileserver.
I don't even bother with initrd's - just deliver a monolithic kernel
to the compute nodes and have them mount their root over NFS. remember:
other than init and something like sshd, the boot sequence for a compute
node can be very abbreviated, not a lot of file IO.
just because the nodes are ephemeral (nothing installed) doesn't mean
that you can't use the disk on a node if it has one. it's great for
swap and /scratch.
More information about the Beowulf