[Gllug] Bootable USB backup like OS X dmg?

[John Hearns]

On 30 May 2011 01:19, Steve Parker <steve at steve-parker.org> wrote:

>
> I've built over 100 Sun/Oracle x86 class servers in the past year or so,
> with Oracle Enterprise Linux (basically RHEL, or CentOS). They all have
> onboard NICs (some have two, others have four) and all have a quad PCI NIC
> for resilience. During the install, they all detect the PCI card first, then
> the onboard, so the onboard is eth4. Long-term, I want the first onboard NIC
> to be called eth0, so I have to mess with udev rules.

You make a good point there.
That is one of the dark arts of building Linux clusters - you get a
PXE network boot and build process which works with a given hardware
combination. You need to get a list of MAC addresses of each server in
the cluster - which you can do by sniffing the packets or tailing logs
on the DHCP server as they are booted up, or you scan barcodes if the
manufacturer iis nice enough to label them for you. The complication
comes as it is quite normal to have two separate ethernet fabrics -
one for general administrative traffic and storage, one for MPI
message passing. If you stick with one motherboard you can get
eth0/eth1 sorted out easily, and write the udev rules with a script if
necessary.
Switch hardware vendors and you have to start watching which is the
one ot tries to boot off first. Yes, I know you can sort this out but
it is a big hassle when you're trying to build a new cluster.
I even had a case several years ago when the sales types switched to a
new model from the same manufacturer. We used IPMI over ethernet,
which shares the NIC with one of the main ethernet ports (ie there is
an internal bridge which filters off the IPMI traffic to the BMC
card). We had a build and cabling which worked fine - but change of
model meant IPMI appeared on the other NIC. Cue wailing and gnashing
of teeth behind the racks.
--
Gllug mailing list  -  Gllug at gllug.org.uk
http://lists.gllug.org.uk/mailman/listinfo/gllug