[Beowulf] Sidebar: Vista Rant

Robert G. Brown rgb at phy.duke.edu
Wed Jul 18 07:51:26 PDT 2007


On Tue, 17 Jul 2007, Jaime Perea wrote:

> Something related, it can't be serious: Any comment from
> the experts? :-)
>
> http://www.linuxtoday.com/high_performance/2007071702826NWHEMS

Only that the article contains several basic contradictions.  "The cost
is about the same..." in one part, and "The cost is $469 per node" in
another.

$469/node is not the same is $0, except in the capable hands of a
Microsoft sales rep, who will point out that it will cost FAR more in
human time to get a grotty old Linux cluster going using RHEL and the
marginally competent MCSE hired by the department to manage the Windows
systems used by the secretaries and staff.  Or perhaps they are
cost-comparing it to e.g. Scyld or RHEL-based nodes.

Note also that I'm sure the $469 price is full retail, leaving a
generous margin for HP and IBM and anybody else who wants to see
pre-configured Win clusters.  This means that in a mass-production
environment (where they can basically flash images directly onto disks,
run a post-install script to complete, and be done at very low cost per
system) they can add $200 or so in pure profit to every node sold.  They
probably can't mark up the linux systems as much as the customers KNOW
that they have no real software costs and will just (re)install them
themselves anyway.  When we buy e.g. Penguin nodes, it's very kind of
them to preinstall e.g. FC on the nodes and then burn them in.  We care
about the burnin, but we're going to reinstall the nodes via
PXE/kickstart when the arrive anyway, so we really don't care about the
install.

People who BELIEVE the FUD about lower Win TCO are a mix of ignorant and
correct.  They are correct in believing it because they are too ignorant
to actually spend the remarkably short time required to learn to install
a linux cluster from scratch on top of any of the major FREE
distributions.

Some years ago when I used to still give people tours of our brahma
cluster (where my own systems were several shelves full of mid-towers
where I've long since gone over to rackmount FF, to give you an idea of
time frame) I would do a little demo.  After running one of the
mandatory demo programs, pvm xep say, so that they could see speedup
happening in a pretty way, I'd say "now let's reinstall a node".  I'd
then proceed to power cycle a node and select the node kickstart from a
PXE boot menu.  I'd then start talking, saying things like "I'm now
reinstalling a node from scratch.  The hard disk is being completely
wiped and re-imaged, and at this point you can see that a whole lot of
packages are being installed.  See the little flashy progress bar?  Each
of the flashes as it fills up is a whole package full of software,
libraries and such.  When this task is done, the node will be configured
IDENTICALLY to the way it was before, and I can resume my applications
as if nothing ever happened.  The install is taking place over a fast
network connection and the server is right over there, and ah, look.
The install is finished and the system is rebooting itself.  There, it
is running a first-boot script that is completing the configuration, and
there's the login prompt.  If we change screens to here (toggle KVM) you
can see it pop back up in the wulfstat display, there, it shifted from
down to running again, and the load average associated with booting is
dropping fast.  Now it's back to idle.  The node is ready for use once
again."

That's how difficult a cluster node was to install maybe six years ago.
Now the network is faster, warewulf has made diskless boots much easier
(and probably much faster than my "install").  One can fit a full linux
distribution on a 4 GB memory stick, a node image on a 1 GB stick.  No
matter what, it is pretty difficult to imagine installing a cluster node
being any EASIER than "turning the node on" after some quite
straightforward and extremely well-documented configuration of a
pxe/dhcp/tftp server and the construction of a node kickstart script.
With yum it is really even easier now, as one can build a SINGLE base
system kickstart script to install all different kinds of systems and
then differentiate the system type with a yum install script that
"finishes off" the system according to usage, desktop server node
whatever.

But if one is ignorant of this, and if one considers the "cost" of
actually learning new things to be high, then one is quite correct --
sticking with Windows is the right thing to do.  Go ahead, pay them
whatever they ask to do your thinking for you, and accept gratefully the
limitations in what they give you and its real comparative cost-benefit.
Not that you are likely to ever know what they are...

    rgb

-- 
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 mailing list