AMD [IBM] press release

Robert G. Brown rgb at phy.duke.edu
Wed Nov 20 14:13:00 PST 2002


On Wed, 20 Nov 2002, Bob Drzyzgula wrote:

> Yes and no. I agree with you that the intensive marketing
> by these companies is a major factor in raising the
> volumes, and by consequence lowering the prices, for
> these chips. And I agree that, for like specifications and
> sufficiently small purchase -- like one or two systems --
> a brand name system can be cost-competitive with privately-
> or locally-assembled systems.

 <deleted detailed description>

> We *do* have to maintain a competent staff and a well-
> equipped integration facility, but tools and ESD benches
> are cheap compared to computers, and the empowerment
> that comes from all this self-reliance does wonders for
> job satisfaction and staff retention.

Beautifully put.

I like to build my own systems as well, and tend to spend MORE time on
systems built by vendors, even very friendly and cooperative local
vendors, than I do on systems I build myself, although it has taken me a
year plus of buying lots of systems and letting the vendor build them
(and then having to mess with them later) to really figure that out.
Unless you have a linux-expert vendor that can do EXACTLY what you have
to do to install the system for you, you end up messing with it more
getting it to where you can install it from an often unknown and
slightly broken initial state OR communicating with the vendor about
what they did wrong than it takes you to just do it, at least with
modest numbers of systems.

With that said, there are some components and circumstances where Dell's
fancy hardware makes sense -- department LAN servers, for example, where
one can minimize the consumption of the scarcest of our resources --
primary sysadmin time -- by buying the highest quality servers, keeping
them under the expensive same-day service contracts, and upgrading them
pretty steadily as they age out.  This is 2x or more expensive at the
hardware side, but can prevent expensive downtime on a major shared
resource.

I think many people confuse HA with HPC.  HA often demands brand name
stuff and all the expensive service deals, because of the
nonlinear/magnified costs associated with downtime.  HPC is USUALLY
fairly insensitive to downtime of single nodes -- it costs you a 1/N
fraction of the total resource, it might cost you one chunk of work at
whatever you've established your checkpoint/task granularity scaled with
N, but it rarely costs you an extended N-scaled loss of resource.

Besides, as Bob points out so ably, a sensible buying/building pattern
can actually significantly reduce downtime even compared to the best of
service contracts.

  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