[rescue] Perverse Question

[Charles Shannon Hendrix]

On Thu, Jun 19, 2003 at 11:17:36PM -0700, Francisco Javier Mesa-Martinez wrote:

> You'd be amazed, that is one of the reasons why SGI is have a hard time
> selling them. Origins can no longer keep up, in most EP fields. Their lack
> of processing power per node can not longer be compesated by the nice
> internodal bandwidth.

Once you get into really large systems, the cost differential starts
melting away though. It's also very easy for the maintenance costs of a
cluster to be higher than that of a big SGI or IBM.

I think what you are really seeing is this:

1) Really cheap clusters built out of bargain-basement PCs. These
are fine for small clusters and short-term use. The traditional high
performance market doesn't even bother with this, and for good reasons.

2) Decent cluster systems at a power level and price point the big
players have ignored, to their detrement. They ignored this because in
the supercomputer boom times, Uncle Sam and big industry was eating up
big machines with no end in sight. This left a large class of users with
the option of either doing without, or borrowing time on someone else's
big system.

Another thing to consider is storage. Truly reliable, large-scale
storage can cost more than the computer system which uses it.

Then there is the cost of maintenance. At some point a cluster starts
costing you more than the same size mainframe or big NUMA system.  This
is even more true if you buy unreliable systems to populate your
cluster.

> > The whole beowulf thing is surrounded by a lot of dangerous hype spread
> > by clueless people who have never actually set one up.
> 
> Same goes for almost any sort of high performance computing...

Yes, but they aren't hyped near as much as the "el-cheap PC cluster"
stuff.


-- 
UNIX/Perl/C/Pizza____________________s h a n n o n at wido !SPAM maker.com