Re: SMP scalability: 8 -> 32 CPUs

ralf@uni-koblenz.de
Thu, 3 Dec 1998 07:23:39 +0100


On Wed, Dec 02, 1998 at 12:00:43AM +0000, Christof Damian wrote:

> > I had a similar sized monster (Convex C1) in my dining room for a year or
> > so. I'm sure had I ever gone through the week-long trial of booting the
> > thing and attaching it to my network of Linux boxes, I would have been
> > disappointed. The mystique of a Cray-compatible vector processing
> > supercomputer would have been shattered in comparison with a modest modern
> > PC.

Only when talking about _outdated_ Crays or maybe the EL series of entry
level / development systems. The comparison of topend vector iron, be it
from Fujitsu, NEC, Cray with PCs still shatters PCs mercyless, especially
when comparing the same program versions, that is the typically vector
optimized Fortran heritage code users of vector machine typically use.

Aside, good old Seymour knew how to build machines that scale for real
world apps. Many comparisions of vector / microprocessor performance I've
seen make the fault of using values close to peak performance for the
micros. Number crunching code not especially tuned for micros often has
hit rates of less then 10% in data caches, goodbye micro ...

> does anybody know how to get one of these monsters ? i got a pretty big
> space and free electricity. i am not a kernel hacker, but would like it
> anyway. i also would be able to put it on my leased line once it is
> running.

Afair a Convex C1 - sucks about 6kw. Don't underestimate the power
requirements of the cooling itself, you can easily end up with something
like 10kw total power requirements. That's about 3 x 16A, 230V. Experience
with a much older machine (Interdata 7/32) was that the lines need to be
significantly stronger to avoid the power on surge blowing the fuses.
That means you'll need at least 3 x 25A or even 3 x 36A 400V if you do not
want the power on surge to blow the fuses. That might end up in quite
a bill for the electric installation and AC if you don't happen to have
a suitable environment already. The other thing is that it is very hard
to keep machines of that caliber alive for a long time without going
bankrupt unless you're a die hard who can dig up the require spare parts
and knows using logic analyzers & soldering iron just as well as joe user
knows a mouse ...

Ralf

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/