Intel is finally shipping the 64-bit Itanium

[Jeff Layton]

Bari Ari wrote:

> Mark Hahn wrote:
>
> >>
> > I can't imagine Itanium being a mass-market item for years, if ever.
> > and I pledge allegiance to the Orthodox Church of Beowulf, which
> > holds that if it's not mass-market, it's not cluster-Kosher ;)
> >
> The AMD Sledge/Hammer series will also be nice for clusters whenever
> they finally make it to market. Hopefully there will be some nice
> chipset support to go along with them. For the time being Mips has the
> price performance edge since nobody has taken the ARM 10 to market yet
> and Intel yanked the FPU out of the XScale before they released it.
>
> It's great to see Beowulf clusters offering similar performance to
> traditional supercomputers for coarse grained applications and even some
> fine grained for a fraction of the cost, but X86 with OTS motherboards
> will also always be a kludge. X86 has 20 years of baggage for legacy
> support and also produce enormous amounts of heat as compared to RISC.
>
> Low cost RISC clusters will outperform any x86 mass-market OTS clusters.
> RISC offers lower cost, smaller footprint, far less heat along with
> higher fixed and floating point performance.

Grasshopper. You forget the wonders of commodity components.
This wonder is what has driven the Beowulf "revolution." There are
many arenas where x86 performs very well. Several billion chips
versus a few hundred thousand allow many great things to develop.
Couple this, grasshopper, with Open-Source OS, compilers, message
passing, queuing systems, and many dedicated people and you have
the Beowulf revolution.

(Removing my teacher mask for a moment). One ALWAYS needs to
benchmarks their app(s) on all potential cluster solutions. For example,
we tested our primary application on Intel and Clusters, SGI Origin
systems, Cray T3E, and IBM SP2s. On a pure performance level, the
Intel cluster outperformed the SGI, Cray, and IBM systems. It also
outperformed the 21164 Alpha clusters. Only when we got to the 21264
Alpha running at higher frequencies than the Intel did we see any
performance gain over the Intels. Remember, this is PERFORMANCE
only. Guess what happens when we went to price-performance?
The Intel clusters where 4 TIMES better than the Alpha clusters.
As for the SGI, Cray, and IBM, forget it. We didn't even compute it.

Again, this is our application. YMMV. On the other hand, I've seen some
benchmark results for another Lockheed application. The Alphas eat
it for lunch. The Intels don't perform too badly though. When one gets
down to price-performance, guess who wins? Alternatively, when we
consider the most power for a fixed price, guess who wins?

I'm not trying to say that Intel is the way to go always. Sometimes you
need as much speed as possible and in this case I would guess that
the RISC stuff will do pretty well (although as I mentioned on our code
it didn't do too well until we hit the bleeding edge of current RISC
CPUs). But, while many of us are performance junkies, we are also
not endowed with large budgets. Therefore, in many cases, commodity
pricing does a wonderful thing for price-performance. Moreover, in
many cases, for a fixed total price, you will get more bang for the buck
from Intels (although that's not always true).

Personally, I try to stay CPU agnostic. I really don't care what I run on
as long as I get the most speed for a fixed price. If it's Intel, AMD,
Alpha, Transmeta, some Russian abomination, a Chinese copy,
whatever, I don't care. As long as I can get an OS to run on it, good
compilers, god message-passing, and someone to support it when I
need it, then it's a candidate.

Enjoy the holiday!

Jeff Layton

P.S. My father is a historian. One of his favorite quotes is, "Those who
refuse to study history are doomed to repeat it." (or something like that).
I like to modify it for those new to Beowulfs, "Those who don't study
commodity components are doomed to be crushed by it."



>
>
> Bari Ari
>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf