Archived from groups: comp.sys.ibm.pc.hardware.chips,comp.sys.intel (
More info?)
Bill Davidsen wrote:
> Robert Myers wrote:
>
>> <quote>
>>
>> The reliance of supercomputers built on clusters, commodity products
>> that operate in parallel, has been criticized before congressional
>> panels by some corporate IT executives, including officials at Ford
>> Motor Co. They see the use of clusters as a setback for supercomputing
>> because the hardware doesn't push technological advances (see story).
>>
>> [http://www.computerworld.com/hardwaretopics/hardware/story/0,10801,94607,00.html]
>
>
>
> Wow! And here I thought we were trying to use cost-effective hardware to
> solve the problems. I thought the technical advances came from solving
> the problems. Sounds like sales talk to me, we would get the same answer
> using more expensive computers?
>
The critic from Ford isn't a computer salesman...he's a buyer, and his
comments are far from unique.
The question isn't, will you get the right answer, but will you get an
answer in time. There are zillions of different ways of using computers
for applications, but I have most often seen computers used successfully
in ways that are very little different from the mythical envelope: as an
aid to visualization and thought. In that mode, it's more important to
get the drawing scratched out when the ideas are hot than it is to make
the lines beautiful. Workstations meet that need nicely. Much of the
space that was once filled by more powerful boxes (say, from SGI) has
been taken over by workstations as workstations have become more powerful.
What do you do, though, when the workstation won't give you the answer
fast enough? You put together a bunch of workstations into a cluster.
Getting the bunch of workstations to produce the answer faster is often
possible, but it takes significant work, and you are definitely out of
back of the envelope territory.
The government has mostly been funding clusters. What industrial users
could most readily use is a bigger, faster envelope.
> Is this the same theory which holds that getting from point A to point B
> is somehow better for your business if you go first class? 'Scuse me if
> I don't buy it, if a computer vendor wants to "push technological
> advances" they can do it on their dime, I want to push my bottom line.
>
By the same logic, we don't need that national labs playing in the
sandbox to build clusters. They are easy enough to build. If the DoE
is going to claim to advance technology, then they should be trying
things that really would advance technology. The DoE's stated desire is
to help make HPC available as a competitive tool for industry. If they
really want to do that, then I would think that what potential
industrial users think would matter a great deal.
>>
>> SGI officials said that criticism doesn't apply to their system, which
>> uses a single image of the Linux operating system in each of its large
>> nodes and has shared memory.
>>
>> David Parry, senior vice president of SGI's server platform group,
>> said the system is "intended to solve very large complex problems that
>> require the sharing of capabilities of all of those processors on a
>> single problem at the same time."
>>
>> </quote>
>>
>> The criticism doesn't apply to their system? Say what? A single
>> system image sounds nice for a gigantic blob of a problem with so many
>> unique pieces that just getting the crudest screwups out of the
>> simulation (never mind how well it corresponds to reality) sounds like
>> a major challenge. Make the problem big enough, though, and a shuttle
>> mission is about as big as it gets in terms of what people are
>> actually going to try, and you still have to cluster.
>
>
> I can find an emoticon which means "I'm looking for your point," but I
> am. It doesn't matter how you solve a problem as long as you get the
> right answer. And using different computers would not help NASA (or
> anyone else) if they ask the wrong questions.
>
It matters a great deal how you get the answer if you've already cut
metal by the time the answer is available.
Some HPC problems are highly-predictable with lots of structure. Those
problems are at least conceptually suited to being partitioned onto a
slew of identical nodes. Big system simulations aren't likely to have
that kind of conveniently crystalline structure, and getting the work
distributed over a large number of nodes in a reasonable fashion could
be comparable in difficulty to the problem itself. Ideal for a big SMP
box, but the problem is so big that you wind up with a cluster, anyway.
In the end, it _is_ a cluster, and the whole simulation could be
bottlenecked by box-to-box communications required by only a small
subset of the simulation.
> This is a group for discussion of Intel compatible hardware, I think we
> can leave the politics out of it, thanks.
Altix uses Itanium, and purchases of very large computers are inherently
political.
RM