18 (Материалы к экзамену), страница 9

The three things most needed in a Supercomputer are fast arithmetic,

large memories, equally fast I/O, and a usable, robust operating

environment.. {So I can't count ... either.} A typical problem may

require between 10^14 and 10^17 arithmetic operations; clearly, the

floating point unit has to be big and fast, and so does the memory

(typically each floating point operation is responsible for 24 bytes

of memory traffic.) * Very large memories are needed to

accommodate the billions of WORDS needed for each data set, and the

memory traffic on average has to move 3 words (24 bytes: 16 in and 8

out) per floating point operation..

It is typical that every data point is re-computed every cycle and

it is usual that on each cycle each point requires tens of thousands

of arithmetical operations. Often there is an enormous gush of

output. Number ranges are often too big for anything other than

floating point. Notwithstanding, there is always a problem of

retaining some numerical significance so multiple precision

arithmetic is vital. Perhaps you can guess that there is just a

limited number of genuine Supercomputing applications, but even this

number shrinks dramatically when we try to see who is willing to pay

for developing a given problem.

Given that there is a problem that needs to be run on a Supercomputer,

it should be obvious that it won't run constantly, so hundreds or

thousands of smaller problems can be run during such gaps. A common

mistake among certain critics is to say that this is a waste of

Supercomputer resources, or that it is simply not cost-effective. It

should be clear from the foregoing that the criticism is simply

wrong-headed. Supercomputing is not defined by these little problems,

but they can and do benefit enormously by being run on a big machines.

A little problem on a big machine is easily managed, but it becomes a

big problem on a little machine, So, instead of wasting a lot of time

trying to figure our how to fit it into a small or otherwise inadequate

machine, the developer is free to concentrate on making the problem do

what he wanted it to do. (I admit that some people like to squeeze

every last twitch out of a computer. To others, such behavior is

unseemingly for real gentlemen.)

The choice of the price of a Supercomputer is not clearly useful. To

see this, we review some (cynical) definitions of Supercomputing.

1: The first idea we considered is that a Supercomputer is that

machine which will run your problem the fastest. Notice it's your

problem, and it's not necessarily the machine you are currently using,

and the concept of "fastest" is one you should control. You might

reasonably decide that it's total through-put time that matters rather

that how fast one can do arithmetic.

2. Showing some real insight, Ken Batcher stated that,

"A Supercomputer is one that will change your compute-bound problem into

one that is I/O-bound." Notice here that the limitations related to

I/O are intrinsically tied to the nature of the problem. Words like

"entry-level," and "mini-super" have entered the lexicon via

over-zealous marketers and salesmen in order to sell things that are

not necessarily supercomputers. (The term itself first appeared about

twelve years ago. Who first used it is subject to argument, but Jack

Worlton was one of the first to speak about such machines, and today is

a leading advocate in trying to change the term to (Ultra) High Speed

or Performance Computing, because as he notes, marketers have

thoroughly trashed the original meaning of "Supercomputer.")

Typically the inadequacies of small computers reside largely in their

inability to support adequate memory traffic and FAST I/O.

3. Still another kind of insight: Neil Lincoln says that a

Supercomputer is one that is only one generation away from being what

you really need.

We can observe here that it is the application that decides if you're

using a Supercomputer, and finally to put extra stress on the idea that

"the concept of "fastest" is one you should control." consider what you

would hold important if all the arithmetic in your problem took zero

time. How fast would your problem run, and now what would you do to

make it go faster.

I hereby apologize for the obscene length of this polemic.

Cray Research managed to build it's first shippable Cray 1 for

under $10Million. Of course, it did not have much software.

Convex managed to build it's first shippable C-1 for around $20 Million,

including software.

Others have also gotten to market with innovative hardware for a lot less

than $100 Million. Critical issues for startups:

1) Know what your initial target market is and understand what it requires.

2) Maintain focus and don't allow significant investment in anything that

does not aid that market.

3) Do everything you can to minimize time to market. Every extra month just

burns more money.

4) Keep your initial staff small (just large enough to do the job).

More people mean more time spent communicating instead of doing.

If you are very selective in hiring, only accepting the top 10% of

potential candidates, and motivating them with "average salaries and

extraordinary stock options, plus exciting work", you should be able

to get several times industry average effort and productivity.

5) Use other people's work whenever possible - like starting with Unix as

an OS instead of inventing your own. Look for strategic partnerships

in as many areas as possible. These efforts will help (3) and (4).

6) Be lucky. :-)

[It helps to be first one to market in your niche, or for your target

market to experience a business boom just as your product is ready.

It also helps for your vendors to deliver what they promise. The

reverse of any of these can sink you. Examples abound.]

Memorial contributions be

made to the Pikes Peak Area Trails Coalition, 1426 N. Hancock, Suite 4,

Colorado Springs, CO 80903 or the Seymour R. Cray memorial at the

University of Minnesota.

The following interview with Seymour Cray appeared in the December 1982

issue of Datamation magazine. The interviewer, Jan Johnson, posed the

same questions to four people: Gene Amdahl, Victor Poor, James Thornton,

and Seymour Cray. There were also follow-up questions addressed to

individuals. Transcribed below are the questions addressed to Cray, and

his answers. All ellipses are in the text as it appeared in Datamation.

A lot of Cray's answers are memorable.

Tom Ace

tea@netcom.com

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

Datamation: What technological developments in the past five to 10 years

have had the biggest impact on your niche of the computer industry?

Cray: I guess there haven't been any.

Datamation: No developments in mathematics, architecture, or in new ways

of looking at things?

Cray: Well, the problem I have with probably most of these questions is

that I don't pay much attention to what is going on in the world. I just

do my own kind of work, so if there were something new in mathematics, I

wouldn't know about it.

Datamation: What have been some of the driving forces behind the changes

in your niche of the industry?

Cray: If Fairchild would quit trying new technology out on us, we'd get

our parts a lot faster. They are always giving us this new technology

and of course it doesn't work, so they have to keep trying it again. It

seems like a real deterrent to getting our job done because it's never

necessary to have the new technology.

Datamation: What do you call state of the art?

Cray: I suppose that it is whatever you can do.