By: Robert Myers (rbmyersusa.delete@this.gmail.com), October 19, 2012 11:40 am
Room: Moderated Discussions
forestlaughing (forestlaughing.delete@this.yahoo.com) on October 19, 2012 9:38 am wrote:
>
> My argument is that *bigger* computers ARE *better* computers,
> at least for solving some problems. I agree with you that they are not better
> for solving all problems. Do all problem scale to 10000 nodes? No, but several
> important problems do. Many more important problems scale to 500 nodes, and
> there are substantial advantages to having 1 10k node system, as compared to 20
> systems, each with 500 nodes. These big systems are very good at solving some of
> the worlds pressing problems, just not the problems you care about.
>
> You also
> misrepresent, to a degree, my main point. It is not that high bandwidth
> computers can't be constructed, it's just that doing so is very expensive, and
> the people who would like such computers are unwilling to provide the money
> necessary to engineer, and also construct them. If you want high bandwidth
> computers, they can be built, you just have to bring a BIG bag of money to the
> table. It's been done in the past, but these huge clusters solve enough of the
> problems, and for relatively little money, that most users learn to live with
> what they can afford.
>
I had no intention of misrepresenting your main point. If people won't provide the money to do what is technologically possible, then, in practice, you "can't" build the better computer, because it is too expensive. I've got that point firmly in my head, which is why, after some very bitter comments about Blue Waters, I've sort of lost interest in the subject.
As to the world's pressing problems, it isn't my place to make the decisive judgments. You know by now that I am very, very skeptical of gigantic computer models. If the problems are important enough to justify the kind of money that is being spent, I'm scared, because my experience would suggest that the chance that any of the simulations will give correct or even useful guidance hovers somewhere in the vicinity of zero. That's a prejudice on my part: the product of actual hands-on experience, listening to all kinds of nonsensical results being presented and incorrectly explained, and nearly every day news about every kind of bug you could possibly imagine in software where the incentive to get it right would seem high enough that, were it possible to get it right, it would have been gotten right.
In practice, we get away with the sloppiness of software because someone makes a mistake, it is discovered by accident, we pay whatever penalty there is for having gotten it wrong, we correct the mistake, and we move on. That approach works just fine, unless the problems really are as important as you claim they are, in which case, the opportunity to correct the mistake and move on may not exist.
I'd prefer that we forget about the headline-grabbing problems and try to do better science, because, without the good science, the "solutions" to the headline-grabbing problems are worthless.
Robert.
>
> My argument is that *bigger* computers ARE *better* computers,
> at least for solving some problems. I agree with you that they are not better
> for solving all problems. Do all problem scale to 10000 nodes? No, but several
> important problems do. Many more important problems scale to 500 nodes, and
> there are substantial advantages to having 1 10k node system, as compared to 20
> systems, each with 500 nodes. These big systems are very good at solving some of
> the worlds pressing problems, just not the problems you care about.
>
> You also
> misrepresent, to a degree, my main point. It is not that high bandwidth
> computers can't be constructed, it's just that doing so is very expensive, and
> the people who would like such computers are unwilling to provide the money
> necessary to engineer, and also construct them. If you want high bandwidth
> computers, they can be built, you just have to bring a BIG bag of money to the
> table. It's been done in the past, but these huge clusters solve enough of the
> problems, and for relatively little money, that most users learn to live with
> what they can afford.
>
I had no intention of misrepresenting your main point. If people won't provide the money to do what is technologically possible, then, in practice, you "can't" build the better computer, because it is too expensive. I've got that point firmly in my head, which is why, after some very bitter comments about Blue Waters, I've sort of lost interest in the subject.
As to the world's pressing problems, it isn't my place to make the decisive judgments. You know by now that I am very, very skeptical of gigantic computer models. If the problems are important enough to justify the kind of money that is being spent, I'm scared, because my experience would suggest that the chance that any of the simulations will give correct or even useful guidance hovers somewhere in the vicinity of zero. That's a prejudice on my part: the product of actual hands-on experience, listening to all kinds of nonsensical results being presented and incorrectly explained, and nearly every day news about every kind of bug you could possibly imagine in software where the incentive to get it right would seem high enough that, were it possible to get it right, it would have been gotten right.
In practice, we get away with the sloppiness of software because someone makes a mistake, it is discovered by accident, we pay whatever penalty there is for having gotten it wrong, we correct the mistake, and we move on. That approach works just fine, unless the problems really are as important as you claim they are, in which case, the opportunity to correct the mistake and move on may not exist.
I'd prefer that we forget about the headline-grabbing problems and try to do better science, because, without the good science, the "solutions" to the headline-grabbing problems are worthless.
Robert.
