By: Robert Myers (rbmyersusa.delete@this.gmail.com), October 19, 2012 3:14 pm
Room: Moderated Discussions
anon (anon.delete@this.anon.com) on October 19, 2012 2:34 am wrote:
>
> 1. Now if you think
> worthwhile problems can be solved with current trends in HPC, why do you
> disagree with the degree of spending? Do you have an economic analysis or
> studies showing a different direction that would give a better return on
> investment for science?
>
You sound like an Ayn Rander. Scientists are not economically rational actors. Bureaucrats are not economically rational actors. Human beings are not economically rational actors. Why people act in the predictably irrational ways that they do is a research subject for people in areas where I have some interest, but no credentials, and no interest in pursuing them.
One of the participants in this conversation, if I have identified him correctly, has difficult choices to make. I do not envy him or anyone else faced with similar choices. It's easy to criticize, as I have done frequently. In general, I would not be spending money on warehouse-sized computers right now, but it is not my decision to make. I have no more rational basis than he does for my prejudices, except that he talks to many more people than I do. If you think crowd-sourcing is an appropriate way to drive science, then his is the better judgment. I sit in the peanut gallery.
> 2. Do you have any studies or data showing highly
> parallel algorithms are inaccurate (problematically less accurate) than less
> parallel ones, for fluid dynamics.
The limitations of the most widely used methods in CFD are well-known. There would be little point in trying to explain them here. Fortunately, fluid mechanics is fairly forgiving in some ways. Some very nice work showing how a bumblebee manages to produce a vortex pair with a single flapping wing was discussed on comp.arch. Understanding the mechanism is a genuine breakthrough. It didn't require a great deal of accuracy, and I doubt if it required much in the way of computational muscle. It required imagination, insight, and understanding, which are far more important than flops, even in CFD.
Other questions in Fluid Mechanics, like the inviscid (infinite Reynolds number) limit of the Navier-Stokes equations or the boundedness of solutions to initial value problems are not so forgiving. One of the Millenium Prizes on offer inquires about the boundedness problem, either for the finite or for the infinite Reynolds number case. In order to gain insight into such things, the lazy methods that are well-suited to warehouse computers are not adequate. The details of how you carry out the differencing have profound implications. One poster (on this forum, I think) claimed that the methods they are using reduce to the proper (Fourier representation) form in the limit of infinite resolution. If you wildly-over-resolve the problem, you can make the problems go away. How much over-resolving do you have to do? When you don't understand the physics in the first place, which is the only reason you're doing the calculations, I don't know how to answer a question like that. I'd rather do the problem using methods where I don't have to make wild guesses about the appropriateness of the method to the mathematical questions I want to ask.
Different people want different things from computational physics. I want to use computers to do the best math possible.
Why do people stand in line to use these big boxes? I have an easy answer for that. In most cases, for most problems, you will get an answer, and it will often produce a photogenic color plot suitable for the alumni bulletin. It's no different from what telescope time is for astronomers, except that astronomers are *always* looking at something real. The same cannot be said for computer simulations.
Robert.
>
> 1. Now if you think
> worthwhile problems can be solved with current trends in HPC, why do you
> disagree with the degree of spending? Do you have an economic analysis or
> studies showing a different direction that would give a better return on
> investment for science?
>
You sound like an Ayn Rander. Scientists are not economically rational actors. Bureaucrats are not economically rational actors. Human beings are not economically rational actors. Why people act in the predictably irrational ways that they do is a research subject for people in areas where I have some interest, but no credentials, and no interest in pursuing them.
One of the participants in this conversation, if I have identified him correctly, has difficult choices to make. I do not envy him or anyone else faced with similar choices. It's easy to criticize, as I have done frequently. In general, I would not be spending money on warehouse-sized computers right now, but it is not my decision to make. I have no more rational basis than he does for my prejudices, except that he talks to many more people than I do. If you think crowd-sourcing is an appropriate way to drive science, then his is the better judgment. I sit in the peanut gallery.
> 2. Do you have any studies or data showing highly
> parallel algorithms are inaccurate (problematically less accurate) than less
> parallel ones, for fluid dynamics.
The limitations of the most widely used methods in CFD are well-known. There would be little point in trying to explain them here. Fortunately, fluid mechanics is fairly forgiving in some ways. Some very nice work showing how a bumblebee manages to produce a vortex pair with a single flapping wing was discussed on comp.arch. Understanding the mechanism is a genuine breakthrough. It didn't require a great deal of accuracy, and I doubt if it required much in the way of computational muscle. It required imagination, insight, and understanding, which are far more important than flops, even in CFD.
Other questions in Fluid Mechanics, like the inviscid (infinite Reynolds number) limit of the Navier-Stokes equations or the boundedness of solutions to initial value problems are not so forgiving. One of the Millenium Prizes on offer inquires about the boundedness problem, either for the finite or for the infinite Reynolds number case. In order to gain insight into such things, the lazy methods that are well-suited to warehouse computers are not adequate. The details of how you carry out the differencing have profound implications. One poster (on this forum, I think) claimed that the methods they are using reduce to the proper (Fourier representation) form in the limit of infinite resolution. If you wildly-over-resolve the problem, you can make the problems go away. How much over-resolving do you have to do? When you don't understand the physics in the first place, which is the only reason you're doing the calculations, I don't know how to answer a question like that. I'd rather do the problem using methods where I don't have to make wild guesses about the appropriateness of the method to the mathematical questions I want to ask.
Different people want different things from computational physics. I want to use computers to do the best math possible.
Why do people stand in line to use these big boxes? I have an easy answer for that. In most cases, for most problems, you will get an answer, and it will often produce a photogenic color plot suitable for the alumni bulletin. It's no different from what telescope time is for astronomers, except that astronomers are *always* looking at something real. The same cannot be said for computer simulations.
Robert.
