By: Robert Myers (rbmyersusa.delete@this.gmail.com), October 18, 2012 9:47 am
Room: Moderated Discussions
anon (anon.delete@this.anon.com) on October 17, 2012 8:36 pm wrote:
>
> You have thin skin for someone who is
> happy to talk about everyone else being wrong. I would have thought you'd be
> quite accustomed to being called to answer for your claims.
>
I have a thin skin for people who adopt a presumption of superior knowledge and who post anonymously. I am *very* accustomed to being challenged, and most of this material has been gone over, and over, and over.
I felt some degree of vindication when a summary high-level report from the Federal government acknowledged that not all problems can be broken down into sub-problems that can be done more-or-less independently, which is the only kind of problem that the gigantic computers being built today can cope with with any degree of efficiency. I cited the actual report and the actual quotation on comp.arch, and I'm just not in the mood to dig it out again.
Right now, we can do two kinds of problems: those that are actually embarrassingly parallel, and those that can be made to appear to be sufficiently nearly embarrassingly parallel that the pathetic interconnect on the current generation of computers isn't a big problem. Even then, as I have pointed out from actual numbers produced by the actual guilty parties in this gigantic scam, these machines are often operating at "efficiencies" (actual flops/unit time) in the single digit range (less than ten percent) that made Congress howl about how its money was being spent at LLNL in the nineties.
I'm not saying that these machines are useless or that none should be built. I am saying that they get way too much attention, there are way too many of them, too much money is being spent on them, and they are creating the illusion that the scalability problem has been solved when it actually hasn't been solved.
I got to thinking about this problem, and I have been thinking about it for a long time, from playing with turbulence and FFT's. That just happens to be where my technical knowledge, experience, and credentials are deepest and most relevant, but I think the problem that worries me is really quite general.
Fluid mechanics has a problem that it shares with most other really interesting problems in theoretical physics. There are at least two length scales (and usually there are more) that are wildly different. The entire enterprise is about figuring out how what happens at big-L affects what happens at little-L, and vice versa. In the case of the boundary layer in high-Reynolds number flow, some really brave mathematics has made amazing progress on dealing with situations where you have at least two wildly differing length scales, but we still don't really understand many things about turbulent boundary layers.
When you do whatever it is you do to break the problem up into bite-sized pieces that your nearly embarrassingly-parallel computer can handle, you are doing arbitrary things to the very physics you want to understand. If you can't handle how big L relates to little L with confidence, there is absolutely no point in building gigantic computers.
I could say much, much more. It's not a problem that is confined, as one poster seems to think, to pseudospectral methods. The brain has some very long-range connections that I suspect are critical to how it operates. Blue-Gene thinking almost guarantees that we will go down the wrong path. The deepest, most interesting questions in quantum mechanics and cosmology are also big L and little L problems. Can you handle big L and little L problems on machines like Blue Gene? Sure, but, as I've argued over and over again, you could probably do them better on machines that will never be on the Top 500 list.
Robert.
>
> You have thin skin for someone who is
> happy to talk about everyone else being wrong. I would have thought you'd be
> quite accustomed to being called to answer for your claims.
>
I have a thin skin for people who adopt a presumption of superior knowledge and who post anonymously. I am *very* accustomed to being challenged, and most of this material has been gone over, and over, and over.
I felt some degree of vindication when a summary high-level report from the Federal government acknowledged that not all problems can be broken down into sub-problems that can be done more-or-less independently, which is the only kind of problem that the gigantic computers being built today can cope with with any degree of efficiency. I cited the actual report and the actual quotation on comp.arch, and I'm just not in the mood to dig it out again.
Right now, we can do two kinds of problems: those that are actually embarrassingly parallel, and those that can be made to appear to be sufficiently nearly embarrassingly parallel that the pathetic interconnect on the current generation of computers isn't a big problem. Even then, as I have pointed out from actual numbers produced by the actual guilty parties in this gigantic scam, these machines are often operating at "efficiencies" (actual flops/unit time) in the single digit range (less than ten percent) that made Congress howl about how its money was being spent at LLNL in the nineties.
I'm not saying that these machines are useless or that none should be built. I am saying that they get way too much attention, there are way too many of them, too much money is being spent on them, and they are creating the illusion that the scalability problem has been solved when it actually hasn't been solved.
I got to thinking about this problem, and I have been thinking about it for a long time, from playing with turbulence and FFT's. That just happens to be where my technical knowledge, experience, and credentials are deepest and most relevant, but I think the problem that worries me is really quite general.
Fluid mechanics has a problem that it shares with most other really interesting problems in theoretical physics. There are at least two length scales (and usually there are more) that are wildly different. The entire enterprise is about figuring out how what happens at big-L affects what happens at little-L, and vice versa. In the case of the boundary layer in high-Reynolds number flow, some really brave mathematics has made amazing progress on dealing with situations where you have at least two wildly differing length scales, but we still don't really understand many things about turbulent boundary layers.
When you do whatever it is you do to break the problem up into bite-sized pieces that your nearly embarrassingly-parallel computer can handle, you are doing arbitrary things to the very physics you want to understand. If you can't handle how big L relates to little L with confidence, there is absolutely no point in building gigantic computers.
I could say much, much more. It's not a problem that is confined, as one poster seems to think, to pseudospectral methods. The brain has some very long-range connections that I suspect are critical to how it operates. Blue-Gene thinking almost guarantees that we will go down the wrong path. The deepest, most interesting questions in quantum mechanics and cosmology are also big L and little L problems. Can you handle big L and little L problems on machines like Blue Gene? Sure, but, as I've argued over and over again, you could probably do them better on machines that will never be on the Top 500 list.
Robert.
