Comment by noobermin | Hacker News Reader

Comment by noobermin | Hacker News Reader

Parent Full thread

I think people know this, they just don't want to admit it. As someone who works with people who used to work with ICF, I've been told the reason that NIF flunked is that it was designed with 1D simulation codes because back then, that's all they could manage. One of the issues they inevitably ran into were plasma instabilities similar to Rayleigh-Taylor instabilities[0] that one only sees in two (or more) dimensional systems, but not 1D. These instabilities reduce the efficacy of heating and compression needed to achieve ignition, and were an unpleasant surprise when they actually performed the experiments. Even to this day, 2D simulations are fairly common and littered throughout the literature, and their justification (arguments of approximate symmetry) are suspect, at least to me. 3D sims are coming on the scene, but they often have to sacrifice things like spatial resolution that hurt's the simulation's physics in other way.

Computational feasibility is one of the things physicists in that field need to admit is holding them back, and may be a little humility in seeking out advice from computer scientists is warranted in my opinion[1].

[0] https://en.wikipedia.org/wiki/Rayleigh%E2%80%93Taylor_instab...

[1] this statement won't give me much criticism here, but it wouldn't go down well in a room of my peers.

Honest question: do you really think talking much with computer scientists would help? My impression is that most of the improvements (besides Moore's law) in the power of simulation of physical systems has come from either new physical insights, or computational insights developed by physicists who were mostly self-taught numerical techniques. It seems very plausible to me that there are honest-to-goodness CS insights that remain unexploited, such as better parallelizing physical simulations, but the track record hasn't been good. Why aren't CS people coming in and showing everyone how it should be done?

(For what it's worth, I'm a physicist, but not one who does anything numerical, and I'm happy to admit arrogance by physicists could be a problem here.)

It's not all about algorithms. Unless current simulations are running on Asics or even GPU's there are probably 10-1,000+x speedups possible.

Why do you assume that the computational fluid dynamics folks are dumb and haven't done this?

CFD is a very vibrant field with applications across the entire spectrum of engineering.

In fact, modern supercomputers are MOSTLY running CFD codes (weather prediction is effectively a gigantic CFD simulation). Nothing else really demands that level of power.

I know some CFD is still run on CPU's and some is run on GPU's. But, I don't know what the fusion tradeoffs are.

I'm quite sure a good number of these are running partly on GPU's. People are quite aware of them; their practical use in research applications is still a bit limited because of how often the code needs to be modified -- but in industry where the same basic model is run over and over they (as well as FPGA's) are pretty common. Not every computational application is the same :)

well, it's not the job of a physicist to understand the many many different sorting algorithms, how to optimise loops so they are cached in the cpu, and plenty more.

I guarantee that most CS professionals (CS; not necessarily developers) are able to write much more efficient code than most professional physicists, because they spent the same amount of time understanding software and hardware that these guys spent building a fusion reactor.

that said, considering the billions that have been spent on fusion research, I doubt they did it in a bubble. it would seem strange if they didn't enlist the help of some CS

Not sure about this project, but a lot of physics simulations are written by computational physicists, who don't really do actual lab work, and instead spend a lot of their time coding and optimizing. These folks are often aware of things like caches or branch prediction penalties or gpu's (mostly I see MPI codes running on largish clusters though). This doesn't mean that there is no room for optimization, most likely there's some because I am yet to see code that doesn't leave any, but it's probably not something an average CS expert would easily find.

CFD is also used in computer graphics (games and film) which is filled with people who focus on low level optimisation.

I'm doing a PhD in physics using using large scale numerical codes. From what I can see more computer-science would mainly improve the maintainability rather than the speed of codes used. They are often already heavily optimized and, for the numerical heavy lifting rely on standard libraries such as Intel's MKL. Also scalability is usually quite good since the codes are routinely developed to run on a large number of CPUs in parallel. GPU support is coming slowly since it's also new to the people who developed Fortran codes for the last 20 years. For example VASP, a popular atomic structure simulation, is getting GPU support just now through a public-private partnership with nVidia. The speed-up for a 16-core sandy-bride machine with and without two GPU cards is about up to factor 8.

CS people aren't coming in and showing how it should be done probably for the same reason that Physicists don't show up at a bridge construction project and show the engineers how it should be done.

Theoretical computer science is quite distinct from coding and large scale numerical computation practice. Just as theoretical physics, or even experimental physics, is quite different from structural engineering.

Except that physicists do come in and show biologists how it's done for all sorts of mathematical problems in biology! People call it "biophysics" but very often it's things where there is really no physics per se, like statistical analysis of genomics.

That might be because the Physicists had help from Mathematicians before and are just passing it on. I don't think that's quite true though, i like to believe that polymaths and university aren't quite dead in academics.

If you ask me they mostly need software engineers, not “CS people” especially. Whenever I talk to a scientist about their software, I see great opportunities and interesting problems that are solved in suboptimal ways. To be blunt, scientists can't code (and neither can most “CS people” for that matter). Every scientist (be it physics or CS) could use a professional software people, who have experience with developing and maintaining complex code bases. The problem is, there is no funding. Let's say a researcher gets a budget for some hardware and 10 PhD students, then there is nothing left to pay the engineer with. We're too expensive.

Essentially what you're talking about is a lab assistant who can code? Sounds like a good idea, and might be something for people who work at companies that do pro-bono charity work.

What's really funny is I started in software (have code committed to some open source projects... BIND, Valgrind), and then moved to Physics. The code is garbage because the constraints are totally different. Experiments and code can be brittle and still work. There's no time to consider user experience for fellow grad student when there's so much else to do... though I do know of SOME students who make time to write good code. Hats off to them.

I'm a mechanical engineer in physics and I've had no formal training in writing code or computer science, yet, I am now doing simulations for atomic physics in Python. I am designing some software library package which will serve as a sort of diagnostic/design tool for future experiments with a machine being built.

I've run into the problem of figuring out how to organize the whole dang thing and sometimes have to go back and re-write a lot of code. Can you recommend any good books that introduce computing concepts for applied physics and engineering?

My favorite general guidebook for writing good, readable, maintainable, defensive code is "Code Complete": http://www.amazon.com/Code-Complete-Practical-Handbook-Const...

It isn't domain specific but there is so much good advice in there around variable naming, code structure, general principles of bug-resistant coding that you will get a tremendous amount from it.

Not sure that would help though. You want someone with the theoretical knowhow to understand the problem at hand, so you can rewrite it to be simpler to program. In other words, a CS major with programming skills. They are in very high demand.

Software quality is just not something academia optimizes for. It only counts whether you can get your paper out. The chances that somebody else actually uses your software are slim, so investing time that could be spent writing more papers into better engineering makes little sense.

Optimizations, GPUs and even parallelization are not always much use in simulations because the computational complexity is usually very high - like O(n^3) or so. That's not even counting the cubing of the problem size in going from 1D to 3D. If you can get 10x the performance out of some optimization, that can allow you to solve a slightly bigger problem - but isn't likely to be dramatic enough to go from 2D to 3D.

There are bigger optimizations than 10X with some of the adaptive grid / wavelet compression stuff. I think these are still 2D, but pretty amazing: https://youtu.be/vYA0f6R5KAI

I was going to link to this too. It's almost as if the internet is calling for us to write an app called "here, the internet wants to help you researchers... have a video..."

> Why aren't CS people coming in and showing everyone how it should be done?

Because CS students don't learn differential equations or real and functional analysis.

Have a B.S. in CS and I learned both differential equations and real analysis.

?

What school did you go to? What school would issue a BSc in computing science without analysis and differential equations?

I did undergrad at UMass Amherst, where computer-science students take math up to multivariable calculus 1 and linear algebra. I then did my MSc at the Technion, where some amount of differential equations and vector calculus is normal for undergrads, but even there, requiring multiple semesters of analysis to grant a degree in computing, where most of the maths are discrete, is considered overkill.

I think you might need hire engineers. What I mean is, someone in cs wouldn't have any papers to produce from applying well known parallelism techniques to your problem. In academia, papers and innovation in one's field are the motivation. What physicists appear to need is well established and understood. Hire a parallelism expert

> Even to this day, 2D simulations are fairly common and littered throughout the literature, and their justification (arguments of approximate symmetry) are suspect, at least to me. 3D sims are coming on the scene, but they often have to sacrifice things like spatial resolution that hurt's the simulation's physics in other way.

Sounds like the joke about the guy who dropped his keys in the bushes but is looking for them under the streetlamp, because there's light there so they should be easier to find.

Reply on news.ycombinator.com