"This comment is nonsense - a good piece of research stands on its own merrits. A large proportion of my fluid dynamics papers contain only a small amount of computational work and actually only a few of my papers contain numerical simulations of the full Navier-Stokes equations.

Following your logic I should stop publishing fluids papers in JFM, EJM/B, QJMAM etc which do not contain huge amounts number crunching on a supercomputer since it would be foolish of me to think that the papers were worth publishing. I have unlimited access to a supercomputer at work and I can tell you that it is very easy to run large simulations but when it comes time to write up the work it is the original idea of why you did the simulations that is important (its just nice if they run quickly)."

What? So now you are comparing with your Own work? My response was to You saying that it's highly unlikely that one might be in a race against time when publishing a paper. So you are doing very few simulation hours and you have unlimited access to supercomputers... good for you, then this might not apply in the same degree to Your work (as stated before).

The "Following your logic" statement makes me believe that you have misread what I have written. I have said nothing about papers containing large ammount of number crunshing etc. Those would benefit more of course from good hardware. I see no reason why you would go into specifics when I give a general statement.

But..

I agree that you are right if you come up with some reference to an investigation stating that CFD researchers generally do Not spend time number crunching. If the general CFD researcher only spends a fraction of his time with an actual simulation (like yourself) then I agree with you. If not then I still say that good hardware is necessary even though You don't need it.

(Obviously I spend alot of time number crunching and hence might not fall into "the good-scientific-work category.")

It doesn't hurt to know something about hardware, and I imagine that any enthusiatic CFD researcher or engineer has some interest in that aspect of CFD, anyway.

However, most people will never be in the position to buy the hardware of their choice. They simply use the best their IT department has to offer.

The argument that you need the fastest machines to win the race to publish your paper is amusing... but not very relevant. I haven't heard of a single case like that. If you publish papers on CFD research, it's either about new methods or new applications. The former case will often only require a proof of concept, spending a lot of time developing the method, certainly not generating huge amounts of data by number crunching. The second case (new applications) is often so specialized that it's not highly contested. In any way, this cannot be the primary reason for knowing hardware.

Not finishing a project, because you're running out of time, because your hardware is too slow....? Unlikely. I can see this happen as you approach the deadline, but most likely it's because you wasted too much time in the beginning of the project, trying to get everything done in the last minute. Getting a job done efficiently is not just about hardware efficiency, but mainly about your own ability to organize your work (with your hardware limits in mind).

But let's say you really are in trouble because you don't have the right hardware. It could happen in some cases. However, in those situations the hardware problem is usually a question of lack of money, not lack of hardware knowledge!

So, what was your point again...? Oh yes, it's important to know hardware. I agree with that, in principle. It's just that the arguments that were brought up are kind of odd.

Well I will continue to get the most out of my hardware regardles of what the "it-department" has offered me. Also I will let my voice be heard before any purchase of hardware that I will use. Most "it-departments" won't lift an extra finger to find the better hardware, they will just order from some predetermined big manufacturer. :P

Also, I very much believe that a 20% simulation time gain is well worth the invested time in learling a bit about HW. And I don't mean cherry-picking components only; how to optimize them is equally important.

Your argument seem to indicate that everyone would be equally happy with gaussian elimination when solving their systems. Why is time important? Why invest several man-hours in developing new algorithms for faster solvers?

With your logic it's like saying that you can dispose your time freely and choose whatever poor algorithm you like that takes forever to solve the problem. Or does it only apply to poor knowledge in hardware?

"Also, I very much believe that a 20% simulation time gain is well worth the invested time in learling a bit about HW. And I don't mean cherry-picking components only; how to optimize them is equally important."

A 20-30% gain/loss can be obtained just by changing the brand of compiler - and probably greater if you choose an algorithm best suited to the computer architecture.

At the end of the day nobody is questioning the need for a knowledge of computer hardware when designing algorithms (a good algorithm on a pc will not be good on a vector supercomputer and what is good on a vector machine is not good on a massively parallel one). The main problem people have is with your comments about "competition and the need for speedier results compared to your competitor". If this were all that was really important then you would never have a hope in science because people like me have unlimited access to far bigger/faster computers than you do.

In industry we adapt the model to the hardware, not the hardware to the model. This is why people did conformal mapping in the the beginning of the 20st century, their only hardware was their own brain, this is where you should start.

In universities or investigation institutes people may be in a lucky position to have better than average equipment at their disposal (usually also more interesting jobs), but even then they won't turn the computer on after having thought about the physics behind the problem.

It's the preparation, the brainstorming, the human errors and the learnign process that takes time in every new design, not running the case. And in the very very very rare ocasion that we really need a strong computer to make something fly, something that is only necessary if the client insists, we RENT some time on a university mainframe. This is reality.

On the other hand, I doubt it that a future is to be found in stronger computers and I wonder whether there exists sort of globus toolkits for aerodyanmics, like they use in astrophysics for parallelising the calculation.

>"And I don't mean cherry-picking components only; how to optimize them is equally important."

You are optimizing your hardware components? How do you do that? Most people optimize their codes, optimize their operating system, BIOS, and other software, but optimizing hardware components by tinkering with the circuits... If you say that's an essential part of CFD I have to disagree. You can get a pretty decent machine by "cherry-picking" a good combination of hardware components. Anyway, it would be fun to see you get the most "out of [your] hardware regardles of what the it-department has offered [you]" by going to your computer lab and tinkering with their stuff. I don't think they will like it.

>"Your argument seem to indicate that everyone would be equally happy with gaussian elimination when solving their systems."

Which argument? I couldn't find anyone in this thread referring to algorithms. You seem confused. We were talking about hardware.

>"With your logic it's like saying that you can dispose your time freely and choose whatever poor algorithm you like that takes forever to solve the problem"

Again, there is a great deal of confusion. No one made such claim or even mentioned anything about algorithms. Everyone understands the importance of efficient algorithms, on any type of hardware, and that's what CFD is really about. You got it.

"A 20-30% gain/loss can be obtained just by changing the brand of compiler - and probably greater if you choose an algorithm best suited to the computer architecture."

Well, yes... How is this not preferable? I don't see your point here.

"The main problem people have is with your comments about "competition and the need for speedier results compared to your competitor". If this were all that was really important then you would never have a hope in science because people like me have unlimited access to far bigger/faster computers than you do."

To answer this I cite myself from the first post. If you read it you will see that I don't imply that hardware knowledge is all-important or that speedier results is "all that was really important". It is just a part of CFD. (You can have the best car in the world and it won't matter if you are a crappy driver, to give another aspect)

"I just want to add one point to this discussion. Hardware knowledge. You might be the best race-car driver in the world, but without a good car you will have a hard time against lesser drivers - with better cars.

You might want do write a paper where you need results from simulations. Considering that there are numerous people doing CFD chances are you are competing against someone else (time). It won't matter if you have perfect BCs etc. if that someone publishes his/her paper before you do. "

With the above I don't connect production speed to hardware alone (at least I don't see it, but if you read it that way then it's my bad and I should be more clear in the future). Of course there are different aspects depending on what you do in you work. Production speed is however almost always a top priority (and No I don't mean that you should abandon quality because of it).

Don't look at it in black and white. You will need all the edge you can get when some of the big countries in the world get up and running CFD on a big scale.

Hey Mani,

I posted above just before reading this but I post some new stuff now.

"Which argument? I couldn't find anyone in this thread referring to algorithms. You seem confused. We were talking about hardware."

The point is that no one would argue in this forum if someone said that a fast and efficient algorithm is important. Right?

"Again, there is a great deal of confusion. No one made such claim or even mentioned anything about algorithms. Everyone understands the importance of efficient algorithms, on any type of hardware, and that's what CFD is really about. You got it. "

Ah, the anwer to the above point.

To state clearly again: I don't claim that hardware knowledge is more or less important than anything in CFD, it is just a part of it, if you like your results faster (than when you don't posess that knowledge).

So you might be limited by your company, no point in Not knowing about hardware just because of it. It's like saying that it is pointless to learn turbulence models outside the scope of your commercial solver package.

If you like I can give you examples where I have increased my hardware output by simple/advanced overclocking (although this is not mainly my point in this discussion). Comparision is made under Suse10.1 (32 and 64bit) and win-xp, using a commercial solver as well as standard benchmarks.

And no I wouldn't temper with something the "it-department" had given me before I started managing my own purchases (through the "it-department"), since it would most likely be a Big-Vendor PC with locked BIOS and crappy motherboard and PSU.

Afraid to overclock, is it something that can damage the computer? Well go and learn something about it on the sites I suggested

yes, overclocking can do some damage if you don't provide for sufficient cooling... and even then there's a point where your system can become less stable, if you're overdoing it.

is this experience or knowledge part of CFD? there's room for interpretation...

to come back to the original question in this post (what do I need to know to run industrial CFD apps), I think a novice can go a long way without knowing which CPU can take how much overclocking on which mobo, and how much cooling is needed to keep it stable...

Dear Mani,

"yes, overclocking can do some damage if you don't provide for sufficient cooling... and even then there's a point where your system can become less stable, if you're overdoing it."

So knowing how to apply it would seem good, right? Btw, not providing sufficient cooling can damage Any system, overclocked or not.

"to come back to the original question in this post (what do I need to know to run industrial CFD apps), I think a novice can go a long way without knowing which CPU can take how much overclocking on which mobo, and how much cooling is needed to keep it stable..."

Agreed, but this is not really the original answer I gave since you base this argument solely on overclocking skills. Though a useful part I meant much more rudimentary knowledge. There are alot of ppl not knowing the difference between a k8 and a net-burst architecture. Those would gladly choose what the "it-department" offered them (which is the Intel alternative 9 out of 10 times). Still speaking of not very high cost equipment of course. And as of previous month a very strong architecture was released, but the main prodction is still in the old poor one (you might know of what I'm speaking). This will undoubtedly lead to alot of poor purchases in the next couple of months just because of lack of hardware knowledge.

The novice might not need this either though. He also might not need any real knowledge in transport phenomena since he can just push buttons and get a result as a previous poster has said.

This discussion is not going anywhere!

The normal situation is that you've got aerodynamic/hydrodynamic guys that know about the physics behind the problem, windtunnel, analytical and numerical (CFD) knowledge, usually spread over a group.

The only one that usually knows something about computer hardware is the CFD guy, because he's closely related to it, but it is certainly NOT a must. I would never contract a guy who knows too much about that, because it usually means that he didn't spend enough time on understanding the physics themselves, the effect of the discretisation, te limitations etc. etc. and I've seen a lot of companies by now.

An IT guy should prepare and know about your hardware and if you're selling software an IT guy should make the GUI and so on, if you're good enough in aerodynamics you simply don't have time to spend on other disciplines.

I personally prefer to develop on machines that are inferior to those of my customers. If my code runs adequately on my machines, it'll fly on the latest hardware.

Well that's another twist to it

(and I certainly see your point here)

However I believe that it is more sensible to develop on latest architectures because of optimazation problems what can occur when transfering your code between old and new platforms. Doesn't apply of course if you by inferior mean "same architecture but slower".

"The only one that usually knows something about computer hardware is the CFD guy, because he's closely related to it, but it is certainly NOT a must."

Never said it was a must. But it's not useless to have as you seem to indicate (or am I wrong?).

"I would never contract a guy who knows too much about that, because it usually means that he didn't spend enough time on understanding the physics themselves"

What? You can't be serious. lol

I imagine the interview with you going:

You - "So do you have any other interests apart from reading about CFD?"

Anxious job-seeker - "Well I do like to hang out with my friends and to excercise."

You - "Aha, I see, ok I think we are done here. Don't call us we call you... (sarcastic grin)"

=)

You just want to be right and have the last word in it.

A company is looking for people with more skills than CFD only and tuning computers is certainly NOT one of those skills we are looking for. It could be music, travelling, sports anything.

To give our poor original poster (from before this thread got completely out of hand!) some idea of where the hardware issue comes in, it is worth looking at where money is spent. In the case of commercial CFD, the cost structure is very roughly something like 60:30:10 between people:software:hardware. This tells you two things:

1. Don't allow expensive people to use up much official time doing non-essential fiddling with hardware

2. Buy decent hardware, and upgrade it frequently. Some good hardware knowledge here is handy, because it will help you identify what you need. Don't agonise about hardware, it should be the least of your concerns. Just make sure you get good equipment. Skimping on hardware is a classic case of penny wise but pound foolish.

In other words, don't mess around overclocking cheap equipment. Buy (or rent time on) proper hardware and focus on getting the job done. Even losing a single run due to an overclock hang-up can cost more than the difference between a bargain CPU and a top-end one.

This of course assumes at least some intelligence and understanding on the part of those doing the budgets and accounting. In my experience those disciplines are without exception totally devoid of anything remotely resembling intelligence, but perhaps you are luckier.

Dear Charles, Mani, Gerrit and Tom.

I sure did like your input and view on the matter, and hopefully our original poster did not take too much offence seing that his thread grew abit large (but interesting imho).

I agree with the 60-30-10, perhaps somewhat less on software and more on people cost. Anyhow, the statement:

"This tells you two things:

1. Don't allow expensive people to use up much official time doing non-essential fiddling with hardware

2. Buy decent hardware, and upgrade it frequently. Some good hardware knowledge here is handy, because it will help you identify what you need. Don't agonise about hardware, it should be the least of your concerns. Just make sure you get good equipment. Skimping on hardware is a classic case of penny wise but pound foolish. "

is abit too general. You can just as easily change hardware for software and get another viewpoint.

Although I have said (perhaps this was not in response to me but it feels that way ) that overclocking was not the main focus of the hardware knowledge, however being an enthusiast I must answer the 1st statement. - Why would overclocking take alot of time? You can soft-clock a computer in less than 2 minutes - if you buy the right Motherboard. I can also add that I have run OCd computers at home for the last 5-6 years without any apparent failure due to the OC (admittedly I'm absolutely not sure about this since a windows-hang-up in the win 98 era was easily dismissed as a software-based problem, not hardware-based.)

"In other words, don't mess around overclocking cheap equipment. Buy (or rent time on) proper hardware and focus on getting the job done. Even losing a single run due to an overclock hang-up can cost more than the difference between a bargain CPU and a top-end one."

I agree here too. I would also say that this is just my point. Hardware knowledge. Without it you might pay for a computer that performs at as low as 60% (perhaps even worse) of a similarly priced one. This is significant for some but not so important for others (though I would call it ignorance of budget). You don't need to be a master-of-hardware (geek?) to qualify for this. A simple browse of hardware pages once in a while will probably be sufficient.

"This of course assumes at least some intelligence and understanding on the part of those doing the budgets and accounting. In my experience those disciplines are without exception totally devoid of anything remotely resembling intelligence, but perhaps you are luckier."

lol