Distributed CFD: Use the CPU power of fans?

[thedutchguy]

Maybe this is a crazy idea, but as far as I understand CFD calculations are done on huge computerfarms, so chopping data up in small packets to be processed independently is possible right?

Why then don't F1 teams use their huge fanbases and develop a tool which fans can install to do parts of the computing on their PC's? What I mean is something like de Folding@home and SETI@home projects. I'm sure there will be challenges to get it done (and to make sure that the data remains safe from competitors eyes) but the potential benefits of vast amounts of 'free' CPU power might be huge.

Interesting thought?

[horse]

Nice idea, but totally insecure. How would you verify that your competitors are not "helping you out"?

In fact, the real issue here is parallel computing architecture. Parallel processing only works well for large jobs if you have fast communication between the computing nodes. If you have distributed memory nodes it tend to be better run many separate jobs rather then one large job. This is because most CFD packages will need to "see the big picture" as it were at least once in a iteration (this may be in matrix inversion - depends on how the algorithm works) and so once you've sent all that data any to be worked on in chunks, you need to get it back quickly also.

For example we have a 50+ node cluster here made up of 8 node boxes. However as soon as you run a single simulation on more than one box (ie more than 8 nodes) simulation time is increased as the communication between the boxes is just ethernet. The fast links between the boxes are very expensive (I'm told) and make a super computer super.

So I suppose my conclusion is distributed computing is cheap but inefficient.

[hollus]

I am in seti at home since ages. The day when they send some work to my computer and get the results the same day is actually the exception. More typical is that one week goes by. Not very fast for F1 development. Also, only very small pieces of work are given to any given computer. In climate at home, which would be more relevant to CFD calculations, months are needed by each "home" computer for just one calculation.

[Giblet]

I had suggested this a while back, when Albert2 was put on line. Made me think that this would be ideal for teams where budgetary concerns were a problem, like Williams.

The data you get is a useless piece of data by itself, you don't get to see the whole thing, evne if you decrpt it, so other teams would have no point in trying to intercept it, as it would be entirely out of context.

Even with SETI, if they didn't give you the lat and long in the cosmos, you would have no idea what you are processing.

[horse]

I had suggested this a while back, when Albert2 was put on line. Made me think that this would be ideal for teams where budgetary concerns were a problem, like Williams.

The data you get is a useless piece of data by itself, you don't get to see the whole thing, evne if you decrpt it, so other teams would have no point in trying to intercept it, as it would be entirely out of context.

Even with SETI, if they didn't give you the lat and long in the cosmos, you would have no idea what you are processing.

Good point. I take back the security argument.

[Jersey Tom]

Parallel processing only works well for large jobs if you have fast communication between the computing nodes.

This.

[The_Man]

I think It'll be simply cheaper for the teams to buy that extra processor than set the entire network up. Also in such architecture it is really difficult to estimate the time a simulation it would require as it is a function of the number of fans online.

Fast communication is not available all over too.

[thedutchguy]

I am in seti at home since ages. The day when they send some work to my computer and get the results the same day is actually the exception. More typical is that one week goes by.

It all depends on how big the chunks are which are sent out for processing and how much computation they require. And if you don't want to run the risk that someone doesn't return results for whatever reason, just make sure that you send each package to multiple users. Obviously that's only possible if you have a large enough installed base, but I guess that won't be an issue for many teams.

[horse]

It all depends on how big the chunks are which are sent out for processing and how much computation they require. And if you don't want to run the risk that someone doesn't return results for whatever reason, just make sure that you send each package to multiple users. Obviously that's only possible if you have a large enough installed base, but I guess that won't be an issue for many teams.

I still think you'll find, that for a time limited industry like F1, this method would be too slow.

[Pandamasque]

I had suggested this a while back, when Albert2 was put on line. Made me think that this would be ideal for teams where budgetary concerns were a problem, like Williams.

The data you get is a useless piece of data by itself, you don't get to see the whole thing, evne if you decrpt it, so other teams would have no point in trying to intercept it, as it would be entirely out of context.

Even with SETI, if they didn't give you the lat and long in the cosmos, you would have no idea what you are processing.

Good point. I take back the security argument.

Unless they would take part just to sabotage the work.

[FuzzyDice]

Someone said, on this forum i think, that teams are limited by the max number of Tflops so its better if the run those flops on their comp.

[DaveKillens]

When Folding@home was released, I was excited in the fact that I may contribute in a very small way in defeating health issues. It cost very little for me, and I was able to be a participant in this wonderful global effort.

Maybe it's impractical to attempt to use distributed computing on important and time-sensitive problems such as aerodynamics, but to allow fans to participate and contribute is something teams should examine carefully. This is a new method to allow fans and "their" teams to form a closer bond, and in which any fan can feel that they have made a minor contribution to "their" team.

I believe this is a wonderful idea, and if I was Monty, Ferrari would be doing it tomorrow, and using distributed computing for such projects like long-term aero research.