The 30/30 Rule

[bhall]

[...]

Though the rule doesn't, your explanation makes sense.

Much obliged.

[rjsa]

What the rule really does is an equivalence formula.

They are matching 8 weeks of non stop processing from a 1 TERAFLOP/s processor with one our worth of wind tunnel.

If you take a 100 TERAFLOP/s system, one continuous week of it's use equals 12.5 hours of wind tunnel.

And then on.

EDIT: Was very confusing and the math is still shaky, thing I got it right this time...

[tim|away]

3.6 The Limit Line is defined as follows :
WT <= WT_limit (1 – CFD/CFD_limit)
Where :
WT = Wind On Time
WT_limit = 30 hours
CFD = CFD Teraflops Usage
CFD_limit = 30 Teraflops

So, we are talking about this:



During the 8 week Aerodynamic Testing Period, teams will have to decide how much 24/7 computing power they want (in Teraflop/sec) and they will get a maximum allowance of hours for the windtunnel as a result. Of course, one could also do it the other way around by choosing an allowance for the windtunnel first and then get a resulting TeraFLOP/sec figure.

At 30 TeraFLOP/sec a team would get no windtunnel time. Equally, at 30 hours of windtunnel, they wouldn't have any allowance for their CFD. In reality therefore, they'd choose something in the middle.

Example: A team decides to run their CFDs with 20 TeraFlop/sec and gets in return an allowance of 10 hours in the windtunnel.



It's actually even more messed up if one cares to be picky: The rules are stating that there is a 30 TeraFLOP/sec limit over 8 weeks, but this is actually calculated by averaging the total FLOP used over the 8 weeks in order to carry out the simlations. In reality, we are talking about 4,838,400 seconds (8 weeks) * 30 TerraFlOP worth of computing power that can done with as much computing performance (read TerraFLOP/sec) as one likes, as long as the 30 TerraFLOP/sec average is not exceeded. Depending on how many total FLOP a team has used up, they will get a windtunnel allowance.

It's a pretty convulated way to put things. A slightly more straight-forward way might be: Every team gets a maximum computing limit of (30 * 4,838,400 * 10^12) = 1.45152 * 10^20 floating-point operations over the 8 weeks of testing. For unused floating point operations a team receives windtunnel time as follows: 1 minute windtunnel time equals (30 * 4,838,400 * 10^12) / (30*60) = 8.064 * 10^16 of unused floating point operations.

[PlatinumZealot]

Basically saying per calculation you can't have use horsepower more than 30 TeraFLOPS. This is saying, in order to cut costs you can't have a super duper tripple flux core capacitor computer.

[beelsebob]

The way this works in cost saving is simple. It's pointless owning a super computer that can perform more than 30TFLOPS. Teams do not need to continuously invest in new computers, because once you have a 30TFLOPS one, you can do all the allowed computation in the allowed 8 weeks.

[mertol]

Yea that's why sauber are upgrading to 57tflops...

[Richard]

As Tim Away highlighted, the rule is based on the total use over 8 weeks. You can use 30 all the time, or a burst of 60 for half the time. This means the analysis time can be interleaved with production, ie analyse a problem, make the part, run it on track or the tunnel, then back to CFD to analyse the next iteration.

There will be multiple development paths, some with fast timescales that rush straight to track, and some will be a slow burn, for example developing the 2015 car. The more advanced CFD packages have processor managing tools to handle parallel work streams so the small quick jobs can flash past the large slow jobs.

I presume that the allowance only relates to actual CFD solvers, there will be a huge amount of time in pre and post processing. That will gobble up some of the computing power in addition to the solving allowance?

In my workplace we're limited by the humans - extra CPUs can be bought off a shelf but the PhD level operators are harder to come by! So the the greatest efficiencies are gained with effective pre and post processing because that requires human interaction. In contrast the CFD can be left to run overnight or weekends. I suspect the smaller teams will still be human limited, while the big teams will be CPU limited with surplus humans.

Edit - There's another factor.... The CFD analysis can be simplified so only looking at once element only requires a small amount of processing, also the mesh size can be adjusted. So CFD is a more subtle tool in terms of managing the allowance. In contrast a wind tunnel is either on or off, you can't use just one blade on the fan for a simple check!

[Callum]

Is it a simple question to ask how long a 1million cell simulation would take to run using 30TeraFlops of power? What sort of cell counts are being used for CFD work in industry these days?

[Richard]

How long is a piece of string? I think mesh sizes would be a more useful metric?

[beelsebob]

Yea that's why sauber are upgrading to 57tflops...

Notice how that says "BMW Sauber" and not "Sauber Ferrari". That's because the article is from a long time ago, when the rules specified a limit that was higher.

Is it a simple question to ask how long a 1million cell simulation would take to run using 30TeraFlops of power? What sort of cell counts are being used for CFD work in industry these days?

What's the stepping interval, and how long are you simulating for? You can make a reasonably informed guess that a single instance of computing a single step of the navies stokes equations is going to take something like 5 vector operations, the multiplies work out to being 5 scalar ops each, and the adds 4 each, so you're looking at around 22 scalar flops per cell per stepping interval. Which means 22MFLOP to compute a single step. Which means you're looking at roughly 1,500,000 steps per second on a machine like this. Of course, this makes some poor assumptions, like that you can make the machine perfectly compute steps all the time. The reality is that there's a lot more to CFD than simply repeatedly applying the equations, and you're also never going to perfectly distribute your computation over all the machines involved, so the reality is that it'll be a lot slower than that. This should give you an idea of the order of magnitude of speed we're talking about though.

[Glyn]

Has anyone got any videos / articles / information explaining the benefits and drawbacks of cfd vs windtunnel. Which order teams will use either. Etc etc. This is all very interesting. But is never mentioned in the media.

[beelsebob]

Has anyone got any videos / articles / information explaining the benefits and drawbacks of cfd vs windtunnel. Which order teams will use either. Etc etc. This is all very interesting. But is never mentioned in the media.

The short answer is this:
The wind tunnel still tends to be more accurate than CFD
CFD allows much much more flexibility in what you're testing - it allows you to deal with the wind coming from odd directions, or even the direction changing as your simulation runs (e.g. it allows you to simulate a car turning).

[Richard]

benefits and drawbacks of cfd vs windtunnel

Imagine you are going to refurbish a room or your garden. You'd find it a lot quicker to draw on paper and adjust that with pencil & eraser than to build a balsa model to see if everything fitted? Its the same for CFD versus physical models.

CFD allows rapid iterations without the cost/time of model making. Also the computing power can be divided to work simultaneously on several jobs but a wind tunnel can only have one model at a time. Given the quota on CPUs and tunnel time, CFD can save precious flops by focussing processing power on part of the model with a detail mesh while the rest of the model is coarse, whereas a wind tunnel is on or off. In extreme, a CFD model might only have half the car (ie split down the centreline) which halves the model size, hence saving flops.

Alas CFD is not as accurate as a wind tunnel, so a tunnel is best at confirming the CFD numbers. That's even true with buildings, they're relatively simple static bluff bodies compared to a car.

[beelsebob]

Another advantage of CFD is that if you cache known flow patterns, you can test tons of parts with only a model of that part, in isolation. You can for example record how the air flows off a front wing/splitter/sidepod assembly, and then test 1000 different diffusers in that flow pattern without ever simulating the front of the car again

[Cold Fussion]

Why is the computing limit so low? A 30 teraflop super computer would surely be quite inexpensive these days.