[AndyR]

Folks,
Perhaps this would be better as a poll, but I don't think we have that for CFD-Online yet.

I have been asked by my management who is CFD illiterate to show where our computational facilities are relative to some "state of the art". I think this might be best described in terms of what a "state of the art" mesh size might be, and what a "a state of the art" turn around time might be, and then compare that to what we can do with our available resources. I thought a time line might also be helpful.

So I ask you;

What would you consider current mesh sizes for:
Grand Challenge
Large but infrequent (once a year within an analysis team)
Large but frequent (4 or 5 times a year within an analysis team)
Daily driver / typical mesh size
Quick and dirty problem

And then turn around time for the problem

My thinking is that for 2010 a state of the art facility should be able to run:

Grand Challenge - 1 billion cells / 2 months
Large infrequent - 50 million cells / 3 to 4 weeks
Large frequent - 10 million cells / 2 weeks
Daily driver - 3 million cells / 1 week
Quick and Dirty - 500K cells / 2 days

Note my times are non-linear. I am assuming that the smaller problems are sharing compute resource with more jobs and the Grand Challenge gets every clock cycle available.

I am still thinking about what those numbers would have been in 2000 and 1990.

Any help is appreciated

Thanks
- AndyR

[ericnutsch]

Originally Posted by AndyR

Perhaps this would be better as a poll, but I don't think we have that for CFD-Online yet.

There is a way to do it, I have never taken the time to figure it out but here is one:
http://www.cfd-online.com/Forums/ope...-openfoam.html

Originally Posted by AndyR

My thinking is that for 2010 a state of the art facility should be able to run:

Grand Challenge - 1 billion cells / 2 months
Large infrequent - 50 million cells / 3 to 4 weeks
Large frequent - 10 million cells / 2 weeks
Daily driver - 3 million cells / 1 week
Quick and Dirty - 500K cells / 2 days

The resolution sounds accurate. I suppose the time depends the solution type, residual targets, transient values, hardware, etc.

I recently built a high end singe processor desktop system with enough ram to handle 12million cells. If the solution time is indeed linear with respect to the number of cells, I should be able solve a 12million cell steady state solution in a little less than 3 hours(running all 4 cores). Obviously those numbers go up when transients are involved, but 2 weeks(336hr) seems pretty significant for a state of the art facility.

Hope that helps!

FYI: http://www.cfd-online.com/Forums/har...benchmark.html

[acampbell]

As always the answer will be "it all depends"!

If it is a complicated multiphysics problem, you can spend a few months getting a commercially reasonable result with a couple of million cells/nodes.
Also the experience of the operator will also impact on the time it takes to complete the work, also if you have a previous example this helps a lot......

[ericnutsch]

AndyR,

I figured out how to start a poll:

Click the "Thread Tools" in the submenu at the top of the forum.
Then click "Add a poll to this thread"

It will only let you do it if you are the thread starter(obvious but relevant).

Cheers!