[jorgem4]

Ok, so I am a mechanical engineer that works for a large corporation, what I do is called Finite Element Analysis. This process requires of solving Finite Element models. The issue is the following:

It takes a lot of time to solve some of the models that we create at work, we use "crappy PCs" (they are not crappy, but to me anything NOT Mac is crappy) to solve de models...also some times if the model is to big we send it to a UNIX machine to solve it, but still takes for ever!

I was thinking if it is possible to solve this models in a G5 (which I think would be much faster) than the UNIX machines we have here. See we can only load our machines with 1 GB of RAM...and the G5 can take more ran and a faster processor.

Do you think it is possible to set a G5 to run these solves and do the crunching much faster?

Thanks.

[-Q-]

I know nothing of solving Finite Element models, but I'd imagine if you throw more power and more RAM at something, it'll get done faster.

I wonder if this might also be good for a 'render farm' type of situation where you string a couple G5 xServes together to do the model crunching?

[Detrius]

I agree with the G5 Xserve proposal. If you guys are doing the programming yourselves, look into lammpi for your cluster programming (you can download and install it using fink). Apple made the Xserve cluster node specifically for this reason. They are $3k each, but they are dual 2GHz w/dual gigabit ethernet.

I looked up the very basics of finite element analysis to get a wee bit more information. It sounds to me like it is an embarassingly parallel algorithm (i.e. the results of one element do not depend on the results of another element). If the algorithms you use are easily parallelizeable, then you can write the program in such a way that you only have to write the program once, but the more machines you throw at it, the faster you will get your results.

However, technically, if you have a bunch of PCs, you could install some form of MPI (or PVM--your choice) on each of the machines and do the above just as easily with the PCs. Personally, I would much rather do this with lammpi on a Unix box (OS X or Linux, etc..).

This is my personal opinion as a Math Major, Computer Science Major, and Apple Certified Technician (note that I actually did cluster programming with the CS degree).

[Scotttheking]

You didn't provide anywhere near enough information to answer the question.
First, there have been many UNIX boxes made in the last 35 years, so you might want to be more specific.
Second, an idea of how big the models are would be nice. (size and time it takes to run, including what specs the machine is)
Third, is the software custom, or is it a commercial package?
Fourth, what kind of load does it put on the CPU? Things like memory usage and bandwidth, FPU, INT, etc.
Fifth, how well does it work over a cluster?

Finally, fix your grammar and spelling. You said you are an engineer, which means you probably have to ask a superior to make a computer change like this. When making a proposal, it always helps to present it well.

[rogerkylin]

We also use finite element codes for electro-magnetic predictions. They do fairly well on the g5 because they use the BLAS library. If your codes use linear algebra packages, then the g5 can surpass even higher clock-speed Xeons... but not always. Depending on the size (memory) of the jobs, our 2 GHz G5 runs from 50% slower, to 40% faster than a 3 GHz Xeon.

There are two primary reasons the G5 is faster. First, altivec. Second, memory bandwidth. We can saturate the memory bus on both the dual G5 and the dual Xeon, but the Xeon hits its ceiling a lot sooner.

[-Q-]

Originally posted by Scotttheking:
Finally, fix your grammar and spelling. You said you are an engineer, which means you probably have to ask a superior to make a computer change like this. When making a proposal, it always helps to present it well.

Well that was rude, unnecessary, and completely unhelpful. The way one writes on a forum posting a question can be as informal as they want. You may not have intended that comment to come across that way but it did. You, of all people, should know about the ambiguities of text.

[jorgem4]

Thank you -Q-,

....and by the way english is my second language...I wonder how your spelling and grammar are on spanish, french and italian? he?

[alien]

The G5s themselves probably would do well with the FEM work, since the PPC 970 has a lot of (potential) floating point performance.

Do you know what kind of programs and code you are using to solve these models? If it's your corporation's C/C++/Fortran code, you could buy IBM's xlc/xlf compilers (they are very good) and recompile on a G5. Mac OS X has Unix underpinnings, if that is necessary for the code. And if you need MPI, it's available too.

If it's third-party apps things get more trickier, since the software makers would have to make an OS X version and take care to optimize for the Mac, not just do a simple port.

I would suggest subscribing to Apple's sci-tech mailing list and ask your questions there, since the people on that list are scientists and engineers. You'll find it at http://lists.apple.com/

[kentuckyfried]

We actually run all our FEA on PC boxes instead of sending them out to UNIX units (well, there is this one code we use that uses Unix but that's going the way of the dinosaur eventually).

What kind of software are you using? It would be nice to convince SDRC/Unigraphics to compile a Mac native version of the I-DEAS solid modeling and FEA packages. =D Fat chance on that happening anytime soon though. *sigh*

That would be sweet to get my company to switch from PC back to G5's.
But there's no way in hell they'ed ever do it unless the performance and price point was far superior to that of using the PC machine.

Originally posted by jorgem4:
Ok, so I am a mechanical engineer that works for a large corporation, what I do is called Finite Element Analysis. This process requires of solving Finite Element models. The issue is the following:

It takes a lot of time to solve some of the models that we create at work, we use "crappy PCs" (they are not crappy, but to me anything NOT Mac is crappy) to solve de models...also some times if the model is to big we send it to a UNIX machine to solve it, but still takes for ever!

I was thinking if it is possible to solve this models in a G5 (which I think would be much faster) than the UNIX machines we have here. See we can only load our machines with 1 GB of RAM...and the G5 can take more ran and a faster processor.

Do you think it is possible to set a G5 to run these solves and do the crunching much faster?

Thanks.

[jorgem4]

We use I-DEAS...it would be great if it could run on a G5 and Mac OS X....

[kentuckyfried]

^ heheh....that would be ideal...who knows when it'll ever happen, if it happens.

So, is this large mechanical corporation in the South Bay? That would be amusing if we were on the same floor.