Memory Requirement

[888]

Hello.

At the place im working we are looking to acquire a new PC to do FEM computations, but we dont know how much memory will be necesary. In another post i read that memory requirements increase linearly with the mesh size, but is there any way i can estimate the necesary memory depending on the number of FE?

We will be using direct methods (UMFPACK).

The idea is to perform 3D computations in relatively large domains compared with the geometries involved, having to use quite large amounts of 3D FE, maybe around 500000 or more.

Maybe someone who works with FEM computations has an idea of a recommended amount of memory. At least to have a reference number.

Thanks!

[raback]

Hi

The iterative methods use much less memory, particularly in 3D umfpack is a memory hog. So you could try some of the Krylov methods for your problem as well. In fact the memory consumption of direct methods may grow superlinearly with the problem size. Why won't you just run some typical case and check the memory consumption while it's running.

-Peter

[888]

Why won't you just run some typical case and check the memory consumption while it's running.

Yes, that is one of the approaches im going to use but i wanted to know if there was some other way to perhaps calculate or estimate the memory consumption.

I didnt expect UMFPACK to be so demanding but if thats the case i would see into the iterative methods.

Thanks

[petroo]

... but i wanted to know if there was some other way to perhaps calculate or estimate the memory consumption.

As memory prices are not so very high at the moment, I'd suggest to buy as much as will fit onto your mainboard, which should amount to 16 or 32 GB presently. On my FEM machine I hit the limit of swapping already with 8 GB for some ill-dimensioned meshes (rather large and rather small geometric details united in one problem). And as soon as swapping starts you may as well stop the calculations since the access time to the swap is about a factor of 1000 larger compared to the RAM access.

Any work you do for getting educated estimations beforehand will cost you presumably much more than an investment in larger RAM. Should the RAM prove too large in the end: Just do several calculations in parallel which should be sensible on any multi-core machine that is set up nowadays.

Regards,

Peter

[888]

Thanks for the response!