[sippanspojk]

Hi foamers,

I have developed a functionObject that reads and works on the face pressure values from some selected patches. At the moment I am working with lists and lists of lists to store my values and to perform my calculations. I have however experienced performance issues (lack of RAM) when my mesh grows. It is working fine on ~100k cells but above that it crashes. I have therefore started to look into the matrix class in OpenFOAM. However, before I start to adapt my code I would like to hear your opinion on this - will my code be more lightweight using the matrix class instead of my current lists of lists definition?

Thanks,
David

EDIT ----------------------------------------------------

To further specify my problem - I need to store 5-8 list of lists throughout the complete simulation that all have the size nCells x nCells. This is eating up the RAM when the number of cells increase and I therefore need to find a more economic way of storing this data.

Thanks again,
David

[HPE]

Hi,

Can't you just write them out (if you really really need temporal-accurate data instead of statistics) instead of keeping them in the memory throughout the simulation? Using List or Matrix would not help if the data keep accumulating as ~ t N^2 (t = time, N = cells).

[hjasak]

You cannot afford to store the list of nCells x nCells. This is a massive size.

Please think what you need this list for. Are there alternative stogare arrangements that would work? Do you REALLY need a list of that size?

fvMatrix is stored in a sparse matrix format, excluding all the entries that you know will be zero. The lookup is slower, but it is feasible. Even on a massive supercomputer, there is not enough memory for nCells x nCells.

Also, please consider carefully what you are trying to do. When running in parallel, does it parallelise, ie do you have one huge global matrix or a smaller matrix on each processor? Read up on parallel communication and sparse matrix formats - otherwise you will always be stuck around 100k cells. By the way, 100k x 100k is 10 BILLION NUMBERS: this is why you run out of memory

[superkelle]

Originally Posted by sippanspojk

Hi foamers,

I have developed a functionObject that reads and works on the face pressure values from some selected patches.

So only for understanding, you are only storing some hundred to some thousand values of (boundary) patches for meshes of a bigger size?

[sippanspojk]

Originally Posted by hjasak

fvMatrix is stored in a sparse matrix format, excluding all the entries that you know will be zero. The lookup is slower, but it is feasible. Even on a massive supercomputer, there is not enough memory for nCells x nCells.

Also, please consider carefully what you are trying to do. When running in parallel, does it parallelise, ie do you have one huge global matrix or a smaller matrix on each processor? Read up on parallel communication and sparse matrix formats - otherwise you will always be stuck around 100k cells. By the way, 100k x 100k is 10 BILLION NUMBERS: this is why you run out of memory

Is fvMatrix taking care of excluding the zero values on its own or do I need to provide more information more than just defining my matrix using the class?

When I run it in parallel at the moment I gather the pressure values onto the master node, perform my matrix multiplications and then scatter it back. Which means that it is the memory on the master node that is doing the job.

Cheers,
David

[sippanspojk]

Originally Posted by superkelle

So only for understanding, you are only storing some hundred to some thousand values of (boundary) patches for meshes of a bigger size?

I select the patches I am interested in (e.g. building, but not inlet, oulet, ground, etc.) and perform my computations on the faces on those patches. And I want to be able to do that on ~1M faces.

Cheers,
David