We've looked at this several times and it's very difficult to estimate
memory growth for complex C++ software. Each data structure that's spread
throughout the code uses their own memory allocator to store information as
needed. Even if you know things like the sizes of individual nodes in a
node based container or the sizes of items and the number of items in a
vector, it's still difficult to capture the amount of memory in use
everywhere as the code changes and new data structures are added or
removed.

As you run in parallel the amount of overhead can change depending on a lot
of factors. Changing out your preconditioner or changing solver strategy
can drastically change your memory usage. So like you, we monitor things
empirically to figure out the amount of memory to reserve and it's a
difficult process. We (MOOSE) have tools that can monitor your memory usage
for a given simulation so you can see trends over time to aid in figuring
this out, but we've never found a bulletproof solution to this problem.

Cody

On Thu, Mar 23, 2017 at 9:39 AM Fabio Canesin <fabio.cane...@gmail.com>
wrote:

> Dear experts,
>
> Is there a api for computing the memory requirements of a FEMSystem or
> EquationSystem ?
> In running a navier-stokes 1MM mesh I constantly hit memory errors, even
> with ParallelMesh.
>
> My intent is to compute how much memory I need in each node for each number
> of mpi tasks before submitting the job (for now, I'm using a trial-error
> empirical relation).
>
> Regards,
> Fábio
>
> ------------------------------------------------------------------------------
> Check out the vibrant tech community on one of the world's most
> engaging tech sites, Slashdot.org! http://sdm.link/slashdot
> _______________________________________________
> Libmesh-users mailing list
> Libmesh-users@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/libmesh-users
>
------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
_______________________________________________
Libmesh-users mailing list
Libmesh-users@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/libmesh-users

Reply via email to