maintenance is the biggest part, starting is easy.

>10.000 lines and it becomes a headache

@finn @Yann quick question, which compressible , single phase solver is the most impactful in the industry? I want to start porting into brae...

Hmmm I'm not sure if it's the most used in industry but I would go for rhoSimple/rhoPimpleFoam

deal

I feel like people are not very happy with sonicFoam/rhoCentralFoam when it comes to transonic/supersonic flows and they tend to look for alternative solvers such as HiSA

I might be wrong though, I don't work much on this kind of applications

okay, these informations are valuable. It avoid me to waste time doing something people don't want or won't used.

But right now, the current work/challenge I have is domain decomposition on 2-8x GPU in a single node.

this one right? https://hisa.gitlab.io/

yep! Keep in mind, I don't work with supersonic flows so it's more of a feeling from the few things I've seen rather than actual experience.

gotcha

are you more into crogenics ?

I do mostly good old incompressible external aerodynamics and a bit of heat transfer (buoyantSimpleFoam and chtMultiRegionFoam)

interesting, i was using recently chtMultiRegionFoam for tube in tube heat exchanger

brae will support more incompressible in few weeks.. if you have GPU, you could try it.

I recently resolved a tube in a tube heat exchanger, by doing strongly coupled set of PF's as well as I had condensation with NRTL equation that was linked. Just some easy python A-matrix setup and DWSIM for thermodynamics

no fancy fluid dynamics needed, or you want to comeup with fancy setups where no one ever have fitted some nusselt correlations for.

Oh i see. so you treated it more like a coupled 1D/process model with thermodynamics from DWSIM, instead of resolving the full 3D flow field or the 2D... mine was 2D only...

There is "real time" and "cpu user time" and "cpu system time" in the solver output at the end. Suppose I run simulation on N cpu cores. Does it mean I should expect Real time = N*(cpu system time+cpu user time) Is this correct or wrong?

cpu user + system time = cpu work accumulated by processes.

user + sys ≈ N × real

because of mpi com, load imbalance, i/o, waiting, and ranks sharing/competing for cores

When i count like that i get N that is much, much less than the number of cpu cores i was using

Does it mean i requested too much cpu cores on that thing? Is there example of a good paper with included scaling study that is properly detailed.

I have found that depending on system, it can be good to leave one or two cores to do OS stuff

If you have a 16 core CPU for exnaple, I would run on only 14 cores, when I have benchmarked it I've found that 15 cores is at best the same speed as 14, and running on 16 can be much slower than on 14

Is this logic applicable for linux operating systems too. I run loads on compute with a multi core 💯 load. I thought I was doing correct? OS is Ubuntu 26.04

Yes, though it's not quite as bad as windows

I'd recommend just trying one of your cases on a few different core counts

Mine is more like "I am running on N cores, but if I do the math based on the real time and cpu times output, it is like it used N/10 (or less) cores". Ah well. I shall attempt a few variations. For example, N/16...