Multi-mesh
- This topic has 7 replies, 3 voices, and was last updated 2 years, 10 months ago by .
-
Posts
-
Hello everyone,
I wonder if openlb can implement multi-meshing tech now.For example,in the boundary layer area we choose a larger resolution comparing with others.If can,how?
Also,I find there are two different resolutions in bstep2d case,but are they both implemented in the case?
Cause I only observe that N=60″ resolution is used.Best Regards,
MikeNo, OpenLB currently doesn’t support local grid refinement. Various approaches exist as prototypes but it is not clear when this will be at a stage where we include it in a release as most efforts are currently focused on improving general performance and supporting GPUs.
Thus the bstep2d case only uses a fixed resolution for the whole lattice. The other
Mresolution is an remainder from development, thanks for pointing out this possible source of confusion!Dear Adrian,
In some example cases, there is a command named “SuperLatticeRefinementMetricKnudsen3D”, isn’t it a method to do the grid refinement?
Best Regards,
mengqiangThis is indeed related to grid refinement.
SuperLatticeRefinementMetricKnudsen3Dimplements a refinement criterion [1] to help determine where to refine to which degree. This is one part of a prototype implementation [2] that I pulled into the release as it is also useful in other contexts. e.g. to check whether a simulation gets close to diverging or to identify problematic regions.[1]: Automatic grid refinement criterion for lattice Boltzmann method [Lagrava et al.]
[2]: Prototype of grid refinement in OpenLB 1.3 written for my Bachelor thesisDear Adrian,
I wonder whether there is an available version of OpenLB to simulation on GPUs? I have tried the paralle simulation on CPUs, but the increase of efficiency is not significant (450s by serial and 350s by paralle of 3 cores). Although that may be because my simulation case is not very complex and the number of cores of my CPU is low. I want to achieve a real-time simulation by OpenLB, just like this paper: http://dx.doi.org/10.1016/j.buildenv.2017.08.048. As you can see, the efficiency of their model is very high. Is that possible to achieve that by OpenLB? Thanks.
Best Regards,
mengqiangNo, currently there is no GPU support in OpenLB – however this is at the moment the main focus of my work. The next release will also include support for vectorization on CPUs which also delivers significant speedups.
Which parallel mode did you use? OpenMPI or OpenMP? For OpenMPI our efficiency is around ~80% so you should be able to get your desired timesteps per second by parallel execution on CPUs. Of course if you want to do this on a single host then there is no real way around GPUs – which is one of the reasons why I am focusing on it. E.g. a GPU LBM code of my own in some cases easily outperforms simulations on ~12 nodes of a small cluster using just a single GPU. This is also driven by the desire to fully utilize our new supercomputer at KIT.
Your work is so great! Wish your job sucessful.
I have chosen the OpenMPI mode, but the efficiency is only increased around 25%. I run OpenLB in cygwin on Win10, I am not sure wheter this has a effect.
Besides, the paralle achievenment in cygwin on Windows system can not realized according to the technical report “https://www.openlb.net/wp-content/uploads/2020/11/olb-tr4.pdf”, I think that’s because the modification of the latest openmpi package in cygwin, and I had done some modifications, otherwise there will be an error ouputed in the console.
I hope you can further guide me to achieve the efficiency increase, thank you.Furthermore, could you kindly tell me when the next version of OpenLb would be released? Very expect! 🙂
Best Regards,
mengqiang.
- You must be logged in to reply to this topic.