Dear Elmer team,
I am currently trying to get a model running in parallel that appears to be working in serial runs but immediately crashes in parallel. The case in question is a 3D parallel sliding contact with periodic BCs applied in sliding direction (X) and Mortar BCs for the two contacting surfaces (in Z) of the two solids. While the periodicity is correctly working in parallel (via "-periodic 1 0 0" during partitioning), the Mortar Contact is ignored. As far as I understood e.g. from viewtopic.php?p=9835#p9835, the Mortar BCs must be connected via "-connect bc1 bc2". Unfortunately, as ElmerGrid complains, the "-connect" flag can obviously not be used in periodic systems.
I have also compiled with Zoltan, but have also not been able to get things running.
Do you have an idea what could do the trick, or is it impossible to run this configuration in parallel?
I would greatly appreciate your help!
Kind regards
Bengt
Periodic BCs and contact in parallel
-
- Posts: 2237
- Joined: 25 Jan 2019, 01:28
- Antispam: Yes
Re: Periodic BCs and contact in parallel
I had a similar problemed and learned that the contact surfaces have to be in the same partition when doing parallel, otherwise they are ignored.
-
- Site Admin
- Posts: 4812
- Joined: 22 Aug 2009, 11:57
- Antispam: Yes
- Location: Espoo, Finland
- Contact:
Re: Periodic BCs and contact in parallel
Hi Bengt
The -connect flag should allow contact mechanics to work in parallel. I think you just need to use dual graph for it, e.g.
What this will do is to enforce bc1 and bc2 to be in same partitions. This might not be ideal for load balancing. For that there is some specialized halo strategies that will duplicate parts of the mesh.
The limitation here is that the mortar integrals are rather tedious and have not been parallelized. So to overcome the problem of communication the partitions are assumed to be setup such that information is available locally.
-Peter
The -connect flag should allow contact mechanics to work in parallel. I think you just need to use dual graph for it, e.g.
Code: Select all
ElmerGrid 2 2 meshdir -partdual -metiskway 8 -connect bc1 bc2
The limitation here is that the mortar integrals are rather tedious and have not been parallelized. So to overcome the problem of communication the partitions are assumed to be setup such that information is available locally.
-Peter
Re: Periodic BCs and contact in parallel
Hi Peter,
thanks a lot for your reply. I have tried, and ElmerSolver complains about the periodicity (imposed on BCs 1 and 2 in x direction).
When using, ElmerGrid complains about
Is there another trick that I could try?
Kind regards
Bengt
thanks a lot for your reply. I have tried
Code: Select all
ElmerGrid 2 2 combinedMesh -partdual -metiskway 10 -connect 5 12
When using
Code: Select all
ElmerGrid 2 2 combinedMesh -partdual -metiskway 10 -connect 5 12 -periodic 1 0 0
.Dual graph not implemented for periodic system!
Is there another trick that I could try?
Kind regards
Bengt
-
- Site Admin
- Posts: 4812
- Joined: 22 Aug 2009, 11:57
- Antispam: Yes
- Location: Espoo, Finland
- Contact:
Re: Periodic BCs and contact in parallel
Hi
Try
If they share nodes you cannot separate the bc sets and have to combine them all
You can look how your partitioning looks like with ElmerGrid 2 5 -> paraview.
You could try having additional flags to divide this large set while including information with halos, e.g. adding "-partzbc 3 -haloz -partlayers 2", for example. This has been used for electrical machines, not contact mechanics but the logic should be the same...
-Peter
Try
Code: Select all
ElmerGrid 2 2 combinedMesh -partdual -metiskway 10 -connect 1 2 -connect 5 12
Code: Select all
ElmerGrid 2 2 combinedMesh -partdual -metiskway 10 -connect 1 2 5 12
You could try having additional flags to divide this large set while including information with halos, e.g. adding "-partzbc 3 -haloz -partlayers 2", for example. This has been used for electrical machines, not contact mechanics but the logic should be the same...
-Peter
Re: Periodic BCs and contact in parallel
Hi Peter,
thanks so much for your helpful suggestions! Unfortunately, connecting all those boundaries and imposing the conforming BCs in sliding direction still appears to be impossible. I have played around with all those additionally available -halo etc. flags, but still haven't had any success.
Maybe I just haven't found the correct combination of partitioning methods and those additional flags for halo elements...
Do you have any further ideas which command syntax could also be used in ElmerGrid? Are special modifications to the sif necessary in order to help ElmerSolver with the conforming BCs?
Kind regards
Bengt
thanks so much for your helpful suggestions! Unfortunately, connecting all those boundaries and imposing the conforming BCs in sliding direction still appears to be impossible. I have played around with all those additionally available -halo etc. flags, but still haven't had any success.
Maybe I just haven't found the correct combination of partitioning methods and those additional flags for halo elements...
Do you have any further ideas which command syntax could also be used in ElmerGrid? Are special modifications to the sif necessary in order to help ElmerSolver with the conforming BCs?
Kind regards
Bengt