Dear Gergely, are you using the same mesh for your serial run as for your parallel run? The mesh shouldn’t require much more memory when it’s distributed. If your serial code runs fine, the same code should also run fine in parallel (at least the basic part. If you don’t have some particular user data you might need to manually distribute that). As long as your macro grid fits into memory it should also be fine to distribute that initially, then refine (each process will refine) and then redistribute to optimise load balance.
Best wishes, Timo > On 13. Jan 2022, at 13:46, Gergely Schmidt > <schm...@hydromech.uni-hannover.de> wrote: > > Dear Timo, > > thank you very much for the detailed explanation! My mesh doesn't seem to fit > into memory, as the simulation freezes before decomposition. I'll stay serial > for a while longer for now, but change in the future to ALUGrid’s internal > mesh format or DGF. > > Kind regards, > Gergely > > Am 12.01.2022 um 17:47 schrieb Timo Koch: >> Dear Gergely, >> >> if the mesh fits into memory, you don’t need to do anything special for >> parallel runs since the default grid manager will call grid.loadBalance() >> which distributes the grid. >> The grid data helper also automatically distributes user data like element >> markers, in case you are using that feature. >> The gmsh grid is read on process 0 and is then distributed to the other >> processors using ALUGrid's default partition strategy. >> The default partition is usually very good for ALU. There might be also >> different strategies you can select. >> There is also the possibility to manually partition the grid using an >> external partitioner (more involved). >> >> If the mesh doesn’t fit into memory, and therefore you can’t read it on >> process 0, there is currently no out-of-the-box helper in Dumux. >> You would have to use ALUGrid’s internal mesh format which allows grid to be >> split into separate pieces. >> How to convert from a distributed gmsh grid to the ALU format you would have >> to ask the AluGrid developers. >> >> Best wishes >> Timo >> >>> On 12. Jan 2022, at 15:51, Gergely Schmidt >>> <schm...@hydromech.uni-hannover.de> wrote: >>> >>> Dear Dumux community, >>> >>> I'd like to ask a question about mesh decomposition. I have a gmsh mesh >>> (.msh) for a simple conduction model that runs without problems serially >>> using ALUGrid. Is it possible to decompose that mesh for parallel runs or >>> is there a workaround with a file conversion maybe or using YASP, UG etc.? >>> I'd like to avoid switching to DGF though if possible. >>> >>> Many thanks! >>> >>> Best, >>> Gergely >>> >>> -- >>> *********************************************************** >>> Gergely Schmidt M.Sc. >>> Wissenschaftl. Mitarbeiter >>> Institut für Strömungsmechanik und Umweltphysik im Bauwesen >>> Leibniz Universität Hannover >>> Appelstr. 9 a >>> 30167 Hannover >>> tel +49 (0) 511 / 762 3709 >>> *********************************************************** >>> >>> _______________________________________________ >>> DuMux mailing list >>> DuMux@listserv.uni-stuttgart.de >>> https://listserv.uni-stuttgart.de/mailman/listinfo/dumux > > -- > *********************************************************** > Gergely Schmidt M.Sc. > Wissenschaftl. Mitarbeiter > Institut für Strömungsmechanik und Umweltphysik im Bauwesen > Leibniz Universität Hannover > Appelstr. 9 a > 30167 Hannover > tel +49 (0) 511 / 762 3709 > *********************************************************** > _______________________________________________ DuMux mailing list DuMux@listserv.uni-stuttgart.de https://listserv.uni-stuttgart.de/mailman/listinfo/dumux