On Dec 9, 2010, at 1:49 PM, Roy Stogner wrote: > You may find it easier to do the mesh creation from another program > that can save to Nemesis format and read it into libMesh from that? > IIRC the mesh reading is currently serialized, but it's done in chunks > (which are freed after being distributed off the root node) to > conserve memory, so if memory-per-node is your limitation then that > will probably get around it.
Nemesis should read each chunk on it's respective processor... However.... my suggestion would be to find a machine with a ton of memory (I mean... how big is your mesh... surely not bigger than like 16-32 GB right?.... I bet there is some big memory machine somewhere at purdue with enough memory to hold your whole mesh at once). Then use that big memory machine to build your mesh using libMesh and discard the nonlocal parts (like Roy suggested)... THEN write it out using parallel XDR format... which should write one file per processor. THEN go back to your normal computer / cluster with your now parallelized mesh and read it in parallel XDR mode (which should suck up one file per processor). Note: I have not actually tried this myself... but based on what I've seen in the code I believe it should work. Ben is the one that wrote the parallel XDR support I believe... so maybe he can chime in. Derek ------------------------------------------------------------------------------ _______________________________________________ Libmesh-users mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/libmesh-users
