On Mon, 18 Aug 2014, John Peterson wrote:
> On Mon, Aug 18, 2014 at 8:58 AM, John Peterson <jwpeter...@gmail.com> wrote: >> On Mon, Aug 18, 2014 at 8:33 AM, Derek Gaston <fried...@gmail.com> wrote: >>> You're right - it doesn't... I misread my old code... I missed my own >>> packing/unpacking of the coordinates from/to Points (read them out into a >>> flat array, broadcast(), recreate Points). >>> >>> So - never mind. I guess I'll do the same thing again for now. >>> >>> How tough would it be to make this work with Point? >> >> Looks like it involves template magic and defining custom MPI types? Pretty much. It involves template magic and custom MPI types that should be 95% there already, though; I'll take a crack at adding whatever's missing this afternoon. >>> Or should I be using allgather_packed_range() for this instead (would that >>> work as is?). >> >> That looks pretty close to what you want, actually... It looks like >> it takes a generic "context" and uses an std::inserter object for >> unpacking. >> >> Here's an example from mesh_communication.C on Node*s... >> >> mesh.comm().allgather_packed_range (&mesh, >> mesh.nodes_begin(), >> mesh.nodes_end(), >> mesh_inserter_iterator<Node>(mesh)) : > > Actually the more I looked into it, it appears that > allgather_packed_range is fairly specific to working with a Mesh > object... so if you just have a vector<Point> you'd like to > allgather(), it doesn't seem to be a viable choice. allgather_packed_range doesn't need to be Mesh specific - that mesh pointer is a generic "context" pointer, necessary for any cases where you're also packing data that isn't directly carried with the objects (in this case boundary info). But all the packed_range code is inefficient for anything that doesn't require packing. With Elem/Node there's data (that boundary info, dof indices, neighbor pointers etc) that's variable size, so serializing is tricky. With Point you get 3*sizeof(Real) bytes regardless, so it should be preferable to use the simpler methods. --- Roy ------------------------------------------------------------------------------ _______________________________________________ Libmesh-devel mailing list Libmesh-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/libmesh-devel
