On 8/24/20 4:57 PM, Matthew Knepley wrote:
On Mon, Aug 24, 2020 at 4:27 PM Jed Brown <[email protected]
<mailto:[email protected]>> wrote:
Cameron Smith <[email protected] <mailto:[email protected]>> writes:
> We made some progress with star forest creation but still have
work to do.
>
> We revisited DMPlexCreateFromCellListParallelPetsc(...) and got it
> working by sequentially partitioning the vertex coordinates across
> processes to satisfy the 'vertexCoords' argument. Specifically,
rank 0
> has the coordinates for vertices with global id 0:N/P-1, rank 1 has
> N/P:2*(N/P)-1, and so on (N is the total number of global
vertices and P
> is the number of processes).
>
> The consequences of the sequential partition of vertex
coordinates in
> subsequent solver operations is not clear. Does it make process i
> responsible for computations and communications associated with
global
> vertices i*(N/P):(i+1)*(N/P)-1 ? We assumed it does and wanted
to confirm.
Yeah, in the sense that the corners would be owned by the rank you
place them on.
But many methods, especially high-order, perform assembly via
non-overlapping partition of elements, in which case the
"computations" happen where the elements are (with any required
vertex data for the closure of those elements being sent to the rank
handling the element).
Note that a typical pattern would be to create a parallel DMPlex
with a naive distribution, then repartition/distribute it.
As Jed says, CreateParallel() just makes the most naive partition of
vertices because we have no other information. Once
the mesh is made, you call DMPlexDistribute() again to reduce the edge cut.
Thanks,
Matt
Thank you.
This is being used for PIC code with low order 2d elements whose mesh is
partitioned to minimize communications during particle operations. This
partition will not be ideal for the field solve using petsc so we're
exploring alternatives that will require minimal data movement between
the two partitions. Towards that, we'll keep pursuing the SF creation.
-Cameron
