#594 Make potential processor rank with which communication can occur available through grid interface
|Reported by||Markus Blatt (firstname.lastname@example.org)|
|Reported at||Sep 2, 2009 19:12|
|Version||Git (pre2.4) [autotools]|
|Operating System||Unspecified / All|
We have the fancy communication interface available in parallel Dune grids. This is all fine and sufficient as long as it is possible to use it for communications. Unfortunately there are scenarios where this is not possible. E.g. when using third party parallel software.
while it is possible to compute all this information based on the global indices (indepedent of the grid interface), doing this usually results in e.g. communicating this information in a ring. Combined with sorting and comparing remote information with local information to setup separate communication structures usually scales very poor for big (that is >2000 processors).
The situation would improve drastically if each processor could get the maximum set of MPI ranks with which communication through the grid interface could occur. Namely instead of ring communication with O(P) complexity we could use asynchronous communication with just the potential neighbours, complexity O(noNeighbours).
I suppose this information is already there in the implementations anyway (Grid maintainers please comment). Evem if not it could be computed.
I therefore would like to have this information available through the grid interface.