loadbalancing dreadfully slow
tommy-engels opened this issue · comments
the loadbalancing routine is dreadfully slow. if the code runs with maxlevel dealiasing, i.e. blocks on jmax are forced to coarsen, then loadbalancing is done only after the grid adaptation step. Hence, in such a case, most of the time the routine does not do much or nothing.
if the maxlevel dealiasing is off, the loadbalancing is also executed in refinement. now, many blocks are shuffeled around, and loadbalancing becomes by far the most expensive part of the code (~70%).
we should hence improve loadbalancing.
with the latest commit, this issue has been improved, but there is still room to do better
nonblocking_vs_blocking.pdf