There seem to be quite large overheads connected with oomph memory allocation. Ideally, performance should be the same in those two scenarios:
- halo exchange with static, pre-allocated messages
- halo exchange with messages allocated each time and moved into the oomph send/recv routines
This seems not to be the case now: ghexbench has much lower performance in the second case. We should debug and improve.
There seem to be quite large overheads connected with oomph memory allocation. Ideally, performance should be the same in those two scenarios:
This seems not to be the case now: ghexbench has much lower performance in the second case. We should debug and improve.