search

cwpearson / tempi

Topology Experiments for MPI
Other
10 stars 4 forks source link

bench-halo-exchage needs some MPI_Barriers #4

Closed cwpearson closed 3 years ago

cwpearson commented 3 years ago

https://github.com/cwpearson/tempi/blob/72811c815de085bbd8db566f0ad95fb15d3b65ab/bin/bench_halo_exchange.cpp#L583-L592

Without barrier, if any thread arrives here early it can spend a comparatively long time here and make the time in exchange show up as quite long. In the baseline implementation the pack/unpack can be very slow, so minor variability in that time can show up as a very long exchange iteration.

cwpearson commented 3 years ago

Fixed in e232cb3c2f8c9ad79adc7fe9380730cdfdbba8e5