GOMP_BARRIER_TYPE – Choose barrier algorithm ¶Choose barrier algorithm.
This variable should be set to one of flat, tree,
centralized, or auto. This choice is only effective on targets
which have the futex_waitv syscall. As a rule of thumb, flat and
tree are faster when GOMP_SPINCOUNT is high enough that no thread
waits passively for some variable to change while centralized is faster
when some threads do wait passively. Accordingly auto chooses between
flat and centralized based on the spin count. However the
microarchitecture of the base CPU and the work a user’s program is running both
matter so this is not a hard rule.
If this is set and GOMP_FORKJOIN_BARRIER_TYPE is not set, then
GOMP_FORKJOIN_BARRIER_TYPE takes its value from this.
OMP_WAIT_POLICY – How waiting threads are handled, GOMP_SPINCOUNT – Set the busy-wait spin count, GOMP_FORKJOIN_BARRIER_TYPE – Choose parallel region implicit barrier algorithm