648 | | Depending on the communication network in use, setting of barriers (MPI_BARRIER) in front of collective MPI operations (MPI_ALLTOALL, MPI_ALLREDUCE) may speedup the code, e.g. by a few percent on SGI-ICE-systems with Infiniband fat tree. Therefore, {{{collective_wait}}} is {{{.TRUE.}}} by default on SGI-systems (if hostname(3:5) = "{{{sgi}}}", given with '''mrun'''-option {{{-h}}}), but {{{.FALSE.}}} on all other systems. |
| 648 | Depending on the communication network in use, setting of barriers (MPI_BARRIER) in front of collective MPI operations (MPI_ALLTOALL, MPI_ALLREDUCE) may speedup the code, e.g. by a few percent on SGI-ICE-systems with Infiniband fat tree, but only for a larger number of PEs (>= 2048) and if dual rail mode is switched on (default on SGI-systems at HLRN). Therefore, {{{collective_wait}}} is {{{.TRUE.}}} by default on SGI-systems (if hostname(3:5) = "{{{sgi}}}", given with '''mrun'''-option {{{-h}}}), but {{{.FALSE.}}} on all other systems. |