在调用Hyper MPI运行应用时,会同时调用MPI_Allreduce、MPI_Bcast、MPI_Barrier和MPI_Alltoallv操作,此时的推荐算法命令示例如下(鲲鹏芯片):
mpirun -np 16 -N 2 --hostfile hf --mca btl ^vader,tcp,openib --map-by socket --rank-by core --bind-to core -x UCX_TLS=sm,rc_x -x UCX_BUILTIN_BCAST_ALGORITHM=3 -x UCX_BUILTIN_ALLREDUCE_ALGORITHM=11 -x UCX_BUILTIN_BARRIER_ALGORITHM=10 -x UCX_BUILTIN_ALLTOALLV_ALGORITHM=2 test_case
mpirun -np 16 -N 2 --hostfile hf --mca btl ^vader,tcp,openib --map-by socket --rank-by core --bind-to core -x UCX_TLS=sm,rc_x -x UCX_BUILTIN_BCAST_ALGORITHM=3 -x UCX_BUILTIN_ALLREDUCE_ALGORITHM=6 -x UCX_BUILTIN_BARRIER_ALGORITHM=5 test_case