I would like to know, which benchmark is reliable on the performance of LARGE Message Allreduce?
As I developed a OpenMPI MPI_Allreduce collective for large message, I want test my own Allreduce on Lammps. But i don’t kown which benchmark is reliable on LARGE Message Allreduce.
OR, How can I configure these benchmarks to make them use Large Message Allreduce as much as possible?
You are asking in the wrong forum. You will be better off to discuss this with the OpenMPI developers.
There are specific MPI benchmarks. A search engine will help you find them.
I don’t think that this is a feature with much value for LAMMPS. Since LAMMPS does domain decomposition the most performance critical communications are point-to-point sends between neighboring sub-domains. And if you use the 3d FFTs with kspace style pppm the dominant calls are Alltoall calls.