Intel MPI Benchmarks (IMB) Keyboard Shortcuts

Point to point

mpirun -np 2 IMB-MPI1 PingPong is the classic latency and bandwidth test between two ranks — place them on different nodes to measure the network, on the same node to measure shared memory — and its small-message time is the fabric's latency. mpirun -np 2 IMB-MPI1 PingPing sends both ways at once, and mpirun -np 4 IMB-MPI1 Sendrecv and mpirun -np 4 IMB-MPI1 Exchange run ring patterns that show how bandwidth holds up under bidirectional load.

Collectives

mpirun -np 8 IMB-MPI1 Allreduce is the pattern behind gradient synchronisation and the one to watch for training clusters; mpirun -np 8 IMB-MPI1 Allgather, mpirun -np 8 IMB-MPI1 Alltoall and mpirun -np 8 IMB-MPI1 Bcast cover the rest, and mpirun -np 8 IMB-MPI1 Barrier measures pure synchronisation cost. Each benchmark prints a table per message size with min, max and average time; the gap between min and max at large process counts reveals stragglers.

Control flags

-msglog min:max restricts message sizes to powers of two in that range, so a run can focus on the small-message regime or the bandwidth regime. -include name adds benchmarks to the default list and -npmin n sets the smallest process count to test; by default IMB runs each collective at several process counts up to the launch size, which is how scaling curves are produced in one job. Placement and binding flags on mpirun apply as they would to any job.