The OSU Micro-Benchmarks (OMB) are used like the Intel MPI Benchmarks (IMB) the measure the achievable latency, bandwidth, … of MPI libraries and interconnects. In contrast to IMB the OSU micro-benchmarks exhibit for several benchmark types a different communication pattern. In the following the patterns used by the latency, bandwidth, and bi-directional bandwidth point-to-point benchmarks is described.

Latency (osu_latency)

In this benchmark latency denotes the time it takes to transfer a message of a certain size from one MPI rank to another. For sending and receiving separate buffers are used, but stay the same during each iteration.

Bandwidth (osu_bw)

Measures the uni-directional bandwidth from one to another MPI rank. Hereby several MPI_Isends are started followed by a MPI_Waitall. There receiving side uses matching MPI_Irecvs with MPI_Waitall. The number of started MPI_Isends is defined as the window_size. The send and receive buffer for each MPI_Isend/MPI_Irecv is the same buffer. One iteration ends when the sending sides gets the receive of all messages acknowledged.

Bi-directional Bandwidth (osu_bibw)

This benchmark works as the uni-directional bandwidth benchmark only that both sides issue first MPI_IRecvs followed by MPI_Isends and MPI_Waitalls. Again send/receive buffers are the same, respectively.