Q-Logic IB6054601-00 D manual Benchmark 2 Measuring MPI Bandwidth Between Two Nodes

A – Benchmark Programs

Benchmark 2: Measuring MPI Bandwidth Between Two Nodes

Q

This benchmark always involves just two node programs. You can run it with the command:

$ mpirun -np 2 -ppn 1 -m mpihosts osu_latency

The -ppn 1 option is needed to be certain that the two communicating processes are on different nodes. Otherwise, in the case of multiprocessor nodes, mpirun might assign the two processes to the same node, and so the result would not be indicative of the latency of the InfiniPath fabric, but rather of the shared memory transport mechanism. Here is what the output of the program looks like:

# OSU MPI Latency Test (Version 2.0)

# Size	Latency (us)
0	1.26
1	1.26
2	1.26
4	1.26
8	1.26
16	1.45
32	1.47
64	1.52
128	1.63
256	1.88
512	2.34
1024	3.25
2048	5.13
4096	7.34
8192	11.58
16384	20.25
32768	37.56
65536	78.69
131072	149.84
262144	287.49
524288	565.84
1048576	1119.18
2097152	2220.18
4194304	4424.59

The first column gives the message size in bytes, the second gives the average (one-way) latency in microseconds. Again, this example is given to show the syntax of the command and the format of the output, and is not meant to represent actual values that might be obtained on any particular InfiniPath installation.

A.2

Benchmark 2: Measuring MPI Bandwidth Between Two Nodes

The osu_bw benchmark is meant to measure the maximum rate at which you can pump data between two nodes. It also uses a ping-pong mechanism, similar to the osu_latency code, except in this case, the originator of the messages pumps a number of them (64 in the installed version) in succession using the non-blocking

A-2 IB6054601-00 D