Q-Logic IB6054601-00 D Manuel D’Utilisation

Page de 122
A – Benchmark Programs
Benchmark 4: Measuring MPI Latency in Host Rings
IB6054601-00 D
A-5
Q
A.4
Benchmark 4: Measuring MPI Latency in Host Rings
The program 
mpi_latency
 can be used to measure latency in a ring of hosts. Its 
syntax is a bit different from Benchmark 1 in that it takes command line arguments 
that let you specify the message size and the number of messages over which to 
average the results. So, for example, if you have a hosts file listing four or more 
nodes, the command: 
mpirun -np 4 -ppn 1 -m mpihosts mpi_latency 100 0
might produce output like this: 
0       1.760125 
This indicates that it took an average of 1.76 microseconds per hop to send a 
zero-length message from the first host, to the second, to the third, to the fourth, 
and then get replies back in the other direction.