Simple MPI Communications Time and Variability Measurements
mmpi.c is the top level program
Measure MPI
This does an all to all tx and rx and times the message delivery time.
The output files are post processed into plots showing core to core travel time
author: Philip James Maechling pjmaechling@yahoo.com
== Example invocation == Example Call: %qsub -A Equake_SS -n 2 -t 10 --mode c16 mmpi # Measures 32 cores on 2 nodes %qsub -A Equake_SS -n 32 -t 10 —mode c1 mmpi # Measures 32 cores on 32 nodes
== Measuring a new machine == To measure a new machine, the most common changes required are: New makefile that works on system under test Job submission script with appropriate parameters such as cores per node The mmpi2mat.py script will often need to be updated to properly scan the measurement output file.
== File Formats ==
- Measurement file format
- matlab file format
- matlab plot formats
== Program Limits ==
Please note the program limits defined in the mmpi.h amd mmsg.h configuration files. These can be changed but then the program must be recompiled
Note MAX_Processors defined in mmsg.h Max processors defines to be 100 Max comm paths 5050