MPI Profiler (alias mprof) is a tool to profile MPI functions. It uses an interposition library, named libmprof.so. A wrapper is also provided, named mprof.
This tool profiles only MPI_COMM_WORLD communication (others communication will work but not profiled) and it is NOT currently thread safe.
Therefore, we recommend you to run this tool only with full MPI applications and with only MPI_COMM_WORLD communicator.
Also, we assume that your code following the few MPI rules:
- you don’t call 2 times or more MPI_Init or MPI_Finalize
- MPI_Init and MPI_Finalize should wrap almost all your code
Please ensure that following dependencies are installed in your system:
- gcc
- make
- MPI library (any vesrion)
- coreutils (install)
$ sudo pacman -S --needed gcc make openmpi coreutils$ git clone https://github.com/Sholde/mprof.git
$ cd mprof
$ make
$ sudo make installIt is easy to use. You just need to put mprof word ahead your MPI execution line. For example:
$ mprof mpiexec -np 4 ./a.out
or
$ mprof mpirun -np 4 ./a.outmprof provides different command line options:
-v, --versionprint the current version-h, --helpprint a friendly help message-l, --listprint the list of environment variable options--options=...allow you to put environment variable options directly in the command line. Note that you must remove hyphen--.--cleanclean all mprof files.--summary=PROCESS_IDprint the file that contain the summary of process PROCESS_ID
mprof allows you to specify different options in an environment variable, named MPROF_OPTIONS.
NOTE: the number of file created correspond to the number of MPI process. We advise you to run mprof with a reduce number of process when you use an environment variable option.
--verboseprint in a file named mprof_ProcessId.out, all call of MPI--barrierprint in a file named mprof_ProcessId.out, all call of MPI_Barrier--initprint in stderr all call of MPI_Init with all argument--finalizeprint in stderr all call of MPI_Finalize--profileprint in a file named mprof_ProcessId.profile, all information of profiling, like contiguous sends
Here it is a code where only the process 0 call MPI_Barrier and others finished.
$ MPROF_OPTIONS="--finalize" mprof mpirun -np 4 ./lbm
=================== CONFIG ===================
iterations = 16
width = 800
height = 160
obstacle_r = 17.000000
obstacle_x = 161.000000
obstacle_y = 83.000000
reynolds = 100.000000
reynolds = 100.000000
inflow_max_velocity = 0.100000
output_filename = result.raw
write_interval = 50
------------ Derived parameters --------------
kinetic_viscosity = 0.034000
relax_parameter = 1.661130
==============================================
RANK 0 ( LEFT -1 RIGHT -1 TOP -1 BOTTOM 1 CORNER -1, -1, -1, -1 ) ( POSITION 0 0 ) (WH 802 42 )
RANK 2 ( LEFT -1 RIGHT -1 TOP 1 BOTTOM 3 CORNER -1, -1, -1, -1 ) ( POSITION 0 80 ) (WH 802 42 )
RANK 1 ( LEFT -1 RIGHT -1 TOP 0 BOTTOM 2 CORNER -1, -1, -1, -1 ) ( POSITION 0 40 ) (WH 802 42 )
RANK 3 ( LEFT -1 RIGHT -1 TOP 2 BOTTOM -1 CORNER -1, -1, -1, -1 ) ( POSITION 0 120 ) (WH 802 42 )
Progress [ 1 / 16]
Progress [ 2 / 16]
Progress [ 3 / 16]
Progress [ 4 / 16]
Progress [ 5 / 16]
Progress [ 6 / 16]
Progress [ 7 / 16]
Progress [ 8 / 16]
Progress [ 9 / 16]
Progress [ 10 / 16]
Progress [ 11 / 16]
Progress [ 12 / 16]
Progress [ 13 / 16]
Progress [ 14 / 16]
Progress [ 15 / 16]
==mprof== Process 3 enter in MPI_Finalize
==mprof== Process 1 enter in MPI_Finalize
==mprof== Process 2 enter in MPI_Finalize
^C$A little benchmark where Process 0 send to Process 1 an array of 2 integer but one by one.
$ MPROF_OPTIONS="--profile" mprof mpirun -np 2 ./a.out
==mprof== PROFILE: Process 0 send independently elements which are contiguous to Process 1
==mprof== - Sending 1 element(s) of MPI_INT
===============================================================================
================================= MPI PROFILER ================================
===============================================================================
==mprof== GLOBAL SUMMARY:
==mprof== running: 74 us
==mprof== message sent: 2 msg take 8 bytes - waiting 26 us in total
==mprof== message recv: 2 msg take 8 bytes - waiting 44 us in total
==mprof== barrier(s) passed: 0 - waiting 0 ns in total
==mprof== warning(s): 1 - 1 contiguous send
==mprof==
==mprof== LOCAL SUMMARY (Process 0):
==mprof== running: 46 us
==mprof== message sent: 2 msg take 8 bytes - waiting 26 us (max: 25 us)
==mprof== message recv: 0 msg take 0 bytes - waiting 0 ns (max: 0 ns)
==mprof== barrier(s) passed: 0 - waiting 0 ns (max: 0 ns)
==mprof== list(s) sent to: 1
==mprof== list(s) recv from:
==mprof== warning(s): 1 - 1 contiguous send
==mprof==
==mprof== LOCAL SUMMARY (Process 1):
==mprof== running: 74 us
==mprof== message sent: 0 msg take 0 bytes - waiting 0 ns (max: 0 ns)
==mprof== message recv: 2 msg take 8 bytes - waiting 44 us (max: 35 us)
==mprof== barrier(s) passed: 0 - waiting 0 ns (max: 0 ns)
==mprof== list(s) sent to:
==mprof== list(s) recv from: 0
==mprof==
==mprof== ERROR SUMMARY:
==mprof== No errors
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent