Skip to content

Latest commit

 

History

History
71 lines (53 loc) · 2.44 KB

README.md

File metadata and controls

71 lines (53 loc) · 2.44 KB
Raw

MPI-Wattmeter

MPI-Wattmeter measures the power consumption of MPI programs.

Running an application

mpirun mpi_wattmeter [options] myappli

The following options can be passed to MPI Interceptor:

  • -v or --verbose
    • Set verbosity level (default level: 0)
  • -d or --details
    • Print detailed measurements (default level: 0)

Building MPI-Wattmeter

git clone https://github.com/trahay/mpi-wattmeter.git
mkdir mpi-wattmeter/build
cd mpi-wattmeter/build
cmake ..  -DCMAKE_INSTALL_PREFIX=$PWD/../install
make install

Example

In this example, we run NAS Parallel Benchmark BT kernel:


$ mpirun -np 4 mpi_wattmeter  ./bin/bt.A.4 
[...]
============================
MPI-Wattmeter summary
============================
#Node:rank      Source          joules          Watt.hour 
3a401-01:0      energy-cores    216.070618      0.060020  
3a401-01:0      energy-pkg      252.904602      0.070251  
3a401-01:0      energy-ram      13.127563       0.003647  
3a401-02:1      energy-cores    194.357422      0.053988  
3a401-02:1      energy-pkg      230.965149      0.064157  
3a401-02:1      energy-ram      12.479248       0.003466  
3a401-03:2      energy-cores    216.447571      0.060124  
3a401-03:2      energy-pkg      253.188782      0.070330  
3a401-03:2      energy-ram      12.913330       0.003587  
3a401-04:3      energy-cores    196.919067      0.054700  
3a401-04:3      energy-pkg      233.507263      0.064863  
3a401-04:3      energy-ram      12.370789       0.003436  

Total:
#Source         total(j)        avg(j)          min(j)          max(j)          total(W.h)      avg(W.h)        min(W.h)        max(W.h)  
energy-cores    823.794678      205.948669      194.357422      216.447571      0.228832        0.057208        0.053988        0.060124  
energy-pkg      970.565796      242.641449      230.965149      253.188782      0.269602        0.067400        0.064157        0.070330  
energy-ram      50.890930       12.722733       12.370789       13.127563       0.014136        0.003534        0.003436        0.003647

Adding the -d option prints the collected data for each machine.

Several hardware counters are collected:

  • energy-cores: energy consumed by the CPU
  • energy-gpu: energy consumed by the integrated GPU
  • energy-ram: energy consumed by the RAM
  • energy-pkg: energy consumed by the whole socket (usually this corresponds to CPU+RAM)
  • energy-psys: total energy consumed by the machine