Skip to content

Latest commit

 

History

History

hw_stats

Instructions on generating the hardware data for correlation

  • Make sure CUDA_INSTALL_PATH is set and bin/lib directories are in PATH and LD_LIBRARY_PATH
  • Build the benchmarks
# Get the applications, their data files and build them:··
git clone https://github.com/accel-sim/gpu-app-collection··
source ./gpu-app-collection/src/setup_environment··
make -j -C ./gpu-app-collection/src rodinia_2.0-ft··
make -C ./gpu-app-collection/src data··
  • Run the hardware correl
# In this folder there are shell scripts that kick off the hardware exectuion, bundle the resutls then publish them online.
# Obviously your commands will not be exact. The most imporant part of this file is the ./run_hw.py command. It takes the exact same benchmark
# definitions as the run_simulations.py command in the util/job_launching folder - ensuring there is a 1:1 relationship between apps run
# in simulation and apps run in hardware.
# To run a set of apps in hardware use the -B command, just like in run_simulations.py. An example:
./run_hw.py -D 0 -B rodinia_2.0-ft -R 10 -c
# ./run_hw.py --help gives a full explanation, but the important switches in this statement are:
# -D <device number returned by the SDK's deviceQuerry (Note: this is often different from nvidia-smi's device number)>
# -B <benchmark list - same format as run_simulations.py>
# -R <number of times to run the cycle tests (not every hardware run takes the same time - error bars will get created on the x-axis of correlation)>
# -c passing this parameter only runs the cycle correlation data. without this swtich both cycle timing AND all the statistics profiling will be done
#   the full stats profiling takes MUCH longer.