/Instruction_roofline_scripts

Primary LanguagePython

Instruction Roofline for Adept

This repository contains the essential scripts of profiling, building, and visualizing the Instruction Roofline Model for the Adept kernels on NVIDIA GPUs.

Usage Instructions

  1. Open the file collect_metric_extended.sh and edit the lines 7 to 14 to reflect Adept application paths.
7 # path to adept's repo home
8 apphome=$HOME/repos/mhaseeb/adept_revamp
9
10 # path to adept_test app and three arguments to it
11 app=${apphome}/build/adept_test
12 arg1=/global/cscratch1/sd/mhaseeb/sw-benchmarks/ref_set_1.fasta
13 arg2=/global/cscratch1/sd/mhaseeb/sw-benchmarks/read_set_1.fasta
14 arg3=${apphome}/build/aligned.out
  1. Allocate a GPU node (preferably with --exclusive) using the following command:
salloc -G 1 -C gpu -t 00:30:00 --exclusive
  1. Run the collect_metric_extended.sh script as:
srun ./collect_metric_extended.sh
  1. After profiler completion, run the merge.sh to obtain the Adept_F.csv and Adept_R.csv files using:
srun ./merge.sh ./output_extended
  1. Finally, run the adept_roofline.py script with proper parameters to build and visualize the roofline for both Adept kernels as:
# get help on input parameters
python ./adept_rooflines.py --help

# run adept_rooflines.py with correct parameters
srun python ./adept_rooflines.py -f ./output_extended/Adept_F.csv --ftime 193.94 -r ./output_extended/Adept_R.csv --rtime 61.853 && ps2pdf ./adept_glob.eps && ps2pdf ./adept_shm.eps

References

  1. Ding, Nan, and Samuel Williams. "An instruction roofline model for gpus." 2019 IEEE/ACM Performance Modeling, Benchmarking and Simulation of High Performance Computer Systems (PMBS). IEEE, 2019