Skip to content

Latest commit

 

History

History
55 lines (45 loc) · 3.06 KB

README.md

File metadata and controls

55 lines (45 loc) · 3.06 KB

Running benchmarks

This section is part of the reproductibility initiative of SC24. Benchmark are implemented using the google benchmark microbenchmark suite.

Prerequisites

  • Apptainer/Singularity must be installed on the machine.
  • Pull the sycl-complete container.
    • apptainer pull docker://ghcr.io/maison-de-la-simulation/sycl-container:llvm-16
  • Build google benchmark library with classical compiler (can be built in the submodule folder thirdparty/benchmark).
  • A cluster with SLURM scheduler

Instructions

  1. Inside the container: compile the parallel-advection project with benchmark option (see README). The file main_bench.cpp contains the main bencmarks.

  2. Update the launch_benchmark.sh script on a SLURM scheduled cluster to submit the benchmarks to the compute nodes

    • Select the right SLURM environment in header
    • Update CONTAINER_RUN (full path to the .sif file)
    • Update OUTFILE_DIR (directory where to output the .json file with benchmark resutls) and BUILD_DIR (directory containing the benchmark executable)
    • Update correct values for HW (hardware), IMPL (sycl implementation), ARGS_GPU_SINGULARITY if GPU benchmarking (typically --nv or --rocm)
    • Use the BENCHMARK_FILTER environment variable. e.g. BENCHMARK_FILTER=BM_Advector/0 or BM_Advector/1 (0 for CPU 1 for GPU)

  3. Once the launch_benchmark.sh script is updated with correct values, submit it to the SLURM scheduler with sbatch launch_benchmarks.sh

Full example: A100 GPU with AdaptiveCpp

# Inside parallel-advection folder with submodules pulled

$ export APPTAINER_BIND="`pwd`:/par-adv"
$ apptainer pull docker://ghcr.io/maison-de-la-simulation/sycl-container:llvm-16
$ apptainer shell sycl-container_llvm-16.sif

#inside container
Singularity> cd /par-adv/thirdparty/benchmark
Singularity> mkdir build && cmake .. -D.... #compile google benchmark apart
Singularity> cd /par-adv
Singularity> ./compile.sh --hw a100 --sycl acpp --benchmark_BUILD_DIR=/par-adv/thirdparty/benchmark/build
Singularity> exit #get out of the container

#On a cluster with SLURM,
#Update values for SLURM scheduler and correct values for benchmark (here we target a100 and acpp, we will need the --nv flag for singularity)
$ vi diag/launch_benchmarks.sh
$ sbatch diag/launch_benchmarks.sh #run the benchmarks for A100

Visualize results

To visualize the results with python (pandas and matplotlib are required):

from utils.py import * #import diag/utils.py functions to vizualise
df_acpp_a100 = get_cleaned_df("LOGS/A100/a100_acpp_50rep.json") #turn json into DataFrame
a100_acpp = create_dict_from_df(df_acpp_a100) #turn DataFrame into a easily plotable list
plot_values(a100_acpp, "A100 GPU ACPP", do_show=true) #plot the values and show

Known issues

  • With AdaptiveCpp on the MI250X, export the ACPP_PERSISTENT_RUNTIME=1 variable because it causes crash with google benchmark framework.