FLAMEGPU2 Circles Benchmark

This repository contains performance benchmarking of a FLAME GPU 2 implementation of the Circles agent based model at various population scales and densities.

In the Circles model is an abstract benchmark model which is used to evaluate neighbourhood search, with agents interacting with other agents within their local neighbourhood. For a more complete description of the model, see:

Chisholm, Robert, Paul Richmond, and Steve Maddock. "A standardised benchmark for assessing the performance of fixed radius near neighbours." European Conference on Parallel Processing. Springer, Cham, 2016., (pdf).

Benchmark Description and Results

Three experiments are undertaken within this benchmark. There is a range of raw data in the sample/data directory with a description of the machine configurations used to generate it in each directory.

The results below are from the V100 runs on the Bessemer HPC system at the University of Sheffield. Job submission scripts are included in the scripts/slurm directory.

A combined figure for use in publication is shown below. For details please refer to the publication.

Fixed Density Benchmark

• Communication Radius is fixed to 2.0
• Agent Density is fixed to 1 agent per unit of volume
• Environment Volume is varied, with values of up to 1000000 units of volume
• 4 Implementations are compared
  • Bruteforce messaging
  • Bruteforce messaging with RTC (run time compilation)
  • Spatial3D messaging
  • Spatial3D messaging with RTC (run time compilation)

[![Fixed Density Benchmark Zoomed](sample/figures/v100-515.65.0A CUDA 11.0

Variable Density Benchmark

• Communication Radius is fixed to 2.0
• Agent Density is varied per unit of volume, from 1 to 4
• Environment Volume is varied up to ~ 500000 units of volume.
• A single implementation is benchmarked
  • Spatial3D messaging with RTC (run time compilation)

Variable Communication Radius Benchmark

• circles_spatial3d_rtc and circles_bruteforce_rtc
• Varied communication radius - shown on the X axis
• Model differentiated by marker
• Agent count is fixed at 64,000
• Environment width is fixed at 40
• Agent Density is 1.0f

Variable Sort Period Benchmark

• Runs for communication radii of 2.0, 4.0, 6.0, 8.0
• Sort period is varied - shown on the X axis - a sort period of 0 represents no sorting, simulation runs for 200 steps in all cases
• Model differentiated by marker
• Agent count is fixed at 64,000
• Environment width is fixed at 40
• Agent density is 1.0f

Building and Running the Benchmark

Detail of dependencies and the cmake build process are described in full in the FLAMEGPU2-example-template Repo and are not repeated here. The benchmark should be built with seatbelts off (e.g. -DFLAMEGPU_SEATBELTS=OFF passed to the cmake configuration step) to disable additional run-time checks. E.g. for Volta (SM_70) GPUs under Linux.

# Configure 
cmake . -B build -DCMAKE_BUILD_TYPE=Release -DFLAMEGPU_SEATBELTS=OFF -DCMAKE_CUDA_ARCHITECTURES=70
# Build
cmake --build build -j`nproc` 

Execution and Data generation

cd build
./bin/Release/circles-benchmark

This will produce a number of .csv files in the build directory.

Note: The FLAMEGPU2_INC_DIR environment variable may need to be set to ./_deps/flamegpu2-src/include/ for run-time compilation (RTC) to succeed if the source directory is not automatically found.

Plotting Results

• A combined figure used for publication (see https://doi.org/10.1002/spe.3207) can be produced using plot_publication.py.
• Individual figures can be generated from data in CSV files via a python script plot.py.
• Figures comparing the performance from multiple runs can be created using plot_comparisons.py

Dependencies

It is recommended to use python virtual environment or conda environment for plotting dependencies.

I.e. for Linux to install the dependencies into a python3 virtual environment and plot the results from all experiments output to the build directory.

# From the root of the repository
# Create the venv
python3 -m venv .venv
# Activate the venv
source .venv/bin/activate
# Install the dependencies via pip
python3 -m pip install -Ur requirements.txt
# Plot using csv files contained within the build directory
python3 plot.py build -o build/figures
# Use -h / --help for more information on optional plotting script parameters.

Publication Plots

Figures of the style used in "FLAME GPU 2: A framework for flexible and performant agent based simulation on GPUs" (https://doi.org/10.1002/spe.3207) can be generate from a directory of CSV files via plot_publication.py.

E.g. The CUDA 11.0 V100 plot in this readme and in the publication was generated via:

python3 plot_publication.py -i sample/data/v100-515.65.01/2.0.0-rc-v100-11.0-beltsoff -o sample/figures/v100-515.65.01/2.0.0-rc-v100-11.0-beltsoff

Individual Plots

Individual figures for each experiment in the benchmark with differing axis can be generated by plot.py. The sample figures in this readme using a CUDA 11.0 on a V100 were generated from the root directory via:

python3 plot.py -i sample/data/v100-515.65.01/2.0.0-rc-v100-11.0-beltsoff -o sample/figures/v100-515.65.01/2.0.0-rc-v100-11.0-beltsoff

Comparison Plots

plot_comparisons.py can be used to compare the performance of multiple runs of individual benchmark experiments, which can be useful when benchmarking different hardware or software versions. To produce comparisons plots, such as those in comparisons/h100-a100-v100-fixed-density:

1. Run plot.py for each individual benchmark run, to generate intermediate data (see above)
2. Write a YAML configuration file to configure the comparison plots and figure properties. e.g. sample/comparisons/h100-a100-v100-fixed-density/config.yml
3. Run plot_comparisons.py, e.g ./plot_comparisons.py -c sample/comparisons/h100-a100-v100-fixed-density/config.yml

This will produce a figure per SIMULATORS entry such as the following:

And a plot showing relative performance compared to a single input directory.

Docker and Singularity/Apptainer

A Dockerfile is provided to enable creation of a container based on a CUDA 11.8 base image, with builds for SM 70, 80 and 90 GPUs.

Once complete, the dockerfile can be converted to apptainer / singularity for execution on HPC systems where docker is unavailable.

For example, the following commands can be used:

# Build the docker container, which may require root.
docker build . -t flamegpu2-circles-benchmark-11.8
# Run the benchmark via docker, may require root
docker run --rm -it --gpus all  -v .:/app -w /app flamegpu2-circles-benchmark-11.8 /opt/FLAMEGPU2-circles-benchmark/build/bin/Release/circles-benchmark
# Create an apptainer container, from the latest version of the docker container built locally. May require root.
apptainer build flamegpu2-circles-benchmark-11.8.sif docker-daemon://flamegpu2-circles-benchmark-11.8:latest
# Run via apptainer, shouldn't require root
apptainer exec --nv --cleanenv flamegpu2-circles-benchmark-11.8.sif /opt/FLAMEGPU2-circles-benchmark/build/bin/Release/circles-benchmark