JUPITER Benchmark Suite: PIConGPU

This benchmark is part of the JUPITER Benchmark Suite. See the repository of the suite for some general remarks.

This repository contains the PIConGPU benchmark. DESCRIPTION.md contains details for compilation, execution, and evaluation.

The source code of PIConGPU is included in the ./src/ subdirectory as a submodule from the upstream PIConGPU repository at github.com/ComputationalRadiationPhysics/picongpu.

Quickstart

# Obtain the benchmark if not present
git clone --recursive $REPO
# Run the benchmark using JUBE
jube run benchmark/jube/benchmark.xml -t {baseline,high_large,high_medium,high_small}

This will obtain the required sources from GitHub, perform a full build and run the benchmark. After the benchmark runs are complete, one can analyse the benchmark runs and generate a results table by executing the following,

jube analyse benchmark/jube/run --id XYZ
jube result benchmark/jube/run --id XYZ

Results

Example results are given in the following

Baseline

The baseline test uses 8 nodes on the JUWELS booster with 4 GPUs per node and uses a grid size -g 1024 1024 512. Following is the example output on the JUWELS booster using "jube run benchmark.xml -t baseline"

version	system	nnodes	nconfig	gridconfig	simtime
26cbd45b2c9	juwelsbooster	8	-d 4 4 2	-g 1024 1024 512	466.155

Large Scale

This is the second mode of the benchmark, designed to run at a scale of around 1000 PFLOP/s. Below is the example out sourced from JUWELS Booster and its validation data.

version	system	nnodes	nconfig	gridconfig	simtime
26cbd45b2c9	juwelsbooster	256	-d 16 8 8	-g 4096 2048 2048	466.567
26cbd45b2c9	juwelsbooster	384	-d 16 16 6	-g 4096 3072 2048	470.815
26cbd45b2c9	juwelsbooster	512	-d 16 16 8	-g 4096 4096 2048	466.618
26cbd45b2c9	juwelsbooster	640	-d 16 16 10	-g 4096 4096 2560	466.785