GPU Accelerated n-body system

3D nbody simulation using the Naive or Barnes-Hut approximation algorithm on CPU or Nvidia GPU. A 2D projection of the bodies is rendered realtime with imgui, OpenGL, glm and GLFW. The CPU version also (optionally) relies on OpenMP.

If a NVIDIA graphics card and CUDA is avilable the computations can be parallelized increasing frame performance significantly.

Running Instructions

git clone https://github.com/benni012/nbody

Make sure to update the arch flag according to your systems Nvidia GPU in CMakeLists.txt (line 67).

cmake .
make

then everything should be set up to run. For example 1k particles with the Barnes-Hut algorrithm on the GPU running for 100 iterations:

./nbody -dgpu -abh -n1e3 -i100

A detailed explanation of the available arguments

Option	Flag	Argument Type	Description	Default
--benchmark	-b	None	Runs in benchmark mode.	false
--device	-d	cpu/gpu	Specifies whether to use CPU or GPU.	cpu
--algo	-a	bh/naive	Selects the algorithm (bh for Barnes-Hut, naive for direct).	naive
--num-particles	-n	Number	Sets the number of particles.	5000
--iters	-i	Number	Defines the number of iterations.	UINT_MAX
--pop-method	-p	Number	Defines the population method (0 Uniform Disk, 1, Uniform disk with central mass, 2 Uniform stable disk with movement, 3 Plummers Model)	3

Benchmark option outputs (and saves to a csv) the average function times, for example:

$ ./nbody -anaive -dgpu -n200 -b -i500
Timings:
Function            Calls     Mean Time (us)      Std Dev (us)
BodiesD2H           500       29                  8                   
PosUpdate_GPU       500       9                   1                   
AccUpdateNaive_GPU  500       112                 183                 

Benchmarking

To recreate the benchmarking plots the Visual Profiler (nvprof) and Python (>= 3) have to be installed. Afterwards run:

python benchmarks/run_benchmark.py

From inside the /benchmarks directory, run:

python plot_roof.py

python plot_runtime_vs_time.py