lab-nbody

N-Body CUDA simulation - Simple all-pairs N-Body algorithm.

Getting started

Build

git clone https://github.com/gcoe-dresden/lab-nbody.git
cd lab-nbody
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make

Run

./nbody [nbodies] [device-index]
# Runs nbody with 8192 nbodies
./nbody 8192

The current version includes validation of the results on the CPU. However, due to the serial performance, validation is restricted to <8193.

Tasks

• Analyze the code for possible performance improvements.
• Choose the most promising optimization and implement it.
• Measure the speedup.
• Repeat the process.

Questions

• What is the maximal performance?
• The measured performance can be higher than the maximal performance. Why?
• Does it scale?
• How is the performance for different blocksizes?
• What is the maximum blocksize?

Profiling CUDA kernels - Current Issues

With CUDA driver 418.43+ admin privileges are required to gather the metrics/events from the device. With older CUDA versions < 10.2, errors look a bit different, e.g. Error: Internal profiling error 4183:7. As of CUDA driver version 418.43 privileges for root/CAP_SYS are required, to use nvprof CUPTI, see:

 https://docs.nvidia.com/cupti/Cupti/r_overview.html#r_whats_new