seticore

A high-performance implementation of some core SETI algorithms, in particular turboSETI's dedoppler algorithm that creates .dat files from .h5 files, and a beamforming algorithm to produce filterbanks from .raw and .bfr5 files.

I want to analyze data generated by seticore

See the Python docs.

I want to run seticore

It should already be installed on the BL machines on MeerKAT, the VLA, and Berkeley. Run seticore with no arguments for help. Ping Kevin or file an issue if this is insufficient.

I want to develop on the seticore codebase

Install dependencies:

sudo apt-get install cmake g++ libboost-all-dev libhdf5-dev pkg-config

You will also need a Python environment for your build tools:

python3 -m pip install meson
python3 -m pip install ninja

Git clone this repo, then update submodules, then run the make scripts:

git submodule init
git submodule update
meson setup build
cd build
meson compile

then run on an h5 file:

./seticore /path/to/your.h5

It will behave roughly like turboseti. If something doesn't immediately work, try the more detailed instructions below.

Fixing hdf5 plugin errors

Depending on how you installed hdf5, you may not have the plugins that you need, in particular the bitshuffle plugin. If seticore tells you that you need to make sure that plugin files are in the plugin directory, check if that directory exists. If not, we'll have to install them.

This is not straightforward because the hdf5 group does not support plugins on Ubuntu, and the bitshuffle plugin developer only supports Python. So we use the Python install and copy its files over.

Create the plugin directory that it expects
pip install hdf5plugin
Find the plugin file location: pip show hdf5plugin | grep Location
Look around for a directory full of .so files. Maybe it's in $LOCATION/hdf5plugin/plugins
Copy all of those .so files to the directory you created

We don't need the hdf5plugin install any more once its files have been copied out.

Testing on your machine

You can run the unit tests with:

./unit_tests.sh

This will automatically download a few gigabytes of test data from Berkeley and ensure you get the expected output.

The main supported target environment is defined by this Dockerfile. If you have the Nvidia container toolkit set up to let Docker access the GPU, you can test the code using this Docker environment with:

./test_docker.sh

Debugging

To get a binary with debug symbols, so you can run gdb:

meson configure --buildtype=debug build

When you're done, configure it back:

meson configure --buildtype=release build

Selecting a Specific GPU Device ID

Reference: https://shawnliu.me/post/nvidia-gpu-id-enumeration-in-linux/

It's easy to screw this up because nvidia-smi doesn't give GPUs the same numbering as the CUDA library does by default. To use the nth GPU, where you've determined n by looking at nvidia-smi:

CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=$n seticore <args>

This makes only the nth GPU visible to seticore. This same procedure works for any binary that uses CUDA, not just seticore.

Profiling

Install some dependencies:

sudo apt-get install perf
pushd ~
git clone git@github.com:brendangregg/FlameGraph.git
popd

Then do a profiling run with your arguments of choice:

perf record -g ./build/seticore <arguments-for-seticore>
perf script > perf.script
~/FlameGraph/stackcollapse-perf.pl perf.script | ~/FlameGraph/flamegraph.pl > perf.svg

Then look at perf.svg in a browser.

Viewer

A Python notebook is provided to show how to analyze the outputs. Set up Python 3.9 in whichever way you prefer (such as a fresh conda env) and then:

cd python
pip install -r requirements.txt
jupyter-lab example.ipynb

You can also read the schemas for the hit file and the stamp file.

n-west / seticore