Millicharged Particle Faint Track

Simulate tracks from a millicharged particle (mCP) crossing the SBND detector, and detect those tracks.

See the technote for a high level description of the design/algorithms/optimization strategies.

Run configurations

There are two main files. The main file is faint_tracks.py, which has all the kernel definitions and different run configurations for benchmarking, profiling, etc. The notebook is where I tested parts individually, and not up to date.

faint_tracks.py

Required command line arguments:

• chunks - How many chunks to use for simulate_tracks. It is useful to benchmark different values as it has a large influence on the runtime of simulate_tracks. See this plot for the effect of different sizes for my configuration.
• threads - How many threads to use for track_sums. Can be 1024 at max. It is useful to benchmark different values as it has a large influence on the runtime of track_sums. See this plot for the effect of different #'s of threads.

For example, python faint_tracks.py 128 32.

Optional command line arguments:

• --mode - What mode to run. The options are
  • bench (default) - Run a bunch of iterations of the algorithm, print timing information, and save a histogram of the distances between found and true start/end pairs.
  • profile - Use this when profiling with ncu, to just run each gpu kernel once. Otherwise, it'll take forever to profile since each gpu function gets way slower.
  • chunks - Use this for testing different values of chunks, will just print chunks mean1 stddev1 mean2 stddev2 (for simulate_tracks, 1 is calculating arrays, 2 is combining them).
  • threads -Use this for testing different values of threads, will just print threads blocks mean stddev ( for track_sums).
  • run - Like bench, but for large numbers of events. Will print
• --no-time-all - For bench mode, use this to just time the total elapsed time and not each individual part.
• --iters - How many iterations to run (default 100) for bench, chunks, or threads.
• -z - The charge of the particle to simulate (defaults to 1.0).
• --gpu - Select which gpu to use, zero indexed. Use nvidia-smi to see details about available gpus. Defaults to 3 because currently on spsfarm, the other 3 gpus are stuck in performance mode P2 (second highest), meaning they aren't able to cool down when not in use, so they are slower.