search

icecube / pisa

Monte Carlo-based data analysis
http://icecube.github.io/pisa/
Apache License 2.0
19 stars 49 forks source link

Set number of threads globally with PISA_NUM_THREADS variable #674

Closed atrettin closed 3 years ago

atrettin commented 3 years ago

This PR allows one to set the number of threads to be used for multithreading with one central environment variable PISA_NUM_THREADS. This variable is used by numba as well as any "custom" multithreading implementation (of which there is currently only one, for fast histograms).

Multithreading can be made much more simple than we used to: All one needs to do is to implement things with simple for-loops and then use numba's prange instead of range in for-loops that can be naively parallelized. This is done for the mceq_barr_red stage as well as the fast lookup procedures. I would recommend going this route as well for prob3, but I don't think that I'm going to do this because I'm not using it myself. I had to remove the target keyword in some places because it no longer works.

nuSQuIDS already support multithreading for the most expensive operation of propagating state densities through matter potentials, and here, too, the PISA_NUM_THREADS variable is used.

Multithreading for fast histograms is possible because the Global Interpreter Lock is released during the main loop, allowing Python to continue during the calculation. The input sample is simply split up in chunks and the histogram is calculated for every chunk, and the outputs are added together. The speedup is not entirely a factor of N for N cores, but there is benefit in it.

philippeller commented 3 years ago

Would be good to throw a warning in __init__.py for when users set PISA_NUM_THREADS > 1, but PISA_TARGET = cpu

philippeller commented 3 years ago

otherwise looks to be working correctly

atrettin commented 3 years ago

Would be good to throw a warning in __init__.py for when users set PISA_NUM_THREADS > 1, but PISA_TARGET = cpu

Okay, done!