Numbagg: Fast N-dimensional aggregation functions with Numba

Fast, flexible N-dimensional array functions written with Numba and NumPy's generalized ufuncs.

Why use numbagg?

Performance

• Outperforms pandas
  • On a single core, 2-10x faster for moving window functions, 1-2x faster for aggregation and grouping functions
  • When parallelizing with multiple cores, 4-30x faster
• Outperforms bottleneck on multiple cores
  • On a single core, matches bottleneck
  • When parallelizing with multiple cores, 3-7x faster
• Outperforms numpy on multiple cores
  • On a single core, matches numpy
  • When parallelizing with multiple cores, 5-15x faster
• ...though numbagg's functions are JIT compiled, so the first run is much slower

Versatility

• More functions (though bottleneck has some functions we don't have, and pandas' functions have many more parameters)
• Functions work for >3 dimensions. All functions take an arbitrary axis or tuple of axes to calculate over
• Written in numba — way less code, simple to inspect, simple to improve

Functions & benchmarks

Summary benchmark

Two benchmarks summarize numbagg's performance — the first with a 1D array of 10M elements without parallelization, and a second with a 2D array of 100x10K elements with parallelization. Numbagg's relative performance is much higher where parallelization is possible. A wider range of arrays is listed in the full set of benchmarks below.

The values in the table are numbagg's performance as a multiple of other libraries for a given shaped array calculated over the final axis. (so 1.00x means numbagg is equal, higher means numbagg is faster.)

Full benchmarks

[^1][^2][^3][^4]

[^1]: Benchmarks were run on a Mac M1 laptop in December 2023 on numbagg's HEAD, pandas 2.1.1, bottleneck 1.3.7, numpy 1.25.2, with python numbagg/test/run_benchmarks.py -- --benchmark-max-time=10. They run in CI, though GHA's low CPU count means we don't see the full benefits of parallelization.

[^2]: While we separate the setup and the running of the functions, pandas still needs to do some work to create its result dataframe, and numbagg does some checks in python which bottleneck does in C or doesn't do. So use benchmarks on larger arrays for our summary so we can focus on the computational speed, which doesn't asymptote away. Any contributions to improve the benchmarks are welcome.

[^3]: In some instances, a library won't have the exact function — for example, pandas doesn't have an equivalent move_exp_nancount function, so we use its sum function on an array of 1s. Similarly for group_nansum_of_squares, we use two separate operations.

[^4]: anynan & allnan are also functions in numbagg, but not listed here as they require a different benchmark setup.

[^5]: This function is not currently parallelized, so exhibits worse performance on parallelizable arrays.

Example implementation

Numbagg makes it easy to write, in pure Python/NumPy, flexible aggregation functions accelerated by Numba. All the hard work is done by Numba's JIT compiler and NumPy's gufunc machinery (as wrapped by Numba).

For example, here is how we wrote nansum:

import numpy as np
from numbagg.decorators import ndreduce

@ndreduce.wrap()
def nansum(a):
    asum = 0.0
    for ai in a.flat:
        if not np.isnan(ai):
            asum += ai
    return asum

Implementation details

Numbagg includes somewhat awkward workarounds for features missing from NumPy/Numba:

• It implements its own cache for functions wrapped by Numba's guvectorize, because that decorator is rather slow.
• It does its own handling of array transposes to handle the axis argument in reduction functions.
• It rewrites plain functions into gufuncs, to allow writing a traditional function while retaining the multidimensional advantages of gufuncs.

Already some of the ideas here have flowed upstream to numba (for example, an axis parameter), and we hope that others will follow.

License

3-clause BSD. Includes portions of Bottleneck, which is distributed under a Simplified BSD license.