Numba-Numpy Benchmarking

Overview

This repository contains benchmarking results for Numba with NumPy. We conducted benchmark tests to evaluate the performance of Numba when using NumPy compared to using raw loops.

Conditions

• Numba version: 0.57.1
• Numpy version: 1.24.4
• Python version: 3.11.4
• ROOT version: 6.29/01
• Each test was run 10 times and an average of the execution time was calculated.
• The tests were run on a linux fedora 38 station.

Real Time Benchmarking:

This figure represents the results of our benchmarking for real-time performance. The x-axis represents the number of threads, and the y-axis represents the average real-time execution in seconds.

Some observations:

• As the number of threads increases, there is a noticeable decrease in average real-time execution for both the loop and Numpy implementations. This suggests that Numba, when combined with NumPy, can take advantage of multi-threading and scale with the number of threads.

• The loop implementation consistently outperforms the Numpy implementation as the number of threads increases.

Speedup Benchmarking:

This figure illustrates the speedup factor when using Numba with NumPy compared to the loop implementation. The x-axis represents the number of threads, and the y-axis represents the speedup factor.

Some observations:

• The speedup factor indicates how much faster the Numba with both NumPy and loop implementations are. Numba had issues with this in the past versions and the scaling wasn't as expected when using Numpy.

• As observed in the figure, the speedup factor increases with the number of threads. For a higher number of threads, 8 in this example, the speedup factor reaches a value of 6.6. This confirms that Numba with NumPy scales linearly with the number of threads, similarily to how the loop implementation scales.