Benchmark results

Note

The results shown here are based on quite naive, straight forward implementations to use the GPU as a calculation backend. The best possible performance is depending on many things, such as the type of the GPU, driver versions, library versions and bandwidth between CPU and GPU contexts for example.

To achieve the best possible performance, care should be taken to keep the amount of data transfer between CPU and GPU contexts as low as possible.

view_names = ["base", "slight01", "medium01"]
host_name = "NUC"
engine_names = ["naive", "cupy", "numba-cuda", "torch-cuda"]

Show code cell source Hide code cell source

from pathlib import Path
import json

import ipywidgets as ipw
import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd

sns.set()
plt.rcParams['figure.dpi'] = 300

results_dir = Path("/Users/jneines/development/pycon_berlin_2023/mandelbrot/results/")

results = []
for results_file in results_dir.glob("*/*/*.json"):
    with results_file.open("r") as fd:
        result = json.loads(fd.read())
        results.append(result)
        
df = pd.DataFrame(results)

dfs = df[
    (df["view"].isin(view_names))
    & (df["hostname"] == host_name)
    & (df["engine"].isin(engine_names))
].reset_index(drop=True)
print(f"We have {len(dfs)} results to analyze.")

tabs = ipw.Tab()

tabs_children = []
tab_names = []
for view_name, _dfs in dfs.groupby("view"):
    tab_names.append(view_name)
    with plt.ioff():
        output = ipw.Output()
        with output:
            fig, axes = plt.subplots()
            sns.violinplot(data=_dfs, x="engine", y="calculation_time", order=engine_names, ax=axes)
            axes.set_ylim(0, None)
            axes.set_ylabel("Calculation time [s]")
            axes.set_xlabel(None)

            axes.set_xticklabels(["\n" * (x_position%2) + label.get_text() for x_position, label in enumerate(axes.get_xticklabels())])
            axes.set_title(view_name)

            plt.tight_layout()
            plt.show()
        tabs_children.append(output)
        
tabs.children = tabs_children
for index, tab_name in enumerate(tab_names):
    tabs.set_title(index, tab_name)

tabs

We have 120 results to analyze.