Last active
April 28, 2023 21:43
-
-
Save braingram/ba388b166bc08b2c42ce1abee317f1b0 to your computer and use it in GitHub Desktop.
Performance of memmapping vs chunking (zarr)
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import os | |
import shutil | |
import time | |
import numpy | |
import zarr | |
n_times = 10 | |
image_size = (4000, 4000, 50) | |
dtype = 'uint8' | |
mm_fn = 'test_memmap.bin' | |
zarr_dir = 'test_zarr' | |
def setup_memmap(): | |
# generate the memmap file | |
arr = numpy.zeros(image_size, dtype) | |
with open(mm_fn, 'wb') as f: | |
f.write(arr.tobytes()) | |
return numpy.memmap(mm_fn, shape=image_size, dtype='uint8') | |
def re_memmap(): | |
return numpy.memmap(mm_fn, shape=image_size, dtype='uint8') | |
def setup_zarr(chunk_shape): | |
# generate the zarr directory store | |
if os.path.exists(zarr_dir): | |
shutil.rmtree(zarr_dir) | |
z = zarr.open(zarr_dir, mode='w', shape=image_size, chunks=chunk_shape, dtype=dtype) | |
z[:] = 0 | |
return z | |
chunk_shapes = [ | |
image_size, # one chunk for whole image | |
[1000, 1000, 50], # first 2 dimensions | |
[4000, 4000, 10], # only z | |
[1000, 1000, 10], # all 3 | |
[300, 300, 1], # very small chunks | |
] | |
accessors = [ | |
((lambda arr: arr[:]), "whole array"), | |
((lambda arr: arr[:, 0, 0]), "first dimension slice"), | |
((lambda arr: arr[0, :, 0]), "second dimension slice"), | |
((lambda arr: arr[0, 0, :]), "third dimension slice"), | |
((lambda arr: arr[10:40, 20:50, 0]), "cutout for 1 z"), | |
((lambda arr: arr[:30, :30, 0]), "cutout aligned to one chunk"), | |
((lambda arr: arr[10:40, 20:50, :]), "chutout for all z"), | |
((lambda arr: arr[100:400, 200:500, 0]), "larger cutout for 1 z"), | |
((lambda arr: arr[:300, :300, 0]), "larger cutout aligned to one chunk"), | |
((lambda arr: arr[100:400, 200:500, :]), "larger cutout for all z"), | |
] | |
def time_it(arr, accessor, n_times, mean=True): | |
ts = [] | |
v = 0 | |
for _ in range(n_times): | |
t0 = time.monotonic() | |
v += numpy.sum(accessor(arr)) | |
t1 = time.monotonic() | |
ts.append(t1 - t0) | |
if not mean: | |
return ts | |
return numpy.mean(ts) | |
if __name__ == '__main__': | |
print(f"Running each test {n_times} times") | |
print(f"Testing image shape: {image_size}") | |
mm = setup_memmap() | |
for cs in chunk_shapes: | |
print(f"\tchunk shape: {cs}") | |
z = setup_zarr(cs) | |
for (a, an) in accessors: | |
# to make this 'fair' make a new memmap every time | |
mm = re_memmap() | |
zt = time_it(z, a, n_times) | |
mmt = time_it(mm, a, n_times) | |
winner = 'memmap' if mmt < zt else 'chunking' | |
print(f"\t\t{winner} win! {an}: chunking={zt}, memmap={mmt}") |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Good point! I updated it to sum the array (and updated the results above).