Slow reading of uncompressed nifti volumes when memory mapping is enabled

[clane9]

It seems that reading uncompressed nifti volumes takes significantly longer when memory mapping is enabled. In this example, I load four large 4D nifti volumes from the ABCD dataset, with and without memory mapping.

import time

import nibabel as nib
import numpy as np

print(nib.__version__)


def benchmark_nibabel_load(path: str, mmap: bool):
    tic = time.monotonic()
    img = nib.load(path, mmap=mmap)
    data = np.asarray(img.get_fdata())
    rt = time.monotonic() - tic
    print(f"mmap: {mmap}, run time: {rt:.3f}s")


paths = [
    "/ABCD/sub-NDARINV0CCVJ39W/ses-2YearFollowUpYArm1/func/sub-NDARINV0CCVJ39W_ses-2YearFollowUpYArm1_task-rest_run-04_bold.nii",
    "/ABCD/sub-NDARINV0YE7L9KU/ses-2YearFollowUpYArm1/func/sub-NDARINV0YE7L9KU_ses-2YearFollowUpYArm1_task-rest_run-04_bold.nii",
    "/ABCD/sub-NDARINV0F82C6R8/ses-2YearFollowUpYArm1/func/sub-NDARINV0F82C6R8_ses-2YearFollowUpYArm1_task-rest_run-02_bold.nii",
    "/ABCD/sub-NDARINV0V80916L/ses-baselineYear1Arm1/func/sub-NDARINV0V80916L_ses-baselineYear1Arm1_task-rest_run-04_bold.nii",
]

mmaps = [True, False, True, False]

for path, mmap in zip(paths, mmaps):
    benchmark_nibabel_load(path, mmap=mmap)

This is what I get, using nibabel v5.2.1:

mmap: True, run time: 90.764s
mmap: False, run time: 3.595s
mmap: True, run time: 37.405s
mmap: False, run time: 4.810s

Any idea why this might happen?