fMRIPrep next preliminary results #3162
Description
I have run two datasets for testing using fMRIPrep 23.1.4 and 23.2.0a2. Just opening up an issue for discussion.
Run commands
| Name | Command |
|---|---|
| Old | time fmriprep-docker -i nipreps/fmriprep:23.1.4 sourcedata/raw . participant -w $SCRATCH --cifti-output --output-spaces MNI152NLin2009cAsym |
| Minimal | time fmriprep-docker -i nipreps/fmriprep:23.2.0a2 sourcedata/raw . participant -w $SCRATCH --cifti-output --output-spaces MNI152NLin2009cAsym --level minimal |
| Full | time fmriprep-docker -i nipreps/fmriprep:23.2.0a2 sourcedata/raw . participant -w $SCRATCH --cifti-output --output-spaces MNI152NLin2009cAsym --level full |
Timing and storage results
| Dataset | Run | Runtime | Scratch Size | Scratch Files | Output Size | Output files |
|---|---|---|---|---|---|---|
| A | Old | 2h24m | 54.8GB | 36.8K | 2.30GB | 176 |
| A | Minimal | 1h35m | 2.91GB | 5.89K | 602MB | 128 |
| A | Full | 1h47m | 19.8GB | 10.0K | 6.37GB | 206 |
| B | Old | 4h25m | 121GB | 157K | 5.10GB | 286 |
| B | Minimal | 1h29m | 1.88GB | 12.0K | 543MB | 206 |
| B | Full | 2h7m | 56.5GB | 19.8K | 14.7GB | 348 |
Notes
An interesting result is the dramatic reduction in scratch size (>50%) between the old and full runs. I would attribute this primarily to the single-shot resampling eliminating the need for splitting, applying and merging.
A surprising result is the increase in the output size. One component was #3159. The other is that the outputs are now saved in float32 instead of the original (int16, in these cases) dtypes.
The runtime reduction is most likely attributable to the elimination of the per-volume fieldmap resampling that was previously implemented in sdcflows. However, I do not have a profile to confirm this.