[antonio-dimeglio]

Summary

A tracking run with no output file kept each batch's streamlines as a tractogram of its own and joined them at the end, so for a moment it held every streamline twice, plus the per-batch copies. On a large run that pushes a 16 GB machine into swap, which slows the end of the run and whatever comes next, such as the viewer's Dissect workspace.

The kept points now go straight from each batch's buffer into one buffer that grows by a quarter when full, in place where the allocator allows it, and is trimmed at the end. Runs written to a file are unchanged. The result is the same tractogram, point for point, with the same end statuses.

Measurements

One million streamlines on the ISMRM example data (det, whole brain mask, seed 1), 431.6 million points (5.18 GB), on an Apple M-series laptop with 16 GB. Peak memory footprint from /usr/bin/time -l, which counts swapped and compressed memory; the same SHA-256 of the points and offsets in both runs.

	Peak footprint	Tracking	Whole run
Before	11.0 GB	71 s	83 s
After	6.5 GB	62 s	64 s

The memory beyond the tractogram itself falls from 5.8 GB to 1.3 GB. A growth of 1.5 instead of 1.25 gave 7.4 GB at the same speed.

Tests

The existing test that compares a run kept in memory with the same run written to TCK and TRK, with batches of seven seeds, covers the buffer growing several times; the tracking, viewer and CLI tests pass.