A/B benchmark where the fix swaps pipes for temp files: put temp files on local disk

When a parallel tool's fix replaces worker->parent pipes with per-worker temp files (e.g. to avoid workers blocking on a full 64 KB pipe buffer while the parent drains another worker), and you then time old vs new on a shared host, check where the temp dir lands. On university/cluster hosts the home directory is often NFS; if TMPDIR or the kit's work dir defaults to home, the new arm pays network write+read latency that the pipe-based old arm never had, understating the fix. Pin the worker temp dir to local disk (e.g. /tmp during the slot), record it in the kit, log df before/after each arm, and keep inputs pre-read so cold NFS caches don't favour whichever arm runs second.