Tier Two prep: fresh native vs origin/main vs Tier One benchmark comparison

[factory-ain3sh]

Tier Two prep — fresh benchmark comparison

Native vs Original AgentFS (origin/main 3a5ed2b, AgentFS 0.6.4)
vs Tier One AgentFS (phase4-north-star-implementation 9be0da4,
the kernel-cache-by-default ship).

All runs on the same machine with no AGENTFS_FUSE_* env vars set, release builds.

---

Headline (ratio of agentfs / native; lower is better)

Workload	Original	Tier One	Δ
Read-heavy (full run incl. startup)	2.51x	3.03x	+21%
Read-heavy (steady-state only)	7.76x	3.79x	−51%
Copy-on-write edit (50 MiB file)	8.19x	5.42x	−34%
Mixed git workload	5.16x	3.21x	−38%

Plus: CoW delta DB growth (overlay copy-up footprint, lower is better):
Original 172.6 MiB → Tier One 50.4 MiB (−71%)

---

Read-heavy detail (read-path-benchmark.py, cold + warm modes)

8 files / 2 dirs / 64 KiB each; 8 iters each of stat-storm, readdir-storm,
open-read-close, repeated-open-read on a steady-state mount.

Phase	Native (s)	Original (s)	Tier One (s)	Orig	Tier One
cold/STARTUP+WORKLOAD total	0.0541	0.1389	0.1263	2.91x	2.34x
cold/STEADY workload	0.0040	0.0360	0.0169	13.35x	4.24x
cold/bounded_file_scan	0.0002	0.0069	0.0008	32.01x	3.52x
cold/open_read_close_loop	0.0012	0.0087	0.0053	12.70x	4.54x
cold/readdir_plus_storm	0.0008	0.0071	0.0036	13.36x	4.56x
cold/readdir_storm	0.0004	0.0057	0.0020	19.47x	4.50x
cold/repeated_read_only_base_open_read_close_loop	0.0004	0.0027	0.0026	9.24x	5.84x
cold/stat_lstat_storm	0.0006	0.0009	0.0012	2.12x	2.02x
cold/tree_discovery	0.0003	0.0039	0.0013	16.77x	4.78x
warm/STARTUP+WORKLOAD total	0.0380	0.1340	0.1151	2.51x	3.03x
warm/STEADY workload	0.0031	0.0220	0.0116	7.76x	3.79x
warm/bounded_file_scan	0.0003	0.0016	0.0008	7.88x	3.13x
warm/open_read_close_loop	0.0008	0.0065	0.0040	8.21x	5.18x
warm/readdir_plus_storm	0.0006	0.0043	0.0017	7.93x	3.03x
warm/readdir_storm	0.0003	0.0028	0.0015	9.36x	4.86x
warm/repeated_read_only_base_open_read_close_loop	0.0003	0.0030	0.0019	10.16x	5.53x
warm/stat_lstat_storm	0.0005	0.0010	0.0005	2.17x	0.97x
warm/tree_discovery	0.0003	0.0029	0.0011	11.51x	4.50x

---

Copy-on-write detail (large-edit-benchmark.py)

50 MiB base file, single-byte edit at file midpoint, then re-read+compare for correctness.

Metric	Native	Original	Tier One
Edit wall time (s)	0.1226	0.5015	0.6650
Wall ratio vs native	1.00x	8.19x	5.42x
Delta DB growth (MiB)	n/a	172.59	50.41
Correctness (outputs match)	n/a	True	True

---

Mixed git-workload detail (git-workload-benchmark-multi.py)

openai/codex (4 643 files, 690 dirs, 63 MiB) bare→working clone, status,
32-file ls-files scan w/ 4 KiB reads, 4 representative edits w/ fsync, diff.
3 measurement iterations + 1 warmup. Medians shown.

Phase	Native (s)	Original (s)	Tier One (s)	Orig	Tier One
checkout	0.1692	0.1725	0.1498	1.07x	0.88x
clone	0.2756	2.3499	2.2126	7.03x	7.65x
diff	0.0226	0.0346	0.1316	2.24x	1.72x
edit	0.0004	0.0013	0.0028	2.38x	6.43x
fsck	0.0000	0.0000	0.0000	0.00x	0.00x
read_search	0.0046	0.0077	0.0097	1.40x	2.11x
status	0.0967	0.2977	0.1646	12.51x	1.70x

---

Per-iteration reproducibility — mixed workload

iter	wall_s (orig)	wall_s (tier1)
1	6.48	9.71
2	4.01	8.61
3	3.59	10.63

Tier One mixed-workload stdev: 0.85x
Original mixed-workload stdev: 1.21x

---

Tier Two focus areas (from this comparison)

1. Clone phase still dominates the mixed wall (~2.2 s of the ~2.9 s
   agentfs median). Native does the same clone in ~0.28 s. The clone phase
   does many small writes through copy-on-write to the SQLite delta;
   batched write path and/or parallel git-pack creation under copy-on-write
   is the next big lever. (median ratio: orig 7.03x → tier1 7.65x; the
   variance is within noise on a 0.28 s native baseline.)

2. Mount startup regressed by ~10-15 ms (read-heavy full-run ratio went
   2.51x → 3.03x — going up) because Tier One mounts now negotiate
   parallel workers + readdirplus + writeback + ABI 7.31 caps at FUSE init.
   For short-lived sandboxes this dominates total wall; for sustained
   workloads it is amortised, which is why steady-state read-storms dropped
   from 7.76x → 3.79x in the same comparison. Tier Two should defer worker
   pool warmup to first request to recover that startup cost.

3. Copy-on-write DB growth is now great (-71%) but the wall-time ratio
   (5.42x) is still the worst-of-three. Chunked copy-up + smarter chunk
   sizing is the obvious next win and would compound with Tier Two prep: fresh native vs origin/main vs Tier One benchmark comparison #1 since
   git-clone bottlenecks on the same path.

4. Steady-state read storms are near best-case: stat_lstat_storm
   (warm) is 0.97x — actually faster than native — because the kernel
   attribute cache absorbs everything past the first lookup. Future
   read-path tuning is diminishing returns; the Tier Two budget should go
   to CoW writes and clone-phase batching.

5. Behaviour to verify before Tier Two ships: the per-iteration
   variance on the mixed workload is high (Tier One stdev 0.85x, Original
   1.21x). A longer-iter (e.g. 10 + 2 warmup) run on a quiet machine would
   tighten the medians; current 3-iter medians are reliable directional
   signal but not paper-grade absolutes.