Repository navigation
Vfs 1.1.1 performance record: chaos scoreboard, clone, mount startup #10
Description
Vfs 1.1.1 performance ledger
The full six-actor chaos workload is the primary scoreboard. Clone and mount startup are focused diagnostics. The agentfs 0.6.4 result is a separate pre-fork reference because it cannot complete the current workload.
v1.1.1 · 4852148 · main CI · release CI
Measurement contract
- Canonical codex fixture:
7d47056ea42636271ac020b86347fbbef49490aa - Chaos: seeds
20260808and20260809, 6 actors, 8 operations each, 48 operations per leg - One discarded warmup per chaos invocation; native/Vfs first-leg order alternates
- Fresh checkout or destination, HOME, TMPDIR, XDG directories, Git config, and Vfs session store per leg
- Profiling disabled for timings; profile runs used only for transport attestation
- Vfs chaos legs require actor read-back, strict fsck, base immutability, integrity, and clean teardown
Chaos absolute_wall_seconds begins after nested checkout preparation and identical cache warmup. Clone, mount startup, verification, integrity, and teardown are outside that timer.
Chaos scoreboard
Lower is better. Absolute workload wall-clock distributions:
| Scenario | n | median | p25 | p75 | min | max | stdev |
|---|---|---|---|---|---|---|---|
| Native controls, pooled | 40 | 0.806s | 0.746 | 0.855 | 0.642 | 1.129 | 0.097 |
| Vfs 1.1.1, legacy FUSE | 20 | 1.568s | 1.511 | 1.725 | 1.343 | 1.960 | 0.162 |
| Vfs 1.1.1, io_uring | 20 | 1.595s | 1.532 | 1.727 | 1.381 | 2.972 | 0.451 |
Paired native medians were 0.810s for legacy and 0.802s for io_uring. Absolute median gaps were +0.757s and +0.793s. Legacy’s Vfs median was 27ms lower than io_uring’s; the distributions overlap. All 80 measured legs passed, covering 3,840 operations.
| Transport | Seed | Native n=10 | Vfs n=10 |
|---|---|---|---|
| Legacy | 20260808 | 0.824s | 1.664s |
| Legacy | 20260809 | 0.790s | 1.520s |
| io_uring | 20260808 | 0.928s | 1.654s |
| io_uring | 20260809 | 0.749s | 1.559s |
Transport profile runs:
| Requested transport | Kernel enable_uring |
fuse_uring_requests |
|---|---|---|
| Legacy | Y |
0 |
| io_uring | Y |
81,603 |
Clone
Timed commands use one prepared local mirror: native git clone --no-hardlinks; Vfs vfs clone. Mirror creation and correctness checks are outside the timer. Every sample passed clean status, strict fsck, tracked-content comparison, and teardown.
| Scenario | n | median | p25 | p75 | min | max | stdev |
|---|---|---|---|---|---|---|---|
| Native clone | 10 | 0.378s | 0.346 | 0.395 | 0.311 | 0.412 | 0.034 |
| Vfs clone | 10 | 1.080s | 1.031 | 1.108 | 0.963 | 1.214 | 0.074 |
Median Vfs clone overhead: 702ms.
Mount startup
Process spawn through the first successful stat of one base file. The native row is the common python -S process/interpreter floor. Two warmups were discarded; each row has 30 measured samples.
| Scenario | n | median | p25 | p75 | min | max | stdev |
|---|---|---|---|---|---|---|---|
| Native process + probe | 30 | 9.42ms | 7.70 | 10.07 | 6.54 | 24.13 | 3.27 |
| Vfs legacy process + mount + probe | 30 | 28.90ms | 26.74 | 29.91 | 25.02 | 41.54 | 3.79 |
| Vfs io_uring process + mount + probe | 30 | 29.99ms | 27.82 | 33.83 | 21.57 | 45.72 | 6.04 |
Median overhead above the native floor: 19.48ms legacy, 20.57ms io_uring. The 1.09ms transport difference is below dispersion. Profile runs verified fuse_uring_requests = 0 for legacy and 6 for io_uring.
Pre-fork agentfs reference
agentfs 0.6.4 fails the current six-actor workload at 40/48 operations. Its timing reference therefore disables the churner and uses a separate 40-operation plan.
| Scenario | n | median | p25 | p75 | min | max | stdev |
|---|---|---|---|---|---|---|---|
| Native reference | 30 | 0.641s | 0.601 | 0.709 | 0.504 | 1.001 | 0.098 |
| agentfs 0.6.4, legacy FUSE | 10 | 2.712s | 2.173 | 2.995 | 2.098 | 4.063 | 0.673 |
agentfs passed 0/5 full-workload samples; Vfs 1.1.1 passed 20/20 per transport.
These benchmarks are serialized local release measurements, not CI timing gates.
Activity
[-]Perf backlog at 1.0.0: trustworthy medians, then clone-phase and mount startup[/-][+]Performance backlog after 1.0.1: legacy read path, clone, mount startup[/+][-]Performance backlog after 1.0.1: legacy read path, clone, mount startup[/-][+]Vfs 1.0.2 performance record: chaos scoreboard, clone, mount startup[/+][-]Vfs 1.0.2 performance record: chaos scoreboard, clone, mount startup[/-][+]Vfs 1.1.0 performance record: chaos scoreboard, clone, mount startup[/+][-]Vfs 1.1.0 performance record: chaos scoreboard, clone, mount startup[/-][+]Vfs 1.1.1 performance record: chaos scoreboard, clone, mount startup[/+]