docs(bench): record blob-prefetch + tokio-runtime benchmark results
Companion benches/*.md (matching the repo convention) capturing the before/after numbers and the honest interpretation behind the read_prefetch 1->2 tuning and the runtime pool sizing. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JG5yYZ9s868mJwqT2Qz7ez
This commit is contained in:
Claude
@@ -0,0 +1,53 @@
|
||||
# Blob download read-ahead benchmark
|
||||
|
||||
Measures the local backend's chunk read-ahead depth — the `buffered(N)`
|
||||
read-ahead in `DedupService::stream_chunks` fed by
|
||||
`BlobStorageBackend::read_prefetch()`. Reproduces the exact production reassembly
|
||||
combinator (`stream::iter(hashes).map(get_blob_stream).buffered(N).try_flatten()`)
|
||||
over a real `LocalBlobBackend` whose chunk files are scattered across the 256
|
||||
hash-prefix dirs, then drains it and reports throughput. `N = 1` is the old
|
||||
production default ("antes"); higher N is the change ("después").
|
||||
|
||||
## Reproduce
|
||||
|
||||
```bash
|
||||
cargo run --release --features bench --example bench_blob_prefetch
|
||||
# tunables: BENCH_FILE_MB=192 BENCH_CHUNK_KB=256 BENCH_PREFETCH=1,2,4,8,16
|
||||
# BENCH_THROTTLE_MBPS=0,300,100 BENCH_REPS=5 BENCH_COLD=1
|
||||
```
|
||||
|
||||
## Results (4-core box, SSD-class storage, 192 MiB in 768×256 KiB chunks)
|
||||
|
||||
Median MB/s over 5 reps; `vs N=1` is the read-ahead gain over the old default.
|
||||
|
||||
| scenario | N=1 | N=2 | N=4 | N=8 | N=16 |
|
||||
|-------------------------|----:|-----:|-----:|-----:|-----:|
|
||||
| warm / unthrottled |1306 | **1460** |1394 |1357 |1248 |
|
||||
| cold / unthrottled | 456 | **489** | 469 | 478 | 473 |
|
||||
| warm / throttled@300MB/s| 167 | 166 | 166 | 166 | 166 |
|
||||
| cold / throttled@300MB/s| 138 | 140 | 135 | 135 | 136 |
|
||||
| warm / throttled@100MB/s| 62 | 62 | 62 | 62 | 62 |
|
||||
| cold / throttled@100MB/s| 57 | 58 | 57 | 57 | 57 |
|
||||
|
||||
(`vs N=1` for the best column N=2: warm/unthrottled **+11.8 %**, cold/unthrottled
|
||||
**+7.2 %**; throttled rows ≈ 0 %. N=16 regresses warm −4.4 %.)
|
||||
|
||||
## Conclusions
|
||||
|
||||
1. **N=2 is the sweet spot, not 8.** It wins or ties in 5 of 6 scenarios at the
|
||||
lowest fan-out: +11.8 % warm and +7.2 % cold on disk-bound reads, neutral when
|
||||
the consumer is the bottleneck. N=8 gives only +3.9 %/+4.8 %; N=16 regresses.
|
||||
So local now defaults to 2 (was 1); S3/Azure keep 8 (request-latency bound).
|
||||
|
||||
2. **The win is disk-bound, not network-bound.** Throttled (network-bound) rows
|
||||
are flat because `buffered(N)` here overlaps the per-chunk `File::open`
|
||||
(cheap on local disk), **not** the data read (which `try_flatten` polls
|
||||
sequentially). The disk-bound rows cover localhost/LAN downloads *and* the
|
||||
internal blob reads that drain as fast as the disk delivers — thumbnail
|
||||
render, transcode, ZIP export, content extraction — all via `stream_chunks`.
|
||||
|
||||
3. **No cold regression on SSD.** The trait doc's "slower cold" worry (concurrent
|
||||
opens → random I/O over scattered chunk files) is an HDD seek-thrash concern;
|
||||
on SSD-class storage cold reads *improved* at N=2. Operators on spinning disks
|
||||
can restore the old behaviour with `OXICLOUD_LOCAL_READ_PREFETCH=1`; NVMe
|
||||
arrays can raise it.
|
||||
@@ -0,0 +1,93 @@
|
||||
# Tokio runtime tuning benchmark
|
||||
|
||||
Measures the two things `build_runtime` (`src/main.rs`) changes versus the bare
|
||||
`#[tokio::main]` defaults, sized by `common::runtime::runtime_pool_sizes`:
|
||||
|
||||
- **Worker count.** `#[tokio::main]` defaults to `available_parallelism()`, which
|
||||
honours CPU *affinity* (`sched_getaffinity`: cpuset, `taskset`) but **ignores
|
||||
the CFS bandwidth quota** (`docker --cpus`, cgroup v2 `cpu.max`, v1
|
||||
`cpu.cfs_quota_us`). On a 2-core-quota container on a many-core host it spawns
|
||||
one worker per *host* core. `effective_parallelism()` folds the quota back in.
|
||||
- **Blocking pool.** `#[tokio::main]` defaults to a flat `max_blocking_threads =
|
||||
512` — a multi-GB RSS blast radius for this heavy `spawn_blocking` user
|
||||
(thumbnails, transcode, zip, PDF/text extraction, Argon2 ≈19 MB/hash). The
|
||||
builder caps it at `max(32, 8 × workers)`.
|
||||
|
||||
## Reproduce
|
||||
|
||||
```bash
|
||||
cargo build --release --features bench --example bench_tokio_runtime
|
||||
# Pin to 2 cores to model a 2-core CPU quota on a bigger host:
|
||||
taskset -c 0,1 ./target/release/examples/bench_tokio_runtime
|
||||
# Part B uses a fixed glibc mmap threshold for a clean RSS read:
|
||||
MALLOC_MMAP_THRESHOLD_=131072 MALLOC_TRIM_THRESHOLD_=131072 \
|
||||
taskset -c 0,1 ./target/release/examples/bench_tokio_runtime
|
||||
# tunables: BENCH_CONCURRENCY=96 BENCH_SECONDS=4 BENCH_BURN_KB=256
|
||||
# BENCH_WORKERS_BEFORE=32 BENCH_BLOCKING_TASKS=96 BENCH_ALLOC_MB=16 BENCH_MAX_BLOCKING_AFTER=16
|
||||
```
|
||||
|
||||
## Results (4-core box, pinned to 2 cores via `taskset -c 0,1`)
|
||||
|
||||
### [A] Worker over-subscription under CPU contention
|
||||
|
||||
96 concurrent async "requests", each an async hop + a 256 KiB BLAKE3 (models a
|
||||
handler that interleaves I/O with on-worker compute), over 4 s.
|
||||
|
||||
| runtime | req/s | p50 µs | p99 µs |
|
||||
|-----------------------|-------:|-------:|-------:|
|
||||
| before: 32 workers | 46 854 | 121 | 60 360 |
|
||||
| after: 2 workers | 42 893 | 2 140 | 4 962 |
|
||||
|
||||
→ **throughput −8.5 %, p99 latency −91.8 %** (after vs before)
|
||||
|
||||
### [B] Blocking-pool RSS blast radius
|
||||
|
||||
96 concurrent `spawn_blocking` tasks, 16 MiB resident each, held 120 ms
|
||||
(fixed glibc mmap threshold so freed allocations leave RSS promptly).
|
||||
|
||||
| max_blocking_threads | peak RSS MiB | vs default |
|
||||
|-----------------------------|-------------:|-----------:|
|
||||
| before: 512 (tokio default) | 1 231 | — |
|
||||
| after: 16 (bounded) | 261 | −970 MiB |
|
||||
|
||||
## Conclusions
|
||||
|
||||
1. **Blocking-pool cap — clear win, no downside.** Bounding 512→16 cut peak RSS
|
||||
under a 96-task flood from **1231 MiB to 261 MiB (−970 MiB)**. The cap only
|
||||
engages under a pile-up; steady-state operation is unaffected, and the app's
|
||||
heaviest blocking consumers are already semaphore-limited (Argon2 = 2,
|
||||
thumbnail decode ≈ cores), so `max(32, 8×workers)` is generous headroom that
|
||||
simply removes the unbounded tail that can OOM-kill the process under a spike.
|
||||
|
||||
2. **Worker sizing — a latency/throughput trade, favourable for a server.**
|
||||
Over-subscription (32 workers on 2 cores, what tokio's default does under a
|
||||
CFS quota) won **+8.5 % peak throughput** but at a **catastrophic p99 of
|
||||
60 ms** (12× the tuned 5 ms) with a bimodal distribution — some requests fly
|
||||
(p50 121 µs), others starve. Sizing to the quota (2 workers) gives uniform,
|
||||
predictable latency at a small throughput cost. For an interactive file
|
||||
server, p99 dominates UX (timeouts, head-of-line blocking), so this is the
|
||||
right trade.
|
||||
|
||||
3. **This microbenchmark is a worst case *for* the tuned config.** It is pure
|
||||
on-worker CPU, which is exactly where over-subscription's throughput edge
|
||||
shows. Real OxiCloud handlers push CPU to `spawn_blocking` and the async
|
||||
workers mostly await I/O (DB, disk) — there the over-subscription throughput
|
||||
edge evaporates (idle workers just park) while its tail-latency penalty
|
||||
remains. Production should see the worker change as ≥ neutral on throughput
|
||||
and strictly better on tail latency.
|
||||
|
||||
4. **No regression off-quota.** `effective_parallelism()` == `available_
|
||||
parallelism()` whenever there is no CFS quota (or affinity already restricts
|
||||
the process), so on bare metal / affinity-pinned deployments the worker count
|
||||
is unchanged from the old default. The change only bites under a CFS quota —
|
||||
precisely the case it fixes.
|
||||
|
||||
5. **Follow-up:** the same `available_parallelism()` blind spot affects the
|
||||
image/rayon pools (`thumbnail_service.rs`, `image_transcode_service.rs`,
|
||||
`di.rs` video) — they over-spawn under a CFS quota too. Switching those to
|
||||
`common::runtime::effective_parallelism()` is the natural next step (left out
|
||||
here to keep this change focused on the runtime).
|
||||
|
||||
Both knobs are env-overridable (`OXICLOUD_WORKER_THREADS` /
|
||||
`OXICLOUD_MAX_BLOCKING_THREADS`) and logged at startup ("Tokio runtime pools
|
||||
sized"), so operators can see and tune what is in effect.
|
||||
Reference in New Issue
Block a user