Machine: Apple M3 Max (Mac15,8), 16 cores, 64 GB · macOS 26.5.1 · NVMe (F_BARRIERFSYNC)
Branch / commit:v1/phase-3-performance @ f13748d (post-Phase-4: weir-ctl, weir-sink-sdk, full DST harness)
Server config:shard_count=4, batch_size=64 (the load suite's bench_preset)
Method:tests/load.rs, release build, 3 serial trials per deadline (--test-threads=1, averaged), at batch_deadline 1 ms and 2 ms. Rendered with deploy/avg_benchmarks.py.
A point-in-time, single-machine snapshot for tracking regressions over time.
Compare like-for-like: this is the Mac dev box (fast F_BARRIERFSYNC).
The CI baseline (latest.md) is a Linux GitHub runner; phase3-results.md
has the slow-storage reference (beast, SATA SSD, fdatasync ~1.4 ms). New
snapshots: docs/benchmarks/snapshot-YYYY-MM-DD-<machine>.md.
The durable path is fsync-bound (the Phase 3 finding still holds). A single Sync record is ~133 µs p50 on this NVMe — almost all of it F_BARRIERFSYNC. That figure is not comparable to a Linux one:F_BARRIERFSYNC orders writes without forcing the drive's volatile cache to the medium, so this box is buying a weaker guarantee rather than performing the same operation faster, and macOS is not power-loss safe at any tier. See Platform support. Batched is statistically identical (shared group-fsync code path). Buffered (memory-only ack) is the fast path at ~26 µs p50.
Group-fsync amortization scales the Sync tier under load: single-thread Sync is ~6.1 k RPS, but the Sync saturation ramp reaches ~80 k RPS @ 64 threads as concurrent records coalesce under one fsync.
Buffered peaks at ~123 k RPS @ 48 threads (the connection cap), holding flat through saturation.
IOPS compression 249:1 — 4,980 records committed in 20 sink commits (the WAB's batch-to-sink amortization).
No perf regression from the Phase 4 / DST work. These numbers match or slightly beat the Phase 3 Mac baseline, so the injectable seams (Arc<dyn> SegmentStore, generic BlockingClock) and the 4.1b flush_batch ack restructure cost nothing measurable on the hot path. (The production binary contains zero sim code.)
The Mac's fast F_BARRIERFSYNC makes its Sync latency ~10× lower than beast's honest SATA fdatasync — beast remains the realistic "slow storage" durability reference. Within the Mac series, this is the number to beat for future runs.
# 3 serial trials per deadline, averaged (matches this snapshot):
SNAP=/tmp/weir_snapshot.jsonl; : > "$SNAP"
for d in 1 2; do for i in 1 2 3; do
WEIR_BENCH_DEADLINE=$d cargo test --release -p weir-server --test load -- \
--nocapture --test-threads=1 2>/dev/null | grep -oE 'BENCH: \{[^}]*\}' >> "$SNAP"
done; done
python3 deploy/avg_benchmarks.py "$SNAP" /tmp/rendered.md && cat /tmp/rendered.md
Note: use grep -oE 'BENCH: \{[^}]*\}' (not grep '^BENCH:') with
--test-threads=1 — serial --nocapture inlines single-line BENCH output
after the test <name> ... prefix, so an anchored grep silently drops every
non-ramp scenario.