Benchmark Snapshot — 2026-06-13 (beast, Linux SATA SSD)

Machine: beast — Intel Core i9-9900K (8c/16t) · 31 GiB · Linux 6.17.0 · ext4 on a Samsung SATA SSD (/dev/sdc2, fdatasync) Branch / commit: v1/phase-3-performance @ c46c0c7 (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.

The honest slow-storage reference. beast's SATA fdatasync (~1.5 ms, no power-loss-protected write cache) is the realistic durability floor — ~10× the Mac dev box's F_BARRIERFSYNC. Compare like-for-like (beast-to-beast for future runs). Cross-environment analysis: snapshot-2026-06-13-comparison.md.

TL;DR

The durable path is fsync-bound — and on honest storage it dominates everything. A single Sync record is 1.5 ms p50, of which ~99% is fdatasync. Single-thread Sync is just 653 RPS.
Batched is identical to Sync (1.5 ms p50) — same group-fsync code path.
Buffered (memory-only ack) is ~80× faster than Sync here — 19 µs p50, 49,725 RPS single-thread — because it never fsyncs. On this box the fast path actually beats the Mac.
Group-fsync amortization is the only thing that scales Sync: single-thread 653 RPS → ~6.8k RPS @ 96 threads as concurrent records coalesce under one fsync (~10× via batching), but the ~1.5 ms fsync floor caps it hard.
Buffered scales to ~155k RPS @ 48 threads (the connection cap).
No regression from the Phase 4 / DST work — these match the Phase 3 beast baseline (single-thread Sync 653 vs 647; Buffered ramp 155k vs 157k).

Throughput — deadline comparison

Scenario	RPS (1ms)	±σ (1ms)	RPS (2ms)	±σ (2ms)	d1ms/d2ms
single_thread_buffered	49,725	±3,752	45,745	±2,727	1.09×
single_thread_sync	653	±19	617	±0	1.06×
thundering_herd_8_threads	1,258	±162	1,255	±235	1.00×
thundering_herd_32_threads	4,445	±507	4,367	±945	1.02×
thundering_herd_64_threads	9,296	±1,416	8,264	±1,586	1.12×
connection_churn	14,394	±967	14,916	±39	0.96×
fire_and_forget_overload	3,899	±8	3,942	±144	0.99×

Latency — single thread, `batch_deadline_ms=1`

Metric	Sync	Batched	Buffered
Min	1.1 ms	1.1 ms	16 µs
Mean	1.6 ms	1.6 ms	19 µs
σ	532 µs	532 µs	6 µs
p50	1.5 ms	1.5 ms	19 µs
p75	1.5 ms	1.6 ms	19 µs
p95	1.7 ms	1.7 ms	22 µs
p99	4.2 ms	4.2 ms	34 µs
p99.9	6.0 ms	6.3 ms	151 µs
Max	7.2 ms	7.0 ms	188 µs

Latency — single thread, `batch_deadline_ms=2`

Metric	Sync	Batched	Buffered
Min	1.2 ms	1.2 ms	16 µs
Mean	1.6 ms	1.6 ms	18 µs
σ	518 µs	515 µs	7 µs
p50	1.5 ms	1.5 ms	17 µs
p75	1.6 ms	1.6 ms	18 µs
p95	1.7 ms	1.7 ms	22 µs
p99	4.2 ms	4.2 ms	40 µs
p99.9	6.1 ms	5.4 ms	124 µs
Max	6.8 ms	5.5 ms	279 µs

Saturation Ramp — Buffered tier

Server started with max_connections = 48. Levels above 48 trigger connection-cap exhaustion; the server must survive every level.

Threads	RPS (d1ms)	RPS (d2ms)	I/O drops	Status
8	30,864	32,579	0	ok
16	62,401	65,178	0	ok
32	116,550	118,661	0	ok
48	155,577	157,044	0	ok
64	149,343	149,293	16	SATURATED ←
96	151,028	154,255	48	SATURATED

Saturation Ramp — Sync tier

Uses Sync durability to stress the group-fsync path under escalating concurrency. The ~1.5 ms fdatasync floor caps throughput an order of magnitude below the Mac.

Threads	RPS (d1ms)	RPS (d2ms)	I/O drops	Status
8	1,199	1,007	0	ok
16	2,282	1,954	0	ok
32	4,399	4,190	0	ok
48	6,178	5,381	0	ok
64	6,618	5,785	16	SATURATED ←
96	6,856	6,155	48	SATURATED

IOPS compression

249:1 — 4,980 records committed in 20 sink commits (content-derived, deterministic; identical to the Mac run).

Reproduce

Same as the Mac snapshot, run on beast (git checkout c46c0c7 first; cargo is under ~/.cargo/bin so . ~/.cargo/env in non-interactive shells):

. ~/.cargo/env; cd /root/weir
SNAP=/tmp/weir_beast.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/r.md && cat /tmp/r.md