strawmANN benchmark report

strawmANN against Qdrant · AMD Ryzen AI 9 HX PRO 370 w/ Radeon 890M (24 logical) · generated 2026-10-03 11:42:56 CEST

Dataset: laion-small-clip
Vectors: 100,000 × 512 · 5,000 held-out queries
Metric: cosine
Storage datatype: strawmANN float32 · Qdrant float32 (default)
Quantization: none (fp32) · separate rows measure binary (1 bit/dim), PQ (product), SQ8 scalar, turboquant

strawmANN: 55/55 ok Qdrant: 55/55 ok

Summary

At equal recall, strawmANN serves 1.53x to 1.81x Qdrant's throughput. That range spans the recall levels measured; allowing for the uncertainty in each recall figure widens it to 0.86x to 3.03x.

throughput at equal recall

1.53 to 1.81x

strawmANN over Qdrant, recall@10 0.963 to 0.998

p99 latency at 90% load

strawmANN2.14 ms

Qdrant4.91 ms

W4-sat90, open loop, one offered rate

upload and index build

strawmANN9 s

Qdrant14 s

W1 + W2

peak memory (RSS)

strawmANN1.5 GiB

Qdrant3.1 GiB

largest before the concurrent-write rows (W11); of it, anonymous 315.8 MiB / 288.5 MiB

written to disk

strawmANN2.3 GiB

Qdrant10.1 GiB

before the concurrent-write rows (W11); they wrote 49.3 MiB / 9.5 GiB more

Throughput against recall

Read it vertically: at any recall both engines reach, the higher curve is faster. Throughput from bfb W10, recall from the conformance sweep, joined on ef. ef is the same unit here: Qdrant held 2 segments for this run, one populated and 1 empty (read back per segment from the engine), and strawmANN held one graph, so equal ef is equal traversal width and the curves may be read at equal x. That is a property of this run, not of the engines.

Every figure is the median of 3 passes per engine, run alternately (A/B/A/B). What this does not establish.

Glossary

recall@10: Of the ten nearest neighbours a query really has, the share the engine returned. 1.0 is a perfect answer. "Really has" is settled by an exhaustive fp64 search, not by the other engine.
ef: The candidate-list size: how many nodes the index keeps in play while searching. Larger is slower and more accurate. It does not mean the same amount of work in both engines, which is why the headline holds recall fixed instead.
MRDE: Mean relative distance error: when the engine returns a wrong neighbour, how much further away it is than the right one. Small numbers mean the misses were near-misses.
qps: Queries per second. On a batched row one request carries several queries, and the request rate is shown under it.
p50 / p99 / p99.9: The latency half of requests beat, that 99% beat, that 99.9% beat. The tail is what a user notices.
closed / open loop: A closed loop sends the next request only when the last one comes back, so a slow server receives less work and its tail looks better than it is. An open loop sends at a fixed rate regardless.
noise floor: How much a number moves between identical runs on this machine. A ratio inside it is shown grey with ≈: no measured difference. Hover a ratio to see its band.
conformance tier: What the two engines were shown to agree on before any speed was quoted. T1 licenses a single engine's own numbers; T3 licenses comparing the two, and requires their recall to be statistically indistinguishable.
§ numbers: Sections of docs/spec.md, the written rule each claim in the appendix is measured against. §7.1 is the host gate, §7.4 the comparison rules, §8 what may be published.

Throughput

Queries per second; higher is better. The ratio is strawmANN over Qdrant: green is faster, red slower, grey ≈ inside the noise floor (hover a ratio for its band). A dash means the pair is not compared, and the note says why. An ef sweep is one row showing its range; every measurement is under All throughput rows.

workload		strawmANN	Qdrant	ratio	notes
W3	search, fp32, single query	1,646	1,637	parity
W4	search, saturating (closed loop)	10,800	10,924	parity
W5	search batched (16 distinct dataset queries per request)	10,643664 requests/s	3,754234 requests/s	2.83x
W6	quantized: scalar	5,507	4,443	1.24x
W6 ef 32 to 512	SQ8 recall control (latency only)	2,072 to 10,847	1,858 to 7,502	1.13 to 1.41x	2 of 5 points not compared recall differs one pass apart: Qdrant pass 3 -21%
W7	quantized: binary + oversampling	5,769	4,616	1.25x
W7-2bit	quantized: binary quantization, 2 bits + oversampling	5,465	4,636	1.18x
W7-1p5bit	quantized: binary quantization, 1.5 bits + oversampling	5,525	4,538	1.22x
W8	quantized: PQ	3,440	2,888	1.19x
W14-1bit	quantized: TurboQuant, 1 bit + oversampling	4,612	4,570	parity
W14-1p5bit	quantized: TurboQuant, 1.5 bits + oversampling	4,356	4,090	1.07x
W14-2bit	quantized: TurboQuant, 2 bits + oversampling	5,464	4,596	1.19x
W14-4bit	quantized: TurboQuant, 4 bits + oversampling	5,086	4,596	1.11x
W9	exact / brute force	311	311	parity
W10 ef 32 to 512	recall control (latency only)	3,423 to 27,901	3,373 to 18,065	1.12 to 1.42x	1 of 5 points at parity 1 of 5 points not compared recall differs
W12-sel1	filtered search, one keyword (~1% of bench12)	7,096	3,368	-	one pass apart: strawmANN pass 1 -16%
W12-sel10	filtered search, any of 10 keywords (~10% of bench12)	1,322	1,803	-	recall differs: 1.0000 vs 0.9854
W12-sel1 ef 32 to 512	filtered recall control, one keyword (latency only)	6,833 to 7,312	1,693 to 5,774	1.26 to 4.04x	1 of 5 points not compared one pass apart: strawmANN pass 1 -39%
W12-sel10 ef 32 to 512	filtered recall control, any of 10 keywords (latency only)	1,319 to 3,212	612 to 3,980	1.29 to 2.16x	3 of 5 points not compared recall differs
W13	scroll / pagination	50,282	46,273	1.09x	strawmANN: 92% is client and socket, not server Qdrant: 82% is client and socket, not server
W11-steady	mixed read/write below the rebuild threshold: search bench2 while 5,000 synthetic points append	10,102	7,170	-	search-during-write; no recall join
W11	mixed read/write: search bench2 while 20,000 synthetic points append (runs last)	9,526	7,151	-	search-during-write; no recall join

Throughput by workload

Higher is better. Hatched bars are rows the table does not compare (W11): each bar is that engine's own rate, and the pair is not a result.

Recall

recall@10 against an exact fp64 search, not against the other engine. Equal ef is not equal work in the two engines, so the comparison that counts holds recall fixed and compares throughput there.

At matched recall

recall@10	measured at	strawmANN q/s	Qdrant q/s	strawmANN / Qdrant	range over recall CI
0.9630	strawmANN at ef=32, Qdrant interpolated	27,901	15,391	1.81x	1.66 – 1.98x
0.9786	Qdrant at ef=64, strawmANN interpolated	20,242	12,438	1.63x	1.48 – 1.79x
0.9852	strawmANN at ef=64, Qdrant interpolated	17,667	9,925	1.78x	1.57 – 2.01x
0.9906	Qdrant at ef=128, strawmANN interpolated	12,840	8,268	1.55x	1.28 – 1.88x
0.9940	strawmANN at ef=128, Qdrant interpolated	10,487	6,166	1.70x	1.39 – 2.09x
0.9956	Qdrant at ef=256, strawmANN interpolated	8,237	5,380	1.53x	1.12 – 2.10x
0.9976	strawmANN at ef=256, Qdrant interpolated	6,012	3,528	1.70x	1.25 – 2.32x
0.9979	Qdrant at ef=512, strawmANN interpolated	5,431	3,373	1.61x	0.86 – 3.03x

How this is computed

Interpolated linear in log(q/s) between the two bracketing measurements, nothing extrapolated past the measured range. Throughput is bfb's W10 sweep, recall the conformance sweep, joined on ef — same m and ef_construct, but separate builds, so the pairing assumes two builds with identical parameters are equivalent.

The last column is why the ratio is not a result on its own. The ratio treats the anchor's recall as exact; it is an estimate, and moving it across its 95% interval moves the interpolated rate with it — near the top of the sweep 0.003 of recall spans a factor of 1.8, so two decimals there quote the interpolation. Even that is the narrow reading: it moves one recall and not the other, adjacent anchors share bracketing segments, and the throughputs behind it are medians of the run's passes.

At matched recall, filtered to 10%

The same reading under a keyword filter over the points the condition matched, 9,943 to 10,109 across 3 builds in strawmANN and 9,917 to 10,066 across 3 builds in Qdrant (bfb draws the keyword payloads unseeded at each upload). The per-row table refuses W12-sel10 a ratio because the two engines land just outside the recall band at the one ef it measures; held at equal recall instead, the comparison exists at every recall both engines reach. Note the shape rather than any single number: one engine's curve is flat in ef and the other's is steep, so where you match decides the ratio.

recall@10	measured at	strawmANN q/s	Qdrant q/s	strawmANN / Qdrant	range over recall CI
0.9808	strawmANN at ef=32, Qdrant interpolated	3,212	1,869	1.72x	1.65 – 1.78x
0.9854	Qdrant at ef=128, strawmANN interpolated	2,736	1,800	1.52x	1.33 – 1.73x
0.9932	strawmANN at ef=64, Qdrant interpolated	2,077	1,239	1.68x	1.49 – 1.89x
0.9972	Qdrant at ef=256, strawmANN interpolated	1,588	1,024	1.55x	1.39 – 1.73x
0.9993	Qdrant at ef=512, strawmANN interpolated	1,379	612	2.25x	2.13 – 2.38x

How this is computed

Interpolated linear in log(q/s) between the two bracketing measurements, nothing extrapolated past the measured range. Throughput is bfb's W12-sel10 sweep, recall the conformance sweep, joined on ef — same m and ef_construct, but separate builds, so the pairing assumes two builds with identical parameters are equivalent.

The last column reads as it does in the fp32 table above.

Recall by `ef`

ef	strawmANN				Qdrant
ef	recall@1	recall@10	recall@100	MRDE	recall@1	recall@10	recall@100	MRDE
ef=32	0.9724	0.9630	-	1.12e-03	0.9676	0.9513	-	1.40e-03
ef=64	0.9872	0.9852	-	3.99e-04	0.9802	0.9786	-	6.50e-04
ef=128	0.9942	0.9940	0.9755	1.51e-04	0.9894	0.9906	0.9677	2.65e-04
ef=256	0.9978	0.9976	0.9912	6.03e-05	0.9946	0.9956	0.9881	1.27e-04
ef=512	0.9990	0.9989	0.9966	2.38e-05	0.9976	0.9979	0.9953	6.26e-05

How recall was measured

5,000 held-out queries, limit 10, ε=9.537e-07, against the fp64 oracle rather than against the other engine. Base checksum 7163721f6ee7759a: the same corpus the latency rows were measured on. Each pass rebuilds the collection. Across 3 independent builds of it, recall@10 at ef=512 spread by: strawmANN 0.00008; Qdrant 0.00006. Over the same builds the engine reported unreachable nodes: strawmANN 0 of 100,000, in-degree zero 4 — a node nothing points at is invisible at any ef, so that is where a graph-quality difference shows rather than being inferred from the recall beside it. Level seed: strawmANN 0x57ea3111. Recorded as provenance: under this engine's level draw, six builds at four seeds sit within 0.00003 of recall@10 at ef 512, well inside the spread above.

What each quantization costs in recall

Colour is the encoding here, not the engine — the engines are the line style. Each sweep is joined on the quantization parameters it actually sent, so a curve speaks only for the search its row ran. Higher is better, and the encodings are not free: read this against the throughput their rows bought.

Latency

Client-side round trip. A closed loop understates the tail (a stalled server stops receiving requests), so this keeps the fixed-rate, open-loop rows, which offer the same load to both engines, plus single-client W3 and batched W5. Every row is under All latency rows.

workload	strawmANN			Qdrant
workload	p50	p99	p99.9	p50	p99	p99.9
W3 search, fp32, single query	574 µs	966 µs	1.07 ms	611 µs	790 µs	860 µs
W4-sat50 search, fixed rate at 50% of saturation (open loop)	696 µs	1.34 ms	1.59 ms	856 µs	1.95 ms	2.39 ms
W4-sat70 search, fixed rate at 70% of saturation (open loop)	861 µs	1.89 ms	2.21 ms	942 µs	2.20 ms	3.30 ms
W4-sat90 search, fixed rate at 90% of saturation (open loop)	913 µs	2.14 ms	4.15 ms	1.33 ms	4.91 ms	8.14 ms
W5 search batched (16 distinct dataset queries per request)	3.00 ms	3.52 ms	3.69 ms	8.50 ms	9.40 ms	9.87 ms

Ingest and index build

Qdrant indexes while it ingests, so its upload time already contains most of the indexing; strawmANN uploads raw and builds afterwards. Compare the sum, not the upload line.

Where the load time goes

Solid is upload, hatched is the wait for Green. The bar's whole length is the sum this section asks you to compare; the split is why the upload line alone is not comparable between these engines.

workload		strawmANN	Qdrant
W0-upload	d=4 floor: graph traversal with the distance taken out	4.01 s	3.01 s †
W1	ingest throughput (no index wait)	0.29 s	1.11 s
W2	index build, time to Green	9.03 s	13.03 s
W6-upload	search, SQ8 scalar quantization	8.03 s	8.03 s
W7-upload	search, binary quantization	5.02 s	8.03 s
W7-2bit-upload	search, binary quantization	6.02 s	8.03 s
W7-1p5bit-upload	search, binary quantization	6.02 s	9.03 s
W8-upload	search, product quantization	45.15 s	45.12 s
W14-1bit-upload	upload	10.03 s	9.03 s
W14-1p5bit-upload	upload	10.03 s	9.03 s
W14-2bit-upload	upload	10.03 s	9.03 s
W14-4bit-upload	upload	10.03 s	9.03 s
W12-upload	upload	10.03 s	21.06 s
W1 + W2	upload and index, together	9.3 s	14.1 s	1.52x

† at the load generator's polling floor. bfb decides a collection is Green by polling once a second and requiring three consecutive Green replies, having slept a second before the first poll, so no Time-to-Green it can report is below 3 s regardless of how fast the build was. A marked cell is an upper bound on the build and a measurement of the polling loop. The bias is a constant added to both engines, so the difference between them survives it and the ratio does not — and it is the faster engine the ratio understates.

Memory and disk

What each engine held in memory and moved to and from disk over the whole run. Per-row figures are under Storage and I/O.

	strawmANN	Qdrant
storage on disk	1.2 GiBallocated at --capacity	1.3 GiB
peak memory (RSS)	1.5 GiB	3.1 GiB
read from disk	0	416.0 MiB
written to disk	2.3 GiB	10.1 GiB

Peak memory and storage on disk are read before the concurrent-write rows (W11): an engine rewriting segments maps old and new files at once, and RSS counts each mapping. Bytes written are summed over the same rows, since W11's volume is set by the harness's write rate; the other disk figures are totals over every row.

Appendix

How the run was set up, every row of every table, and the diagnostics behind them. Charts on a logarithmic axis say so on the axis: read the positions there, not the distances.

Run conditions and conformance

Measured as-deployed: the scheduler was left as it ships, so a difference is what a user would see rather than the engine in isolation, and ambient load is part of the measurement. Qdrant ran equal-work: asked for one populated graph (--segments 1), as strawmANN serves, so ef means the same thing on both sides. This is the configuration §8's comparative licensing is built around, and it is a control rather than a deployment. Its optimizer treats the count as a target; what it actually held is read back from the engine and stated under Collection config.

Conformance tier reached: T4 quantization fidelity. T1 passed, so single-engine performance rows are licensed (§8). T3 passed, so a strawmANN-vs-Qdrant throughput comparison is licensed: the engines are at equal recall (§7.4). Conformance hash c7c43624d600016f.

On exact search the two engines' scores differ by at most 2.980e-07 (p99 1.788e-07), against a calibrated absolute ε of 9.537e-07. §8.1: bit-exactness is unachievable between two different summation orders, so this — equal values within a measured tolerance — is the claim the speed rests on.

The conformance harness's single-client rate agrees on the ordering (2.37x to 2.72x); a different instrument, so no ratio is formed from it.

Every figure is the median of 3 passes per row per engine, alternated A/B/A/B (§7.4), and the spread is this run's own over 42 of 55 rows. The ingest and index-build rows (W0-upload, W1, W2, W6-upload, W7-upload, W7-2bit-upload, W7-1p5bit-upload, W8-upload, W14-1bit-upload, W14-1p5bit-upload, W14-2bit-upload, W14-4bit-upload, W12-upload) have none, so no build time carries a verdict. Each pass rebuilds the index, so the bands include build variance and are wider than a floor measured against one standing graph.

Units are queries per second. bfb reports rps, which counts batch requests: at --search-batch-size 16 the two differ by 16x, and reading one as the other once turned a 2.2x speedup into an apparent 7x regression. The tables show the request rate wherever it diverges.

The open-loop rows are not a speed. W4-sat50/70/90 offer a fixed fraction of measured saturation; serving it means the engine kept up, not that it was faster, so no ratio is printed. They exist because --parallel is a closed loop, where a stalled server stops receiving requests and understates its own tail.

ef is the same unit here: Qdrant held 2 segments for this run, one populated and 1 empty (read back per segment from the engine), and strawmANN held one graph, so equal ef is equal traversal width and the curves may be read at equal x. That is a property of this run, not of the engines.

What was measured builds, dataset, host

	strawmANN `sm-laion-perf-1003`	Qdrant `qd-laion-perf-1003`
engine	commit `f346aea94dbf` `ReleaseFast`, native build, vnni on binary sha256 `aca46d195c7e22bc` built 2026-10-02T14:42:33Z with zig 0.16.0	version 1.19.2-dev binary `~/.cache/strawmann/qdrant-dbeb0f73/qdrant` sha256 `dbeb0f73dea2d371` build `878843e6` (from the server's banner, not a checkout) native binary outside a checkout: sha256 is the identity (§8.9) network `native` no container in the path, like strawmann
measured	2026-10-03T06:30:49Z to 2026-10-03T08:25:46Z (3 passes)	2026-10-03T06:54:01Z to 2026-10-03T08:50:18Z (3 passes)
profile	`as-deployed` · both engines
§7.1 gate	pass · both engines
environment hash	`373e8531fd3d1bf6` · both engines
load generator	bfb dev @ 29240511 (qdrant/bfb#183; carries #172, findings 32's --rps reaping fix) · both engines
client	qdrant-client 1.16.1-dev (git dev branch) · both engines
launched as	`~/Workspace/strawmann/zig-out/bin/strawmann --port 6334 --capacity 125000 --connections 64 --workers 7 --io-threads 1 --pin --cpus 4-11 --data-dir ~/.cache/strawmann/strawmann-storage --default-placement cached`	`~/.cache/strawmann/qdrant-dbeb0f73/qdrant`
cores the engine could use	`4-11` (8 cores) requested by the harness `4-11` (8 cores) observed on 8 threads `0-23` (24 cores) observed on 1 thread	`4-11` (8 cores) requested by the harness `4-11` (8 cores) observed on 52 threads

Dataset

laion-small-clip, 100000 × 512, cosine, 5000 held-out queries

ground truth: shipped, and diffed against our fp64 recompute

1 file, each pinned by sha256 in datasets.json

Host

AMD Ryzen AI 9 HX PRO 370 w/ Radeon 890M

24 logical cores · 58.6 GB · kernel 7.0.0-38-generic

memory bandwidth: 75.7 GB/s aggregate (24 threads) · 45.2 GB/s single core (60% of bus)

strawmann startup banner, 2026-10-03T06:30:49Z; the Qdrant run agrees within 10%

All throughput rows every measurement, with its notes

workload		strawmANN	Qdrant	ratio	notes
W0	d=4 floor: graph traversal with the distance taken out	4,824	4,075	1.18x	1.18x is 1.60x less work per query x 0.66x cores busy during the row (0.66 against 1.00) x 1.00x clock x 1.12x counted on-CPU share: the middle term is occupancy, not search speed d=4 floor: this measures graph traversal and heaps with the distance taken out, not request plumbing (findings 28: 76% of the server's CPU was HNSW search, 24% kernel; the RPC path is a small remainder)
W3	search, fp32, single query	1,646	1,637	parity	within the ±2.3% band this dataset's noise floor puts on W3: no measured difference, not a small one
W4	search, saturating (closed loop)	10,800	10,924	parity	within the ±2.0% band this dataset's noise floor puts on W4: no measured difference, not a small one
W4-sat50	search, fixed rate at 50% of saturation (open loop)	5,489	5,489	offered
W4-sat70	search, fixed rate at 70% of saturation (open loop)	7,685	7,685	offered
W4-sat90	search, fixed rate at 90% of saturation (open loop)	9,879	9,879	offered
W5	search batched (16 distinct dataset queries per request)	10,643664 requests/s	3,754234 requests/s	2.83x	per-batch latency (16 queries/request); queries: dataset, random-sample 2.83x is 0.78x less work per query x 3.58x cores busy during the row (7.04 against 1.97) x 1.00x clock x 1.01x counted on-CPU share: the middle term is occupancy, not search speed batched 16, so qps and rps differ by 16x; the ratio uses qps
W6	quantized: scalar	5,507	4,443	1.24x
W6-ef32	SQ8 recall control, ef=32 (latency only)	10,847	7,502	-	recall unequal: 0.9624 vs 0.9504; §7.4 compares at equal recall; the rescore pools match, so the encoders differ
W6-ef64	SQ8 recall control, ef=64 (latency only)	8,452	5,988	1.41x	1.41x is 1.72x less work per query x 0.77x cores busy during the row (1.51 against 1.97) x 1.00x clock x 1.07x counted on-CPU share: the middle term is occupancy, not search speed
W6-ef128	SQ8 recall control, ef=128 (latency only)	5,402	4,469	1.21x
W6-ef256	SQ8 recall control, ef=256 (latency only)	3,410	3,025	1.13x
W6-ef512	SQ8 recall control, ef=512 (latency only)	2,072	1,858	-	Qdrant pass 3 -21% against two passes that agree: one pass set the spread on this row, so a band built from it is not a noise band
W7	quantized: binary + oversampling	5,769	4,616	1.25x
W7-2bit	quantized: binary quantization, 2 bits + oversampling	5,465	4,636	1.18x
W7-1p5bit	quantized: binary quantization, 1.5 bits + oversampling	5,525	4,538	1.22x
W8	quantized: PQ	3,440	2,888	1.19x
W14-1bit	quantized: TurboQuant, 1 bit + oversampling	4,612	4,570	parity	within the ±2.0% band this dataset's noise floor puts on W14-1bit: no measured difference, not a small one
W14-1p5bit	quantized: TurboQuant, 1.5 bits + oversampling	4,356	4,090	1.07x
W14-2bit	quantized: TurboQuant, 2 bits + oversampling	5,464	4,596	1.19x
W14-4bit	quantized: TurboQuant, 4 bits + oversampling	5,086	4,596	1.11x
W9	exact / brute force	311	311	parity	within the ±12.2% band this dataset's noise floor puts on W9: no measured difference, not a small one brute force over the whole collection, no index involved
W10-ef32	recall control, ef=32 (latency only)	27,901	18,065	-	strawmANN: -n 100000 (2x the table's 50000) recall unequal: 0.9630 vs 0.9513; §7.4 compares at equal recall
W10-ef64	recall control, ef=64 (latency only)	17,667	12,438	1.42x
W10-ef128	recall control, ef=128 (latency only)	10,487	8,268	1.27x
W10-ef256	recall control, ef=256 (latency only)	6,012	5,380	1.12x
W10-ef512	recall control, ef=512 (latency only)	3,423	3,373	parity	within the ±2.0% band this dataset's noise floor puts on W10-ef512: no measured difference, not a small one
W12-sel1	filtered search, one keyword (~1% of bench12)	7,096	3,368	-	strawmANN pass 1 -16% against two passes that agree: one pass set the spread on this row, so a band built from it is not a noise band filtered search over a keyword payload index both engines were asked to build; whether each did is read back from the engine and refused beside the row when it did not. Recall is measured under the filter, against ground truth restricted to the matching set
W12-sel10	filtered search, any of 10 keywords (~10% of bench12)	1,322	1,803	-	recall unequal: 1.0000 vs 0.9854; §7.4 compares at equal recall filtered search over a keyword payload index both engines were asked to build; whether each did is read back from the engine and refused beside the row when it did not. Recall is measured under the filter, against ground truth restricted to the matching set
W12-sel1-ef32	filtered recall control, one keyword, ef=32 (latency only)	7,291	5,774	1.26x
W12-sel1-ef64	filtered recall control, one keyword, ef=64 (latency only)	7,312	4,527	1.62x
W12-sel1-ef128	filtered recall control, one keyword, ef=128 (latency only)	7,255	3,339	2.17x	2.17x is 2.72x less work per query x 0.80x cores busy during the row (1.60 against 2.00) x 1.00x clock: the middle term is occupancy, not search speed
W12-sel1-ef256	filtered recall control, one keyword, ef=256 (latency only)	7,076	2,441	-	strawmANN pass 1 -39% against two passes that agree: one pass set the spread on this row, so a band built from it is not a noise band
W12-sel1-ef512	filtered recall control, one keyword, ef=512 (latency only)	6,833	1,693	4.04x	4.04x is 5.32x less work per query x 0.79x cores busy during the row (1.59 against 2.00) x 1.00x clock x 0.96x counted on-CPU share: the middle term is occupancy, not search speed
W12-sel10-ef32	filtered recall control, any of 10 keywords, ef=32 (latency only)	3,212	3,980	-	recall unequal: 0.9808 vs 0.8164; §7.4 compares at equal recall
W12-sel10-ef64	filtered recall control, any of 10 keywords, ef=64 (latency only)	2,077	2,780	-	recall unequal: 0.9932 vs 0.9326; §7.4 compares at equal recall
W12-sel10-ef128	filtered recall control, any of 10 keywords, ef=128 (latency only)	1,319	1,800	-	recall unequal: 1.0000 vs 0.9854; §7.4 compares at equal recall
W12-sel10-ef256	filtered recall control, any of 10 keywords, ef=256 (latency only)	1,319	1,024	1.29x
W12-sel10-ef512	filtered recall control, any of 10 keywords, ef=512 (latency only)	1,320	612	2.16x
W13	scroll / pagination	50,282	46,273	1.09x	strawmANN: -n 200000 (4x the table's 50000) Qdrant: -n 100000 (2x the table's 50000) strawmANN: 0.13 ms of the client's 0.14 ms p50 is not server time (13x), so 92% of this row is the load generator and the socket Qdrant: 0.13 ms of the client's 0.16 ms p50 is not server time (6x), so 82% of this row is the load generator and the socket 1.09x is 3.29x less work per query x 0.33x cores busy during the row (1.38 against 4.24) x 1.00x clock x 1.01x counted on-CPU share: the middle term is occupancy, not search speed
W11-steady	mixed read/write below the rebuild threshold: search bench2 while 5,000 synthetic points append	10,102	7,170	-	strawmANN: qps over the 25.0 s of the append the search saw (10,152 over the whole search); append 200 points/s Qdrant: qps over the 25.0 s of the append the search saw (7,523 over the whole search); append 200 points/s search-during-write; no recall join
W11	mixed read/write: search bench2 while 20,000 synthetic points append (runs last)	9,526	7,151	-	strawmANN: qps over the 66.7 s of the append the search saw (9,566 over the whole search); append 300 points/s Qdrant: qps over the 66.7 s of the append the search saw (7,528 over the whole search); append 300 points/s search-during-write; no recall join

W10: throughput against ef

All latency rows closed loop included, p50 to max

A row flagged “… not server time” is one where the client saw far more than the server reported: below the rate at which a queue can form, what is left is the load generator's own scheduling.

workload		strawmANN p50	strawmANN p95	strawmANN p99	strawmANN p99.9	strawmANN max	Qdrant p50	Qdrant p95	Qdrant p99	Qdrant p99.9	Qdrant max
W0	d=4 floor: graph traversal with the distance taken out closed loop	205 µs	232 µs	250 µs	305 µs	3.61 ms	242 µs	274 µs	300 µs	364 µs	2.19 ms
W3	search, fp32, single query closed loop	574 µs	888 µs	966 µs	1.07 ms	2.69 ms	611 µs	747 µs	790 µs	860 µs	3.62 ms
W4	search, saturating (closed loop) closed loop	5.90 ms	6.34 ms	6.54 ms	19.21 ms	24.65 ms	5.36 ms	10.67 ms	14.89 ms	26.91 ms	37.17 ms
W4-sat50	search, fixed rate at 50% of saturation (open loop) open loop	696 µs	1.10 ms	1.34 ms	1.59 ms	4.44 ms	856 µs	1.56 ms	1.95 ms	2.39 ms	6.06 ms
W4-sat70	search, fixed rate at 70% of saturation (open loop) open loop	861 µs	1.58 ms	1.89 ms	2.21 ms	5.27 ms	942 µs	1.81 ms	2.20 ms	3.30 ms	10.24 ms
W4-sat90	search, fixed rate at 90% of saturation (open loop) open loop	913 µs	1.68 ms	2.14 ms	4.15 ms	6.03 ms	1.33 ms	3.02 ms	4.91 ms	8.14 ms	19.53 ms
W5	search batched (16 distinct dataset queries per request) closed loop	3.00 ms	3.36 ms	3.52 ms	3.69 ms	8.21 ms	8.50 ms	9.16 ms	9.40 ms	9.87 ms	12.68 ms
W6	quantized: scalar closed loop	355 µs	463 µs	538 µs	660 µs	1.36 ms	447 µs	502 µs	534 µs	632 µs	1.71 ms
W6-ef32	SQ8 recall control, ef=32 (latency only) closed loop	180 µs	224 µs	247 µs	280 µs	990 µs	260 µs	310 µs	344 µs	389 µs	1.43 ms
W6-ef64	SQ8 recall control, ef=64 (latency only) closed loop	229 µs	296 µs	324 µs	369 µs	975 µs	327 µs	389 µs	437 µs	491 µs	1.45 ms
W6-ef128	SQ8 recall control, ef=128 (latency only) closed loop	360 µs	485 µs	544 µs	605 µs	1.41 ms	445 µs	500 µs	531 µs	618 µs	1.67 ms
W6-ef256	SQ8 recall control, ef=256 (latency only) closed loop	576 µs	766 µs	868 µs	950 µs	1.41 ms	662 µs	745 µs	779 µs	842 µs	1.96 ms
W6-ef512	SQ8 recall control, ef=512 (latency only) closed loop	964 µs	1.15 ms	1.31 ms	1.53 ms	1.87 ms	1.08 ms	1.23 ms	1.28 ms	1.36 ms	4.22 ms
W7	quantized: binary + oversampling closed loop	341 µs	425 µs	469 µs	629 µs	1.51 ms	423 µs	514 µs	550 µs	607 µs	1.76 ms
W7-2bit	quantized: binary quantization, 2 bits + oversampling closed loop	361 µs	441 µs	474 µs	537 µs	1.26 ms	420 µs	514 µs	548 µs	609 µs	1.65 ms
W7-1p5bit	quantized: binary quantization, 1.5 bits + oversampling closed loop	357 µs	441 µs	477 µs	541 µs	1.06 ms	431 µs	526 µs	563 µs	621 µs	1.81 ms
W8	quantized: PQ closed loop	580 µs	686 µs	726 µs	780 µs	1.36 ms	692 µs	776 µs	816 µs	897 µs	1.99 ms
W14-1bit	quantized: TurboQuant, 1 bit + oversampling closed loop	433 µs	507 µs	538 µs	604 µs	1.17 ms	429 µs	520 µs	554 µs	610 µs	1.62 ms
W14-1p5bit	quantized: TurboQuant, 1.5 bits + oversampling closed loop	458 µs	545 µs	576 µs	631 µs	1.52 ms	484 µs	577 µs	655 µs	726 µs	1.68 ms
W14-2bit	quantized: TurboQuant, 2 bits + oversampling closed loop	362 µs	444 µs	476 µs	519 µs	1.13 ms	425 µs	518 µs	551 µs	607 µs	3.75 ms
W14-4bit	quantized: TurboQuant, 4 bits + oversampling closed loop	387 µs	482 µs	516 µs	572 µs	1.16 ms	429 µs	499 µs	557 µs	623 µs	1.64 ms
W9	exact / brute force closed loop	25.51 ms	32.73 ms	35.49 ms	39.53 ms	39.81 ms	26.05 ms	29.72 ms	31.80 ms	36.92 ms	40.22 ms
W10-ef32	recall control, ef=32 (latency only) closed loop	279 µs	360 µs	403 µs	499 µs	2.06 ms	411 µs	649 µs	801 µs	1.11 ms	2.42 ms
W10-ef64	recall control, ef=64 (latency only) closed loop	447 µs	583 µs	644 µs	798 µs	2.89 ms	595 µs	984 µs	1.19 ms	1.55 ms	2.81 ms
W10-ef128	recall control, ef=128 (latency only) closed loop	757 µs	1.02 ms	1.14 ms	1.29 ms	4.59 ms	885 µs	1.55 ms	1.84 ms	2.35 ms	3.87 ms
W10-ef256	recall control, ef=256 (latency only) closed loop	1.32 ms	1.80 ms	2.02 ms	2.28 ms	6.84 ms	1.36 ms	2.32 ms	2.78 ms	3.48 ms	5.42 ms
W10-ef512	recall control, ef=512 (latency only) closed loop	2.33 ms	3.16 ms	3.56 ms	4.06 ms	8.28 ms	2.24 ms	3.60 ms	4.18 ms	5.27 ms	8.66 ms
W12-sel1	filtered search, one keyword (~1% of bench12) closed loop	271 µs	340 µs	449 µs	698 µs	1.52 ms	591 µs	645 µs	676 µs	751 µs	2.23 ms
W12-sel10	filtered search, any of 10 keywords (~10% of bench12) closed loop	1.51 ms	1.61 ms	1.64 ms	1.70 ms	4.74 ms	1.11 ms	1.28 ms	1.34 ms	1.42 ms	4.68 ms
W12-sel1-ef32	filtered recall control, one keyword, ef=32 (latency only) closed loop	269 µs	309 µs	339 µs	560 µs	1.26 ms	341 µs	385 µs	413 µs	489 µs	1.67 ms
W12-sel1-ef64	filtered recall control, one keyword, ef=64 (latency only) closed loop	269 µs	304 µs	335 µs	569 µs	1.26 ms	438 µs	484 µs	510 µs	559 µs	1.89 ms
W12-sel1-ef128	filtered recall control, one keyword, ef=128 (latency only) closed loop	271 µs	307 µs	331 µs	504 µs	1.33 ms	596 µs	654 µs	689 µs	761 µs	2.23 ms
W12-sel1-ef256	filtered recall control, one keyword, ef=256 (latency only) closed loop	273 µs	341 µs	372 µs	629 µs	1.34 ms	815 µs	886 µs	928 µs	998 µs	2.60 ms
W12-sel1-ef512	filtered recall control, one keyword, ef=512 (latency only) closed loop	284 µs	340 µs	402 µs	629 µs	1.21 ms	1.18 ms	1.27 ms	1.33 ms	1.46 ms	3.34 ms
W12-sel10-ef32	filtered recall control, any of 10 keywords, ef=32 (latency only) closed loop	627 µs	723 µs	775 µs	935 µs	1.66 ms	495 µs	587 µs	638 µs	716 µs	1.87 ms
W12-sel10-ef64	filtered recall control, any of 10 keywords, ef=64 (latency only) closed loop	974 µs	1.13 ms	1.17 ms	1.23 ms	2.30 ms	714 µs	829 µs	886 µs	979 µs	3.29 ms
W12-sel10-ef128	filtered recall control, any of 10 keywords, ef=128 (latency only) closed loop	1.52 ms	1.61 ms	1.65 ms	1.73 ms	4.71 ms	1.11 ms	1.28 ms	1.35 ms	1.45 ms	3.26 ms
W12-sel10-ef256	filtered recall control, any of 10 keywords, ef=256 (latency only) closed loop	1.52 ms	1.61 ms	1.64 ms	1.73 ms	4.87 ms	1.95 ms	2.26 ms	2.37 ms	2.50 ms	4.53 ms
W12-sel10-ef512	filtered recall control, any of 10 keywords, ef=512 (latency only) closed loop	1.51 ms	1.61 ms	1.65 ms	1.72 ms	4.69 ms	3.26 ms	3.73 ms	3.90 ms	4.12 ms	6.70 ms
W13	scroll / pagination closed loop	139 µs	207 µs	235 µs	275 µs	1.05 ms	159 µs	224 µs	296 µs	440 µs	2.28 ms
W11-steady	mixed read/write below the rebuild threshold: search bench2 while 5,000 synthetic points append closed loop	771 µs	1.04 ms	1.34 ms	3.00 ms	10.06 ms	932 µs	1.71 ms	2.26 ms	3.83 ms	24.44 ms
W11	mixed read/write: search bench2 while 20,000 synthetic points append (runs last) closed loop	809 µs	1.10 ms	1.98 ms	3.91 ms	12.33 ms	937 µs	1.70 ms	2.25 ms	6.85 ms	135.73 ms

Storage and I/O block layer and syscalls, per row

The syscall rows count every descriptor, sockets included, so on a search row they measure the network rather than the disk.

	strawmANN	Qdrant
storage on disk	1.2 GiBbefore W11-steady, W11allocated at --capacity	1.3 GiBbefore W11-steady, W11
peak RSS	1.5 GiB	5.0 GiB3.1 GiB before the writers
of which anonymous	315.8 MiB	288.5 MiB
of which file-backed	1.2 GiB	2.2 GiB
disk read bytes	0	416.0 MiB
disk write bytes	2.3 GiB	19.6 GiB
disk read ops	0	12,702
disk write ops	16,702	348,284
syscall reads (all fds)	16,037,871	13,742
syscall writes (all fds)	7,446,285	9,995,151

measured via strawmANN: proc+cgroup / Qdrant: proc+cgroup. proc supplies syscall counts and block-layer bytes, cgroup supplies block-layer operations and bytes, so a row one interface does not carry reads n/a via …. unknown means it was not measured, and 0 means it was: an engine started with no --data-dir has no store, which is the row this section exists for. storage on disk is the level before the rows with a concurrent writer: during those an engine that rewrites segments is caught mid-rewrite, and the same row has read 3.47 and 10.11 GiB on two runs of one binary. peak RSS does include them, being a peak.

Per workload

A search row doing block-layer reads is an engine going to disk to answer a query.

workload	strawmANN written	strawmANN write ops	Qdrant read	Qdrant written	Qdrant read ops	Qdrant write ops
W0-upload	6.1 MiB	0	10.2 MiB	35.0 MiB	375	950
W0	0	99	0	0	0	0
W1	195.4 MiB	0	768.0 KiB	436.5 MiB	57	4,083
W2	195.4 MiB	3	11.9 MiB	843.1 MiB	411	14,911
W3	0	3,156	0	0	0	6
W4	0	0	0	0	0	0
W4-sat50	0	0	0	0	0	0
W4-sat70	0	0	0	0	0	0
W4-sat90	0	0	0	0	0	0
W5	0	0	0	0	0	0
W6-upload	195.4 MiB	0	12.7 MiB	992.2 MiB	435	16,643
W6	0	3,153	0	0	0	0
W6-ef32	0	0	0	0	0	6
W6-ef64	0	3	0	0	0	0
W6-ef128	0	0	0	0	0	0
W6-ef256	0	0	0	0	0	0
W6-ef512	0	0	0	0	0	0
W7-upload	195.4 MiB	0	18.7 MiB	973.9 MiB	630	17,414
W7	0	3,153	0	0	0	0
W7-2bit-upload	195.4 MiB	0	14.1 MiB	902.0 MiB	474	16,056
W7-2bit	0	3	0	0	0	0
W7-1p5bit-upload	195.4 MiB	0	15.6 MiB	903.3 MiB	543	16,239
W7-1p5bit	0	3	0	0	0	0
W8-upload	195.4 MiB	3,154	11.6 MiB	814.3 MiB	357	14,401
W8	0	3	0	0	0	6
W14-1bit-upload	195.4 MiB	0	16.1 MiB	898.4 MiB	525	16,084
W14-1bit	0	0	0	0	0	9
W14-1p5bit-upload	195.4 MiB	0	15.2 MiB	880.6 MiB	483	15,638
W14-1p5bit	0	0	0	0	0	0
W14-2bit-upload	195.4 MiB	0	12.8 MiB	889.9 MiB	420	15,805
W14-2bit	0	0	0	0	0	3
W14-4bit-upload	195.4 MiB	0	11.7 MiB	912.4 MiB	402	16,137
W14-4bit	0	0	0	0	0	6
W9	0	0	0	0	0	0
W10-ef32	0	0	0	0	0	0
W10-ef64	0	0	0	0	0	0
W10-ef128	0	0	0	0	0	0
W10-ef256	0	0	0	0	0	0
W10-ef512	0	0	0	0	0	0
W12-upload	195.4 MiB	0	22.1 MiB	837.4 MiB	684	15,314
W12-sel1	0	3,153	0	0	0	0
W12-sel10	0	3	0	0	0	0
W12-sel1-ef32	0	0	0	0	0	0
W12-sel1-ef64	0	0	0	0	0	0
W12-sel1-ef128	0	0	0	0	0	0
W12-sel1-ef256	0	0	0	0	0	0
W12-sel1-ef512	0	0	0	0	0	0
W12-sel10-ef32	0	0	0	0	0	0
W12-sel10-ef64	0	0	0	0	0	0
W12-sel10-ef128	0	0	0	0	0	0
W12-sel10-ef256	0	0	0	0	0	0
W12-sel10-ef512	0	0	0	0	0	0
W13	0	0	0	0	0	0
W11-steady	9.9 MiB	164	95.2 MiB	3.7 GiB	2,787	66,097
W11	39.4 MiB	652	147.3 MiB	5.8 GiB	4,119	102,476

Scheduler and memory CPU use, run-queue wait, migrations

Whether the engine was running while it ran. of wall is CPU over elapsed, waiting is runnable-but-not-scheduled, migrations checks the pinning claim, and switches gives voluntary over involuntary — the scheduler taking the core away against the engine choosing to sleep, which per unit of work is the cheapest signal of lock contention there is. A dash is not a zero: an index build's threads exit before the row does, and threads shows what fraction of the row the survivors account for.

Time spent waiting for a core

Lower is better; the bar between a pair is the gap.

Peak memory per row

Peak RSS while the row ran, from the engine's own process. Unlike the disk counters this is not refused across the two engines: residency decides where bytes live, and this is what the process held either way.

workload	strawmANN							Qdrant
workload	cpu	of wall	waiting	switches vol/invol	migrations	faults min/maj	threads	cpu	of wall	waiting	switches vol/invol	migrations	faults min/maj	threads
W0-upload	8.5 s	138%	3.4 ms	4,121 / 36	0	12,642 / 0	9 2%	11.5 s	198%	857.4 ms	16,417 / 1,812	1,632	46,470 / 111,010	43 14%
W0	7.1 s	68%	13.1 ms	168,401 / 7	0	0 / 0	9	12.6 s	103%	3.7 ms	561,307 / 19	146	377 / 0	37 0%
W1	0.6 s	27%	0.2 ms	35,840 / 0	0	56,822 / 0	9 99%	7.1 s	222%	2,458.8 ms	14,357 / 2,626	3,176	50,609 / 118,243	53 41%
W2	56.5 s	497%	27.6 ms	36,348 / 268	0	62,863 / 0	9 1%	71.3 s	433%	3,346.4 ms	17,108 / 4,268	3,685	88,140 / 160,650	44 6%
W3	26.7 s	88%	6.9 ms	185,798 / 23	0	6 / 0	9	31.4 s	103%	3.7 ms	560,703 / 51	144	363 / 0	39 0%
W4	34.1 s	730%	2.5 ms	65,942 / 46	0	65 / 0	9	35.6 s	771%	82,694.5 ms	57,535 / 41,720	15,794	2,042 / 0	76
W4-sat50	117.4 s	321%	14.9 ms	348,316 / 63	0	8 / 0	9	148.1 s	405%	119,429.2 ms	1,210,934 / 124,899	439,062	314 / 0	68 89%
W4-sat70	131.2 s	501%	3.7 ms	278,186 / 54	0	28 / 0	9	153.0 s	585%	169,227.7 ms	1,104,084 / 172,332	452,191	587 / 0	68
W4-sat90	126.6 s	621%	15.2 ms	253,342 / 86	0	56 / 0	9	139.2 s	683%	122,345.4 ms	529,807 / 133,561	178,031	1,282 / 0	44 0%
W5	33.0 s	702%	2.1 ms	9,862 / 54	0	0 / 0	9	26.4 s	198%	4.7 ms	35,577 / 180	1,355	585 / 0	42 0%
W6-upload	40.0 s	385%	17.0 ms	7,384 / 191	0	128,700 / 0	9 1%	34.5 s	309%	1,680.1 ms	18,744 / 2,047	2,722	122,288 / 160,677	42 0%
W6	14.9 s	164%	11.8 ms	168,070 / 11	0	0 / 0	9	22.8 s	202%	12.3 ms	502,120 / 36	10,087	430 / 0	44 0%
W6-ef32	6.2 s	133%	2.1 ms	140,698 / 7	0	0 / 0	9	13.1 s	196%	10.5 ms	466,292 / 32	10,584	73 / 0	44
W6-ef64	9.0 s	151%	1.5 ms	157,603 / 3	0	0 / 0	9	16.6 s	198%	12.3 ms	477,760 / 24	9,873	87 / 0	44
W6-ef128	15.3 s	165%	5.7 ms	170,493 / 7	0	0 / 0	9	22.6 s	202%	11.4 ms	501,314 / 43	10,050	69 / 0	44
W6-ef256	25.9 s	176%	1.3 ms	177,766 / 11	0	0 / 0	9	33.6 s	203%	11.4 ms	525,535 / 49	12,715	172 / 0	44
W6-ef512	44.8 s	186%	5.8 ms	182,326 / 36	0	0 / 0	9	54.6 s	202%	15.2 ms	539,775 / 185	17,661	358 / 0	44
W7-upload	17.1 s	232%	7.8 ms	8,615 / 82	0	64,793 / 0	9 2%	31.1 s	269%	1,866.1 ms	20,159 / 2,366	2,983	107,187 / 196,234	42 0%
W7	14.1 s	162%	9.1 ms	168,802 / 6	0	1 / 0	9	21.7 s	200%	13.7 ms	495,377 / 53	10,273	440 / 0	46 0%
W7-2bit-upload	19.0 s	226%	7.4 ms	8,250 / 99	1	66,716 / 0	9 2%	28.9 s	250%	1,449.4 ms	20,120 / 2,158	2,745	100,202 / 180,734	50 23%
W7-2bit	14.9 s	163%	3.1 ms	171,867 / 14	0	0 / 0	9	21.6 s	200%	11.9 ms	491,664 / 40	10,027	380 / 0	50 0%
W7-1p5bit-upload	19.4 s	232%	7.5 ms	7,574 / 92	1	66,708 / 0	9 2%	33.2 s	266%	1,604.3 ms	20,125 / 2,152	2,756	100,267 / 179,982	51 0%
W7-1p5bit	14.7 s	162%	3.4 ms	170,077 / 9	0	0 / 0	9	22.2 s	201%	15.1 ms	495,584 / 38	10,269	463 / 0	50 0%
W8-upload	342.9 s	722%	104.2 ms	8,832 / 1,583	0	121,701 / 0	9 0%	271.2 s	563%	1,586.7 ms	106,277 / 11,324	8,238	92,993 / 126,801	42 0%
W8	25.8 s	177%	1.4 ms	176,141 / 17	0	0 / 0	9	35.1 s	202%	18.7 ms	512,151 / 102	12,396	724 / 0	46
W14-1bit-upload	57.8 s	466%	25.7 ms	7,831 / 248	0	67,661 / 0	9 1%	33.9 s	267%	1,597.1 ms	22,091 / 7,169	3,844	99,537 / 179,408	51 0%
W14-1bit	18.2 s	167%	11.2 ms	171,064 / 12	0	0 / 0	9	22.1 s	201%	13.4 ms	495,888 / 43	10,473	364 / 0	50 0%
W14-1p5bit-upload	58.7 s	473%	27.5 ms	7,897 / 284	0	69,155 / 0	9 1%	35.9 s	274%	1,853.0 ms	21,159 / 3,930	3,132	99,394 / 163,226	50 22%
W14-1p5bit	19.5 s	169%	3.3 ms	174,050 / 11	0	0 / 0	9	24.7 s	202%	9.1 ms	515,129 / 41	11,740	511 / 0	50 0%
W14-2bit-upload	58.1 s	468%	27.2 ms	8,521 / 275	0	71,668 / 0	9 1%	36.5 s	281%	2,027.8 ms	24,068 / 8,112	4,012	100,282 / 163,988	50 20%
W14-2bit	15.2 s	165%	2.5 ms	170,628 / 11	0	0 / 0	9	21.9 s	200%	15.0 ms	491,285 / 39	9,711	154 / 0	50 0%
W14-4bit-upload	58.8 s	475%	26.8 ms	7,657 / 276	0	79,674 / 0	9 1%	37.2 s	279%	1,927.8 ms	30,707 / 16,230	4,040	113,836 / 167,086	51 0%
W14-4bit	16.3 s	165%	3.9 ms	172,180 / 8	0	0 / 0	9	21.9 s	201%	15.2 ms	494,084 / 54	9,940	481 / 0	50 0%
W9	45.4 s	704%	4.3 ms	3,835 / 73	0	0 / 0	9	49.8 s	772%	7,100.6 ms	15,586 / 7,223	5,751	80 / 0	65
W10-ef32	21.4 s	584%	3.3 ms	162,478 / 17	0	0 / 0	9	16.1 s	575%	6,265.7 ms	232,650 / 18,876	65,474	856 / 0	48 0%
W10-ef64	18.9 s	659%	2.2 ms	90,642 / 16	0	0 / 0	9	24.0 s	592%	14,521.1 ms	274,069 / 34,041	111,342	524 / 0	72
W10-ef128	33.3 s	693%	3.8 ms	97,283 / 32	0	0 / 0	9	36.6 s	602%	26,270.7 ms	295,037 / 35,762	117,455	214 / 0	66 0%
W10-ef256	58.8 s	704%	4.7 ms	104,254 / 64	0	0 / 0	9	58.6 s	628%	38,470.1 ms	309,047 / 40,585	143,422	179 / 0	66
W10-ef512	103.6 s	708%	8.4 ms	111,925 / 135	0	0 / 0	9	96.4 s	649%	53,106.6 ms	331,443 / 48,884	169,278	293 / 0	66
W12-upload	56.5 s	450%	24.9 ms	10,585 / 256	0	63,572 / 0	9 1%	85.5 s	351%	1,653.3 ms	31,032 / 3,092	3,459	112,773 / 154,555	51 0%
W12-sel1	11.2 s	159%	10.8 ms	177,424 / 9	0	1 / 0	9	29.7 s	200%	15.7 ms	498,516 / 55	9,230	621 / 0	56
W12-sel10	71.2 s	188%	7.6 ms	130,386 / 87	0	0 / 0	9	56.2 s	202%	13.6 ms	537,860 / 133	17,354	381 / 0	56
W12-sel1-ef32	11.0 s	160%	1.2 ms	174,196 / 5	0	0 / 0	9	17.2 s	198%	12.5 ms	478,083 / 59	9,810	36 / 0	56
W12-sel1-ef64	11.0 s	160%	2.1 ms	176,346 / 8	0	0 / 0	9	22.2 s	200%	13.0 ms	490,172 / 68	9,451	62 / 0	56
W12-sel1-ef128	11.1 s	159%	1.9 ms	177,201 / 4	0	0 / 0	9	30.2 s	201%	16.8 ms	499,372 / 106	10,052	251 / 0	56
W12-sel1-ef256	11.3 s	159%	2.2 ms	176,052 / 4	0	0 / 0	9	41.3 s	201%	22.1 ms	504,917 / 166	10,027	218 / 0	56
W12-sel1-ef512	11.7 s	159%	1.8 ms	175,182 / 3	0	0 / 0	9	59.5 s	201%	41.6 ms	509,883 / 256	11,588	313 / 0	56
W12-sel10-ef32	27.7 s	178%	3.0 ms	178,558 / 12	0	0 / 0	9	25.5 s	202%	10.8 ms	515,979 / 103	12,232	195 / 0	56
W12-sel10-ef64	44.7 s	185%	2.1 ms	179,629 / 23	0	0 / 0	9	36.5 s	202%	13.1 ms	528,464 / 178	14,006	292 / 0	56
W12-sel10-ef128	71.3 s	188%	8.4 ms	133,979 / 83	0	0 / 0	9	56.2 s	202%	16.4 ms	538,067 / 294	17,649	260 / 0	56
W12-sel10-ef256	71.3 s	188%	5.8 ms	133,831 / 72	0	0 / 0	9	98.6 s	202%	53.8 ms	548,624 / 505	19,972	756 / 0	56
W12-sel10-ef512	71.2 s	188%	4.7 ms	134,101 / 79	0	0 / 0	9	162.6 s	199%	91.6 ms	561,935 / 1,032	21,393	1,381 / 0	56
W13	5.6 s	137%	2.5 ms	245,517 / 7	0	0 / 0	9	9.1 s	409%	922.2 ms	382,586 / 5,038	113,126	622 / 0	67
W11-steady	228.5 s	685%	3,877.6 ms	641,262 / 1,386	16	23,682 / 0	9 98%	277.1 s	624%	247,440.9 ms	2,134,190 / 335,030	894,806	376,849 / 480,766	74 95%
W11	634.6 s	689%	18,357.3 ms	1,669,390 / 7,157	60	89,971 / 0	9 97%	720.3 s	617%	572,136.9 ms	5,316,734 / 716,405	2,221,296	658,653 / 439,853	74 95%

Waiting and stalls where the time went when the engine was not running

waiting is runnable and not scheduled, blocked on disk is not runnable at all, and pressure is PSI for the engine's cgroup (some = at least one task stalled, full = every runnable task).

Time stalled on I/O

Time every task in the engine's cgroup was stalled on I/O. Lower is better. 40 rows never stalled on I/O and are not drawn.

workload	strawmANN					Qdrant
workload	waiting	blocked on disk	cpu some/full	io some/full	memory some/full	waiting	blocked on disk	cpu some/full	io some/full	memory some/full
W0-upload	3.4 ms	0 ms	2 / 2 ms	0 / 0 ms	0 / 0 ms	857.4 ms	0 ms	96 / 2 ms	13 / 13 ms	0 / 0 ms
W0	13.1 ms	0 ms	48 / 48 ms	0 / 0 ms	0 / 0 ms	3.7 ms	0 ms	69 / 69 ms	0 / 0 ms	0 / 0 ms
W1	0.2 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	2,458.8 ms	0 ms	227 / 2 ms	14 / 9 ms	0 / 0 ms
W2	27.6 ms	0 ms	7 / 4 ms	0 / 0 ms	0 / 0 ms	3,346.4 ms	0 ms	326 / 5 ms	53 / 45 ms	0 / 0 ms
W3	6.9 ms	0 ms	48 / 48 ms	0 / 0 ms	0 / 0 ms	3.7 ms	0 ms	85 / 85 ms	0 / 0 ms	0 / 0 ms
W4	2.5 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	82,694.5 ms	0 ms	3,283 / 3 ms	0 / 0 ms	0 / 0 ms
W4-sat50	14.9 ms	0 ms	86 / 86 ms	0 / 0 ms	0 / 0 ms	119,429.2 ms	0 ms	8,875 / 109 ms	0 / 0 ms	0 / 0 ms
W4-sat70	3.7 ms	0 ms	55 / 55 ms	0 / 0 ms	0 / 0 ms	169,227.7 ms	0 ms	11,072 / 75 ms	0 / 0 ms	0 / 0 ms
W4-sat90	15.2 ms	0 ms	28 / 28 ms	0 / 0 ms	0 / 0 ms	122,345.4 ms	0 ms	8,508 / 40 ms	0 / 0 ms	0 / 0 ms
W5	2.1 ms	0 ms	1 / 1 ms	0 / 0 ms	0 / 0 ms	4.7 ms	0 ms	14 / 13 ms	0 / 0 ms	0 / 0 ms
W6-upload	17.0 ms	0 ms	4 / 2 ms	0 / 0 ms	0 / 0 ms	1,680.1 ms	0 ms	162 / 4 ms	55 / 54 ms	0 / 0 ms
W6	11.8 ms	0 ms	30 / 30 ms	0 / 0 ms	0 / 0 ms	12.3 ms	0 ms	56 / 55 ms	0 / 0 ms	0 / 0 ms
W6-ef32	2.1 ms	0 ms	23 / 23 ms	0 / 0 ms	0 / 0 ms	10.5 ms	0 ms	47 / 46 ms	0 / 0 ms	0 / 0 ms
W6-ef64	1.5 ms	0 ms	23 / 23 ms	0 / 0 ms	0 / 0 ms	12.3 ms	0 ms	52 / 51 ms	0 / 0 ms	0 / 0 ms
W6-ef128	5.7 ms	0 ms	30 / 30 ms	0 / 0 ms	0 / 0 ms	11.4 ms	0 ms	55 / 55 ms	0 / 0 ms	0 / 0 ms
W6-ef256	1.3 ms	0 ms	37 / 37 ms	0 / 0 ms	0 / 0 ms	11.4 ms	0 ms	60 / 59 ms	0 / 0 ms	0 / 0 ms
W6-ef512	5.8 ms	0 ms	36 / 36 ms	0 / 0 ms	0 / 0 ms	15.2 ms	0 ms	69 / 68 ms	0 / 0 ms	0 / 0 ms
W7-upload	7.8 ms	0 ms	2 / 2 ms	0 / 0 ms	0 / 0 ms	1,866.1 ms	0 ms	227 / 5 ms	37 / 34 ms	0 / 0 ms
W7	9.1 ms	0 ms	30 / 30 ms	0 / 0 ms	0 / 0 ms	13.7 ms	0 ms	57 / 56 ms	0 / 0 ms	0 / 0 ms
W7-2bit-upload	7.4 ms	0 ms	2 / 2 ms	0 / 0 ms	0 / 0 ms	1,449.4 ms	0 ms	172 / 5 ms	49 / 48 ms	0 / 0 ms
W7-2bit	3.1 ms	0 ms	32 / 32 ms	0 / 0 ms	0 / 0 ms	11.9 ms	0 ms	57 / 56 ms	0 / 0 ms	0 / 0 ms
W7-1p5bit-upload	7.5 ms	0 ms	3 / 2 ms	0 / 0 ms	0 / 0 ms	1,604.3 ms	0 ms	166 / 5 ms	47 / 45 ms	0 / 0 ms
W7-1p5bit	3.4 ms	0 ms	29 / 29 ms	0 / 0 ms	0 / 0 ms	15.1 ms	0 ms	58 / 57 ms	0 / 0 ms	0 / 0 ms
W8-upload	104.2 ms	0 ms	15 / 4 ms	0 / 0 ms	0 / 0 ms	1,586.7 ms	0 ms	342 / 35 ms	51 / 46 ms	0 / 0 ms
W8	1.4 ms	0 ms	32 / 32 ms	0 / 0 ms	0 / 0 ms	18.7 ms	0 ms	70 / 69 ms	0 / 0 ms	0 / 0 ms
W14-1bit-upload	25.7 ms	0 ms	5 / 2 ms	0 / 0 ms	0 / 0 ms	1,597.1 ms	0 ms	221 / 6 ms	70 / 67 ms	0 / 0 ms
W14-1bit	11.2 ms	0 ms	34 / 34 ms	0 / 0 ms	0 / 0 ms	13.4 ms	0 ms	58 / 57 ms	0 / 0 ms	0 / 0 ms
W14-1p5bit-upload	27.5 ms	0 ms	5 / 2 ms	7 / 7 ms	0 / 0 ms	1,853.0 ms	0 ms	231 / 7 ms	96 / 94 ms	0 / 0 ms
W14-1p5bit	3.3 ms	0 ms	33 / 33 ms	0 / 0 ms	0 / 0 ms	9.1 ms	0 ms	57 / 56 ms	0 / 0 ms	0 / 0 ms
W14-2bit-upload	27.2 ms	0 ms	5 / 2 ms	1 / 1 ms	0 / 0 ms	2,027.8 ms	0 ms	260 / 7 ms	58 / 54 ms	0 / 0 ms
W14-2bit	2.5 ms	0 ms	27 / 27 ms	0 / 0 ms	0 / 0 ms	15.0 ms	0 ms	57 / 56 ms	0 / 0 ms	0 / 0 ms
W14-4bit-upload	26.8 ms	0 ms	5 / 2 ms	0 / 0 ms	0 / 0 ms	1,927.8 ms	0 ms	237 / 8 ms	55 / 52 ms	0 / 0 ms
W14-4bit	3.9 ms	0 ms	32 / 32 ms	0 / 0 ms	0 / 0 ms	15.2 ms	0 ms	56 / 55 ms	0 / 0 ms	0 / 0 ms
W9	4.3 ms	0 ms	1 / 1 ms	0 / 0 ms	0 / 0 ms	7,100.6 ms	0 ms	823 / 11 ms	0 / 0 ms	0 / 0 ms
W10-ef32	3.3 ms	0 ms	21 / 21 ms	0 / 0 ms	0 / 0 ms	6,265.7 ms	0 ms	715 / 16 ms	0 / 0 ms	0 / 0 ms
W10-ef64	2.2 ms	0 ms	11 / 11 ms	0 / 0 ms	0 / 0 ms	14,521.1 ms	0 ms	1,351 / 18 ms	0 / 0 ms	0 / 0 ms
W10-ef128	3.8 ms	0 ms	10 / 10 ms	0 / 0 ms	0 / 0 ms	26,270.7 ms	0 ms	2,265 / 21 ms	0 / 0 ms	0 / 0 ms
W10-ef256	4.7 ms	0 ms	9 / 9 ms	0 / 0 ms	0 / 0 ms	38,470.1 ms	0 ms	3,489 / 24 ms	0 / 0 ms	0 / 0 ms
W10-ef512	8.4 ms	0 ms	7 / 7 ms	0 / 0 ms	0 / 0 ms	53,106.6 ms	0 ms	5,089 / 27 ms	0 / 0 ms	0 / 0 ms
W12-upload	24.9 ms	0 ms	6 / 3 ms	0 / 0 ms	0 / 0 ms	1,653.3 ms	0 ms	164 / 9 ms	44 / 43 ms	0 / 0 ms
W12-sel1	10.8 ms	0 ms	25 / 25 ms	0 / 0 ms	0 / 0 ms	15.7 ms	0 ms	58 / 56 ms	0 / 0 ms	0 / 0 ms
W12-sel10	7.6 ms	0 ms	35 / 35 ms	0 / 0 ms	0 / 0 ms	13.6 ms	0 ms	71 / 70 ms	0 / 0 ms	0 / 0 ms
W12-sel1-ef32	1.2 ms	0 ms	22 / 22 ms	0 / 0 ms	0 / 0 ms	12.5 ms	0 ms	51 / 50 ms	0 / 0 ms	0 / 0 ms
W12-sel1-ef64	2.1 ms	0 ms	22 / 22 ms	0 / 0 ms	0 / 0 ms	13.0 ms	0 ms	55 / 53 ms	0 / 0 ms	0 / 0 ms
W12-sel1-ef128	1.9 ms	0 ms	24 / 24 ms	0 / 0 ms	0 / 0 ms	16.8 ms	0 ms	58 / 57 ms	0 / 0 ms	0 / 0 ms
W12-sel1-ef256	2.2 ms	0 ms	24 / 24 ms	0 / 0 ms	0 / 0 ms	22.1 ms	0 ms	62 / 60 ms	0 / 0 ms	0 / 0 ms
W12-sel1-ef512	1.8 ms	0 ms	24 / 24 ms	0 / 0 ms	0 / 0 ms	41.6 ms	0 ms	71 / 67 ms	0 / 0 ms	0 / 0 ms
W12-sel10-ef32	3.0 ms	0 ms	34 / 34 ms	0 / 0 ms	0 / 0 ms	10.8 ms	0 ms	57 / 56 ms	0 / 0 ms	0 / 0 ms
W12-sel10-ef64	2.1 ms	0 ms	34 / 34 ms	0 / 0 ms	0 / 0 ms	13.1 ms	0 ms	62 / 61 ms	0 / 0 ms	0 / 0 ms
W12-sel10-ef128	8.4 ms	0 ms	35 / 35 ms	0 / 0 ms	0 / 0 ms	16.4 ms	0 ms	70 / 69 ms	0 / 0 ms	0 / 0 ms
W12-sel10-ef256	5.8 ms	0 ms	35 / 35 ms	0 / 0 ms	0 / 0 ms	53.8 ms	0 ms	112 / 107 ms	0 / 0 ms	0 / 0 ms
W12-sel10-ef512	4.7 ms	0 ms	35 / 35 ms	0 / 0 ms	0 / 0 ms	91.6 ms	0 ms	142 / 135 ms	0 / 0 ms	0 / 0 ms
W13	2.5 ms	0 ms	30 / 30 ms	0 / 0 ms	0 / 0 ms	922.2 ms	0 ms	128 / 27 ms	0 / 0 ms	0 / 0 ms
W11-steady	3,877.6 ms	0 ms	830 / 66 ms	0 / 0 ms	0 / 0 ms	247,440.9 ms	0 ms	17,953 / 152 ms	201 / 46 ms	0 / 0 ms
W11	18,357.3 ms	0 ms	3,736 / 176 ms	10 / 10 ms	0 / 0 ms	572,136.9 ms	0 ms	44,599 / 386 ms	348 / 82 ms	0 / 0 ms

A dash under blocked on disk is not a zero: most hosts ship with kernel.task_delayacct=0, which reports the field as a permanent zero and would manufacture the strongest claim here — that the memory-resident engine never waits for a device — out of a sysctl. bench/setup.py apply turns it on.

Hardware counters cycles, DRAM loads, TLB walks, IPC per query

What the core did per query, from perf stat attached to the engine for the same window as the /proc counters. §5's cost model is a set of claims about instructions per cycle, DRAM traffic and TLB reach; these are those quantities, measured on the row the headline quotes rather than on a microbenchmark.

What one query cost, in cycles

Lower is better; the bar between a pair is the gap. Load rows are absent: they have no queries to divide by, and an absolute count under a per-query axis would be a different quantity wearing this one's label.

Demand loads from DRAM per query

Demand loads served from DRAM, times the cache line this host reports. Lower is better. Demand only: the hardware prefetcher's fills are a separate counter and are not in this number, so a row that streams — an exact scan above all — moved far more than this says. On the graph rows, where there is little for a prefetcher to predict, it is most of the traffic and is the outstanding-miss quantity §5.2 argues the engine is limited by.

TLB walks per query

Data TLB misses that reached a page walk, per query. Lower is better. §5.5 argues a memory-resident index needs hugepage care; this is what not taking it costs, and `--no-huge-pages` is the A/B arm that prices it.

Instructions per cycle

How well the core was fed while it ran. Not a score: a low IPC on a memory-bound row is what §5.2 predicts, and a high one on a row that does more work per query is not a win. Read it beside the two charts above.

Branch mispredictions per 1k instructions

A graph traversal is a chain of data-dependent branches and none of them is predictable from the last query, so this is the column that separates a slow kernel from an unpredictable walk. Per thousand instructions rather than per branch: `branches` costs a PMU counter that `ref-cycles` needs more, and MPKI is the comparable form anyway.

workload	strawmANN						Qdrant
workload	IPC	GHz	branch MPKI	cycles/query	demand DRAM/query	TLB walks/query	IPC	GHz	branch MPKI	cycles/query	demand DRAM/query	TLB walks/query
W0-upload	1.95	2.011.005x nominal	5.37	-	-	-	1.86	2.011.006x nominal	4.49	-	-	-
W0	1.50	2.011.006x nominal	5.35	233,469	4.4 KiB	582.9	1.43	2.011.006x nominal	3.16	374,574	7.2 KiB	725.9
W1	1.85	2.011.006x nominal	0.84	-	-	-	1.85	2.011.005x nominal	2.26	-	-	-
W2	0.77	2.011.005x nominal	3.78	-	-	-	0.93	2.011.006x nominal	3.39	-	-	-
W3	0.70	2.011.006x nominal	5.69	985,886	1,953.5 KiB	2,214.8	0.94	2.011.006x nominal	3.26	1,128,593	2,073.0 KiB	3,147.9
W4	0.50	2.011.006x nominal	4.87	1,356,665	1,838.9 KiB	1,682.0	0.69	2.011.006x nominal	3.78	1,371,556	1,717.6 KiB	2,313.0
W4-sat50	0.60	2.011.006x nominal	5.39	1,146,737	1,895.8 KiB	2,026.4	0.77	2.011.006x nominal	3.84	1,360,280	1,833.2 KiB	2,864.4
W4-sat70	0.52	2.011.006x nominal	5.32	1,294,216	1,898.4 KiB	1,918.8	0.75	2.011.007x nominal	3.83	1,396,471	1,786.9 KiB	2,757.8
W4-sat90	0.54	2.011.007x nominal	4.89	1,259,213	1,873.3 KiB	1,820.7	0.73	2.011.007x nominal	3.77	1,333,580	1,732.3 KiB	2,486.2
W5	0.47	2.011.006x nominal	5.16	1,325,995	1,910.1 KiB	1,671.3	0.77	2.011.006x nominal	4.34	1,037,501	2,140.9 KiB	2,182.3
W6-upload	1.21	2.011.007x nominal	3.57	-	-	-	2.34	2.011.006x nominal	2.84	-	-	-
W6	1.48	2.011.006x nominal	5.51	547,931	243.8 KiB	2,262.4	1.73	2.011.006x nominal	2.97	797,572	254.8 KiB	2,579.5
W6-ef32	1.46	2.011.007x nominal	3.27	219,352	60.4 KiB	960.3	1.73	2.011.006x nominal	2.17	422,321	68.4 KiB	1,214.5
W6-ef64	1.50	2.011.006x nominal	4.06	325,096	113.3 KiB	1,406.8	1.71	2.011.006x nominal	2.56	559,582	131.8 KiB	1,762.6
W6-ef128	1.44	2.011.007x nominal	5.55	562,388	243.5 KiB	2,265.7	1.74	2.011.007x nominal	2.98	793,797	256.0 KiB	2,583.4
W6-ef256	1.48	2.011.007x nominal	6.05	973,275	466.1 KiB	3,698.4	1.76	2.011.006x nominal	3.35	1,225,670	497.2 KiB	3,949.7
W6-ef512	1.54	2.011.007x nominal	6.31	1,731,137	919.0 KiB	6,511.6	1.74	2.011.006x nominal	3.67	2,057,181	975.3 KiB	6,289.6
W7-upload	1.81	2.011.005x nominal	6.70	-	-	-	2.70	2.011.005x nominal	3.77	-	-	-
W7	1.33	2.011.006x nominal	7.98	517,567	191.0 KiB	1,523.3	1.75	2.011.006x nominal	3.46	756,751	204.4 KiB	1,992.1
W7-2bit-upload	1.70	2.011.005x nominal	6.34	-	-	-	2.50	2.011.006x nominal	3.83	-	-	-
W7-2bit	1.31	2.011.005x nominal	7.87	546,372	196.8 KiB	1,860.8	1.68	2.011.006x nominal	3.42	754,912	207.4 KiB	2,304.4
W7-1p5bit-upload	1.73	2.011.006x nominal	6.19	-	-	-	2.61	2.011.006x nominal	3.77	-	-	-
W7-1p5bit	1.34	2.011.006x nominal	7.67	543,907	196.9 KiB	1,885.3	1.72	2.011.006x nominal	3.45	774,544	215.6 KiB	2,237.1
W8-upload	5.98	2.011.006x nominal	0.54	-	-	-	5.82	2.011.006x nominal	0.25	-	-	-
W8	3.40	2.011.006x nominal	1.61	971,825	199.6 KiB	1,976.0	2.86	2.011.007x nominal	1.16	1,276,102	204.9 KiB	2,477.8
W14-1bit-upload	0.88	2.011.005x nominal	3.30	-	-	-	2.70	2.011.005x nominal	2.62	-	-	-
W14-1bit	1.60	2.011.006x nominal	4.69	675,022	189.9 KiB	1,727.9	1.66	2.011.006x nominal	3.03	768,426	208.8 KiB	1,913.3
W14-1p5bit-upload	0.92	2.011.006x nominal	3.13	-	-	-	2.67	2.011.005x nominal	2.27	-	-	-
W14-1p5bit	1.73	2.011.006x nominal	4.16	726,348	195.2 KiB	1,917.3	1.59	2.011.005x nominal	2.78	874,531	258.7 KiB	2,038.6
W14-2bit-upload	0.88	2.011.005x nominal	3.33	-	-	-	2.78	2.011.006x nominal	1.94	-	-	-
W14-2bit	1.70	2.011.006x nominal	4.96	559,282	199.4 KiB	1,965.0	1.66	2.011.006x nominal	2.99	762,046	216.7 KiB	2,121.7
W14-4bit-upload	0.90	2.011.005x nominal	3.35	-	-	-	2.62	2.011.006x nominal	2.03	-	-	-
W14-4bit	1.64	2.011.006x nominal	4.94	600,297	208.3 KiB	2,215.8	1.66	2.011.006x nominal	2.95	764,630	234.3 KiB	2,331.9
W9	0.45	2.011.005x nominal	0.20	45,430,233	38,805.8 KiB	13,036.9	0.55	2.011.006x nominal	0.23	49,311,818	14,364.0 KiB	50,750.0
W10-ef32	0.69	2.011.007x nominal	3.46	409,203	629.3 KiB	633.4	0.96	2.011.007x nominal	2.72	580,475	680.1 KiB	1,027.0
W10-ef64	0.57	2.011.007x nominal	4.19	738,651	1,059.4 KiB	990.2	0.85	2.011.006x nominal	3.29	869,671	1,094.7 KiB	1,642.2
W10-ef128	0.52	2.011.007x nominal	4.87	1,316,102	1,871.8 KiB	1,720.8	0.77	2.011.006x nominal	3.80	1,352,164	1,800.8 KiB	2,622.0
W10-ef256	0.50	2.011.007x nominal	5.41	2,340,542	3,351.8 KiB	3,254.3	0.73	2.011.007x nominal	4.19	2,193,403	3,059.3 KiB	4,225.3
W10-ef512	0.51	2.011.007x nominal	5.77	4,131,218	5,975.3 KiB	6,506.7	0.71	2.011.006x nominal	4.49	3,689,417	5,217.7 KiB	6,967.8
W12-upload	0.77	2.011.006x nominal	3.78	-	-	-	1.18	2.011.006x nominal	3.26	-	-	-
W12-sel1	0.90	2.011.007x nominal	3.54	409,321	1,385.0 KiB	1,351.3	1.15	2.011.006x nominal	4.16	1,084,037	1,061.7 KiB	2,369.7
W12-sel10	0.83	2.011.007x nominal	1.16	2,809,433	9,484.6 KiB	8,566.1	1.16	2.011.007x nominal	2.75	2,120,999	2,510.2 KiB	4,006.1
W12-sel1-ef32	0.92	2.011.006x nominal	3.43	399,898	1,358.8 KiB	1,350.6	1.18	2.011.006x nominal	2.92	583,982	526.1 KiB	1,542.1
W12-sel1-ef64	0.92	2.011.006x nominal	3.43	401,349	1,360.8 KiB	1,345.2	1.15	2.011.006x nominal	3.52	779,509	770.0 KiB	1,943.1
W12-sel1-ef128	0.92	2.011.005x nominal	3.43	402,494	1,363.0 KiB	1,346.3	1.14	2.011.006x nominal	4.20	1,096,158	1,066.9 KiB	2,369.2
W12-sel1-ef256	0.89	2.011.006x nominal	3.49	411,211	1,360.0 KiB	1,363.0	1.21	2.011.006x nominal	4.81	1,538,110	1,356.5 KiB	2,796.1
W12-sel1-ef512	0.87	2.011.006x nominal	3.59	425,118	1,376.1 KiB	1,368.7	1.30	2.011.007x nominal	5.30	2,262,190	1,630.4 KiB	3,189.2
W12-sel10-ef32	1.28	2.011.007x nominal	3.98	1,047,546	1,199.4 KiB	2,529.9	1.19	2.011.007x nominal	2.35	904,771	898.0 KiB	2,036.0
W12-sel10-ef64	1.27	2.011.007x nominal	4.27	1,727,693	2,023.7 KiB	3,981.3	1.16	2.011.006x nominal	2.53	1,339,219	1,495.5 KiB	2,789.9
W12-sel10-ef128	0.83	2.011.006x nominal	1.15	2,812,021	9,484.7 KiB	8,566.4	1.16	2.011.007x nominal	2.75	2,125,320	2,509.3 KiB	4,004.9
W12-sel10-ef256	0.83	2.011.006x nominal	1.15	2,809,822	9,472.2 KiB	8,564.9	1.08	2.011.007x nominal	3.27	3,780,587	4,341.7 KiB	5,854.9
W12-sel10-ef512	0.83	2.011.007x nominal	1.16	2,812,624	9,476.0 KiB	8,562.5	1.11	2.011.006x nominal	3.46	6,315,394	6,872.6 KiB	8,444.3
W13	1.36	2.011.007x nominal	1.31	42,563	1.0 KiB	14.6	1.45	2.011.006x nominal	1.50	140,129	1.6 KiB	40.0
W11-steady	0.52	2.011.006x nominal	4.83	1,363,652	1,951.3 KiB	1,792.2	0.85	2.011.006x nominal	3.48	1,532,615	1,896.2 KiB	2,737.5
W11	0.51	2.011.007x nominal	4.73	1,451,374	2,074.0 KiB	1,927.6	0.82	2.011.007x nominal	3.40	1,526,540	1,952.6 KiB	2,803.9

Nothing here is scaled. Where perf had to multiplex the group, the values it prints are extrapolations from the fraction of the row each counter was on, and they are withheld rather than shown — the same rule the scheduler table applies to a thread that exited. An event this host does not implement is likewise blank, never zero.

The two instruments disagree on these rows. Qdrant: perf and /proc disagree on minor faults by up to 35% on W0-upload, W1, W2, W6-upload, W7-upload, W7-2bit-upload, W7-1p5bit-upload, W8-upload, W14-1bit-upload, W14-1p5bit-upload, W14-2bit-upload, W14-4bit-upload, W12-upload, W11-steady, W11; Qdrant: perf and /proc disagree on context switches by up to 70% on W1, W7-upload, W7-2bit-upload, W7-1p5bit-upload, W8-upload, W14-1bit-upload, W14-1p5bit-upload, W14-2bit-upload, W14-4bit-upload. perf keeps an exiting thread's counts and the /proc sums do not, so a gap here is usually the same lost-thread effect the scheduler table reports as coverage.

Collection config what each engine says it built

Read back from the engine rather than taken from the request. --segments 1 sets Qdrant's default_segment_number, which its optimizer treats as a target, and the segment count is the largest confound in the ef comparison. strawmANN held 1 segment; Qdrant held 2 segments.

Where the engines disagree a cell reads strawmANN / Qdrant, and is marked.

collection	segments	populated segments	requested segments	points	indexed vectors	vector size	hnsw m	hnsw ef_construct	quantization	shards	vector residency
bench0	1 / 2	- / 1	- / 1	100,000	100,000	4	16	100	-	1	cached
bench1	1 / 5	- / 5	- / 1	100,000	0 / 19,100	512	16	100	-	1	cached
bench2	1 / 2	- / 1	- / 1	100,000	100,000	512	16	100	-	1	cached
bench6	1 / 2	- / 1	- / 1	100,000	100,000	512	16	100	scalar	1	cached
bench7	1 / 2	- / 1	- / 1	100,000	100,000	512	16	100	binary	1	cached
bench7b2	1 / 2	- / 1	- / 1	100,000	100,000	512	16	100	binary	1	cached
bench7b15	1 / 2	- / 1	- / 1	100,000	100,000	512	16	100	binary	1	cached
bench8	1 / 2	- / 1	- / 1	100,000	100,000	512	16	100	product	1	cached
bench14t1	1 / 2	- / 1	- / 1	100,000	100,000	512	16	100	turboquant	1	cached
bench14t15	1 / 2	- / 1	- / 1	100,000	100,000	512	16	100	turboquant	1	cached
bench14t2	1 / 2	- / 1	- / 1	100,000	100,000	512	16	100	turboquant	1	cached
bench14t4	1 / 2	- / 1	- / 1	100,000	100,000	512	16	100	turboquant	1	cached
bench12	1 / 2	- / 1	- / 1	100,000	100,000	512	16	100	-	1	cached

After the mutating rows bench2 held strawmANN 125,000 (+25,000), 125,000 of them indexed; Qdrant 125,000 (+25,000), 125,000 of them indexed. The table above is the state the throughput, latency and recall rows searched; this is what W11 left, and is that row's subject rather than theirs.

bench1 was read back straight after its row and dropped, with no settle between, so an engine still building it reports a snapshot taken mid-ingest. Its cells are not marked as a disagreement.

Per workload every metric of every row, engine beside engine

Metric down, engine across, so a comparison is two adjacent cells. Metrics a row did not measure are dropped rather than shown empty.

W0-upload transport floor: load

upload · collection=bench0 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=100,000

metric	strawmANN	Qdrant
wall clock	6.2 s	5.8 s	0.94x
of which upload	0.1 s	0.5 s	4.20x
of which index wait	4.0 s	3.0 s	0.75x
time to Green	4.0 s	3.0 s (at bfb's polling floor)	0.75x
cpu	8.5 s	11.5 s	1.35x
cpu, % of wall	138%	198%
waiting for a core	3.4 ms	857.4 ms	252.69x
migrations	0	1,632
faults min/maj	12,642 / 0	46,470 / 111,010
peak RSS	144.6 MiB	688.1 MiB	4.76x
disk written	6.1 MiB	35.0 MiB
disk read	0	10.2 MiB

W0 d=4 floor: graph traversal with the distance taken out 1.18x

closed-loop · collection=bench0 · client -p 1 -t 2 -c 1 (-t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	4,824	4,075	1.18x
client p50	205 µs	242 µs	1.18x
client p99	250 µs	300 µs	1.20x
client p99.9	305 µs	364 µs	1.19x
server p50	112 µs	133 µs	1.19x
server p99	122 µs	169 µs	1.38x
wall clock	10.4 s	12.3 s	1.18x
queries sent	50,000	50,000
cpu	7.1 s	12.6 s	1.78x
cpu, % of wall	68%	103%
waiting for a core	13.1 ms	3.7 ms	0.28x
migrations	0	146
faults min/maj	0 / 0	377 / 0
peak RSS	144.6 MiB	688.1 MiB	4.76x
disk written	0	0
disk read	0	0

1.18x is 1.60x less work per query x 0.66x cores busy during the row (0.66 against 1.00) x 1.00x clock x 1.12x counted on-CPU share: the middle term is occupancy, not search speed d=4 floor: this measures graph traversal and heaps with the distance taken out, not request plumbing (findings 28: 76% of the server's CPU was HNSW search, 24% kernel; the RPC path is a small remainder)

W1 ingest throughput

upload · collection=bench1 · client -p 8 -t 8 -c 1 (-c bfb default) · n=100,000

metric	strawmANN	Qdrant
wall clock	2.3 s	3.2 s	1.37x
of which upload	0.3 s	1.1 s	3.86x
cpu	0.6 s	7.1 s	11.25x
cpu, % of wall	27%	222%
waiting for a core	0.2 ms	2,458.8 ms	13,114.14x
migrations	0	3,176
faults min/maj	56,822 / 0	50,609 / 118,243
peak RSS	321.8 MiB	1,013.4 MiB	3.15x
disk written	195.4 MiB	436.5 MiB
disk read	0	768.0 KiB

W2 index build time

upload · collection=bench2 · client -p 8 -t 8 -c 1 (-c bfb default) · n=100,000

metric	strawmANN	Qdrant
wall clock	11.4 s	16.4 s	1.45x
of which upload	0.3 s	1.2 s	4.16x
of which index wait	9.0 s	13.0 s	1.44x
time to Green	9.0 s	13.0 s	1.44x
cpu	56.5 s	71.3 s	1.26x
cpu, % of wall	497%	433%
waiting for a core	27.6 ms	3,346.4 ms	121.08x
migrations	0	3,685
faults min/maj	62,863 / 0	88,140 / 160,650
peak RSS	346.5 MiB	1.9 GiB	5.48x
disk written	195.4 MiB	843.1 MiB
disk read	0	11.9 MiB

W3 search, fp32, single query parity

ef=128 · closed-loop · collection=bench2 · client -p 1 -t 2 -c 1 (-t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	1,646	1,637	parity
recall@10	0.9940	0.9906	~equal
client p50	574 µs	611 µs	1.06x
client p99	966 µs	790 µs	0.82x
client p99.9	1.07 ms	860 µs	0.80x
server p50	471 µs	511 µs	1.08x
server p99	852 µs	683 µs	0.80x
wall clock	30.4 s	30.6 s	~equal
queries sent	50,000	50,000
cpu	26.7 s	31.4 s	1.18x
cpu, % of wall	88%	103%
waiting for a core	6.9 ms	3.7 ms	0.54x
migrations	0	144
faults min/maj	6 / 0	363 / 0
peak RSS	346.5 MiB	1.9 GiB	5.48x
disk written	0	0
disk read	0	0

within the ±2.3% band this dataset's noise floor puts on W3: no measured difference, not a small one

W4 search, saturating (closed loop) parity

ef=128 · closed-loop · collection=bench2 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	10,800	10,924	parity
recall@10	0.9940	0.9906	~equal
client p50	5.90 ms	5.36 ms	0.91x
client p99	6.54 ms	14.89 ms	2.28x
client p99.9	19.21 ms	26.91 ms	1.40x
server p50	5.80 ms	4.35 ms	0.75x
server p99	6.42 ms	11.31 ms	1.76x
wall clock	4.7 s	4.6 s	~equal
queries sent	50,000	50,000
cpu	34.1 s	35.6 s	1.04x
cpu, % of wall	730%	771%
waiting for a core	2.5 ms	82,694.5 ms	32,895.25x
migrations	0	15,794
faults min/maj	65 / 0	2,042 / 0
peak RSS	346.5 MiB	1.9 GiB	5.48x
disk written	0	0
disk read	0	0

within the ±2.0% band this dataset's noise floor puts on W4: no measured difference, not a small one

W4-sat50 search, fixed rate at 50% of saturation (open loop) offered

ef=128 · open-loop · offered=5,489/s (50% of saturation, pinned reference 10,978 qps) · collection=bench2 · client -p n/a under --rps -t 16 -c 2 · n=200,000

metric	strawmANN	Qdrant
queries/second	5,489	5,489	offered
recall@10	0.9940	0.9906	~equal
client p50	696 µs	856 µs	1.23x
client p99	1.34 ms	1.95 ms	1.45x
client p99.9	1.59 ms	2.39 ms	1.51x
server p50	583 µs	694 µs	1.19x
server p99	1.19 ms	1.74 ms	1.46x
wall clock	36.6 s	36.6 s	~equal
queries sent	200,000	200,000
cpu	117.4 s	148.1 s	1.26x
cpu, % of wall	321%	405%
waiting for a core	14.9 ms	119,429.2 ms	8,035.02x
migrations	0	439,062
faults min/maj	8 / 0	314 / 0
peak RSS	346.5 MiB	1.9 GiB	5.48x
disk written	0	0
disk read	0	0

W4-sat70 search, fixed rate at 70% of saturation (open loop) offered

ef=128 · open-loop · offered=7,685/s (70% of saturation, pinned reference 10,978 qps) · collection=bench2 · client -p n/a under --rps -t 16 -c 2 · n=200,000

metric	strawmANN	Qdrant
queries/second	7,685	7,685	offered
recall@10	0.9940	0.9906	~equal
client p50	861 µs	942 µs	1.09x
client p99	1.89 ms	2.20 ms	1.16x
client p99.9	2.21 ms	3.30 ms	1.50x
server p50	742 µs	757 µs	1.02x
server p99	1.70 ms	1.89 ms	1.11x
wall clock	26.2 s	26.2 s	~equal
queries sent	200,000	200,000
cpu	131.2 s	153.0 s	1.17x
cpu, % of wall	501%	585%
waiting for a core	3.7 ms	169,227.7 ms	45,936.04x
migrations	0	452,191
faults min/maj	28 / 0	587 / 0
peak RSS	346.5 MiB	1.9 GiB	5.48x
disk written	0	0
disk read	0	0

W4-sat90 search, fixed rate at 90% of saturation (open loop) offered

ef=128 · open-loop · offered=9,880/s (90% of saturation, pinned reference 10,978 qps) · collection=bench2 · client -p n/a under --rps -t 16 -c 2 · n=200,000

metric	strawmANN	Qdrant
queries/second	9,879	9,879	offered
recall@10	0.9940	0.9906	~equal
client p50	913 µs	1.33 ms	1.46x
client p99	2.14 ms	4.91 ms	2.30x
client p99.9	4.15 ms	8.14 ms	1.96x
server p50	790 µs	976 µs	1.23x
server p99	1.93 ms	3.60 ms	1.86x
wall clock	20.4 s	20.4 s	~equal
queries sent	200,000	200,000
cpu	126.6 s	139.2 s	1.10x
cpu, % of wall	621%	683%
waiting for a core	15.2 ms	122,345.4 ms	8,051.04x
migrations	0	178,031
faults min/maj	56 / 0	1,282 / 0
peak RSS	346.5 MiB	1.9 GiB	5.48x
disk written	0	0
disk read	0	0

W5 search batched (16 distinct dataset queries per request) 2.83x

ef=128 · closed-loop · collection=bench2 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	10,643	3,754	2.83x
recall@10	0.9940	0.9906	~equal
client p50	3.00 ms	8.50 ms	2.83x
client p99	3.52 ms	9.40 ms	2.67x
client p99.9	3.69 ms	9.87 ms	2.68x
server p50	2.81 ms	8.00 ms	2.85x
server p99	3.32 ms	8.86 ms	2.67x
wall clock	4.7 s	13.3 s	2.83x
queries sent	50,000	50,000
cpu	33.0 s	26.4 s	0.80x
cpu, % of wall	702%	198%
waiting for a core	2.1 ms	4.7 ms	2.22x
migrations	0	1,355
faults min/maj	0 / 0	585 / 0
peak RSS	346.5 MiB	1.9 GiB	5.48x
disk written	0	0
disk read	0	0

per-batch latency (16 queries/request); queries: dataset, random-sample 2.83x is 0.78x less work per query x 3.58x cores busy during the row (7.04 against 1.97) x 1.00x clock x 1.01x counted on-CPU share: the middle term is occupancy, not search speed batched 16, so qps and rps differ by 16x; the ratio uses qps

W6-upload scalar quantization: load

upload · collection=bench6 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=100,000

metric	strawmANN	Qdrant
wall clock	10.4 s	11.2 s	1.07x
of which upload	0.3 s	1.0 s	3.16x
of which index wait	8.0 s	8.0 s	~equal
time to Green	8.0 s	8.0 s	~equal
cpu	40.0 s	34.5 s	0.86x
cpu, % of wall	385%	309%
waiting for a core	17.0 ms	1,680.1 ms	98.65x
migrations	0	2,722
faults min/maj	128,700 / 0	122,288 / 160,677
peak RSS	877.6 MiB	2.5 GiB	2.96x
disk written	195.4 MiB	992.2 MiB
disk read	0	12.7 MiB

W6 quantized: scalar 1.24x

ef=128 · closed-loop · collection=bench6 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	5,507	4,443	1.24x
recall@10	0.9937	0.9908	~equal
client p50	355 µs	447 µs	1.26x
client p99	538 µs	534 µs	~equal
client p99.9	660 µs	632 µs	0.96x
server p50	264 µs	345 µs	1.30x
server p99	427 µs	412 µs	0.96x
wall clock	9.1 s	11.3 s	1.24x
queries sent	50,000	50,000
cpu	14.9 s	22.8 s	1.53x
cpu, % of wall	164%	202%
waiting for a core	11.8 ms	12.3 ms	1.04x
migrations	0	10,087
faults min/maj	0 / 0	430 / 0
peak RSS	877.6 MiB	2.5 GiB	2.96x
disk written	0	0
disk read	0	0

W6-ef32 SQ8 recall control, ef=32 (latency only)

ef=32 · closed-loop · collection=bench6 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000 · oversampling=3.2 · rescore=true

metric	strawmANN	Qdrant
queries/second	10,847	7,502	-
recall@10	0.9624	0.9504
client p50	180 µs	260 µs
client p99	247 µs	344 µs
client p99.9	280 µs	389 µs
server p50	91 µs	159 µs
server p99	140 µs	220 µs
wall clock	4.6 s	6.7 s
queries sent	50,000	50,000
cpu	6.2 s	13.1 s
cpu, % of wall	133%	196%
waiting for a core	2.1 ms	10.5 ms
migrations	0	10,584
faults min/maj	0 / 0	73 / 0
peak RSS	877.6 MiB	2.5 GiB
disk written	0	0
disk read	0	0

recall unequal: 0.9624 vs 0.9504; §7.4 compares at equal recall; the rescore pools match, so the encoders differ

W6-ef64 SQ8 recall control, ef=64 (latency only) 1.41x

ef=64 · closed-loop · collection=bench6 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000 · oversampling=6.4 · rescore=true

metric	strawmANN	Qdrant
queries/second	8,452	5,988	1.41x
recall@10	0.9850	0.9783	~equal
client p50	229 µs	327 µs	1.43x
client p99	324 µs	437 µs	1.35x
client p99.9	369 µs	491 µs	1.33x
server p50	142 µs	226 µs	1.59x
server p99	218 µs	317 µs	1.45x
wall clock	6.0 s	8.4 s	1.41x
queries sent	50,000	50,000
cpu	9.0 s	16.6 s	1.85x
cpu, % of wall	151%	198%
waiting for a core	1.5 ms	12.3 ms	8.09x
migrations	0	9,873
faults min/maj	0 / 0	87 / 0
peak RSS	877.6 MiB	2.5 GiB	2.96x
disk written	0	0
disk read	0	0

1.41x is 1.72x less work per query x 0.77x cores busy during the row (1.51 against 1.97) x 1.00x clock x 1.07x counted on-CPU share: the middle term is occupancy, not search speed

W6-ef128 SQ8 recall control, ef=128 (latency only) 1.21x

ef=128 · closed-loop · collection=bench6 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	5,402	4,469	1.21x
recall@10	0.9937	0.9908	~equal
client p50	360 µs	445 µs	1.24x
client p99	544 µs	531 µs	0.98x
client p99.9	605 µs	618 µs	1.02x
server p50	269 µs	343 µs	1.28x
server p99	435 µs	410 µs	0.94x
wall clock	9.3 s	11.2 s	1.21x
queries sent	50,000	50,000
cpu	15.3 s	22.6 s	1.48x
cpu, % of wall	165%	202%
waiting for a core	5.7 ms	11.4 ms	2.00x
migrations	0	10,050
faults min/maj	0 / 0	69 / 0
peak RSS	877.6 MiB	2.5 GiB	2.96x
disk written	0	0
disk read	0	0

W6-ef256 SQ8 recall control, ef=256 (latency only) 1.13x

ef=256 · closed-loop · collection=bench6 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000 · oversampling=25.6 · rescore=true

metric	strawmANN	Qdrant
queries/second	3,410	3,025	1.13x
recall@10	0.9975	0.9958	~equal
client p50	576 µs	662 µs	1.15x
client p99	868 µs	779 µs	0.90x
client p99.9	950 µs	842 µs	0.89x
server p50	483 µs	559 µs	1.16x
server p99	757 µs	663 µs	0.88x
wall clock	14.7 s	16.6 s	1.13x
queries sent	50,000	50,000
cpu	25.9 s	33.6 s	1.29x
cpu, % of wall	176%	203%
waiting for a core	1.3 ms	11.4 ms	8.80x
migrations	0	12,715
faults min/maj	0 / 0	172 / 0
peak RSS	877.6 MiB	2.5 GiB	2.96x
disk written	0	0
disk read	0	0

W6-ef512 SQ8 recall control, ef=512 (latency only)

ef=512 · closed-loop · collection=bench6 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000 · oversampling=51.2 · rescore=true

metric	strawmANN	Qdrant
queries/second	2,072	1,858	-
recall@10	0.9989	0.9980
client p50	964 µs	1.08 ms
client p99	1.31 ms	1.28 ms
client p99.9	1.53 ms	1.36 ms
server p50	868 µs	975 µs
server p99	1.20 ms	1.16 ms
wall clock	24.2 s	27.0 s
queries sent	50,000	50,000
cpu	44.8 s	54.6 s
cpu, % of wall	186%	202%
waiting for a core	5.8 ms	15.2 ms
migrations	0	17,661
faults min/maj	0 / 0	358 / 0
peak RSS	877.6 MiB	2.5 GiB
disk written	0	0
disk read	0	0

Qdrant pass 3 -21% against two passes that agree: one pass set the spread on this row, so a band built from it is not a noise band

W7-upload binary quantization: load

upload · collection=bench7 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=100,000

metric	strawmANN	Qdrant
wall clock	7.4 s	11.5 s	1.56x
of which upload	0.3 s	1.4 s	4.15x
of which index wait	5.0 s	8.0 s	1.60x
time to Green	5.0 s	8.0 s	1.60x
cpu	17.1 s	31.1 s	1.82x
cpu, % of wall	232%	269%
waiting for a core	7.8 ms	1,866.1 ms	238.27x
migrations	0	2,983
faults min/maj	64,793 / 0	107,187 / 196,234
peak RSS	974.0 MiB	3.1 GiB	3.22x
disk written	195.4 MiB	973.9 MiB
disk read	0	18.7 MiB

W7 quantized: binary + oversampling 1.25x

ef=128 · closed-loop · collection=bench7 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	5,769	4,616	1.25x
recall@10	0.9150	0.9162	~equal
client p50	341 µs	423 µs	1.24x
client p99	469 µs	550 µs	1.17x
client p99.9	629 µs	607 µs	0.97x
server p50	249 µs	319 µs	1.28x
server p99	360 µs	424 µs	1.18x
wall clock	8.7 s	10.9 s	1.25x
queries sent	50,000	50,000
cpu	14.1 s	21.7 s	1.54x
cpu, % of wall	162%	200%
waiting for a core	9.1 ms	13.7 ms	1.51x
migrations	0	10,273
faults min/maj	1 / 0	440 / 0
peak RSS	974.0 MiB	3.1 GiB	3.22x
disk written	0	0
disk read	0	0

W7-2bit-upload binary quantization, 2 bits: load

upload · collection=bench7b2 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=100,000

metric	strawmANN	Qdrant
wall clock	8.4 s	11.6 s	1.38x
of which upload	0.3 s	1.3 s	3.90x
of which index wait	6.0 s	8.0 s	1.33x
time to Green	6.0 s	8.0 s	1.33x
cpu	19.0 s	28.9 s	1.52x
cpu, % of wall	226%	250%
waiting for a core	7.4 ms	1,449.4 ms	196.07x
migrations	1	2,745	2,745.00x
faults min/maj	66,716 / 0	100,202 / 180,734
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	195.4 MiB	902.0 MiB
disk read	0	14.1 MiB

W7-2bit quantized: binary quantization, 2 bits + oversampling 1.18x

ef=128 · closed-loop · collection=bench7b2 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	5,465	4,636	1.18x
recall@10	0.9696	0.9674	~equal
client p50	361 µs	420 µs	1.16x
client p99	474 µs	548 µs	1.15x
client p99.9	537 µs	609 µs	1.13x
server p50	266 µs	317 µs	1.19x
server p99	364 µs	423 µs	1.16x
wall clock	9.2 s	10.8 s	1.18x
queries sent	50,000	50,000
cpu	14.9 s	21.6 s	1.45x
cpu, % of wall	163%	200%
waiting for a core	3.1 ms	11.9 ms	3.77x
migrations	0	10,027
faults min/maj	0 / 0	380 / 0
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	0	0
disk read	0	0

W7-1p5bit-upload binary quantization, 1.5 bits: load

upload · collection=bench7b15 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=100,000

metric	strawmANN	Qdrant
wall clock	8.4 s	12.5 s	1.49x
of which upload	0.3 s	1.3 s	4.14x
of which index wait	6.0 s	9.0 s	1.50x
time to Green	6.0 s	9.0 s	1.50x
cpu	19.4 s	33.2 s	1.71x
cpu, % of wall	232%	266%
waiting for a core	7.5 ms	1,604.3 ms	212.53x
migrations	1	2,756	2,756.00x
faults min/maj	66,708 / 0	100,267 / 179,982
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	195.4 MiB	903.3 MiB
disk read	0	15.6 MiB

W7-1p5bit quantized: binary quantization, 1.5 bits + oversampling 1.22x

ef=128 · closed-loop · collection=bench7b15 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	5,525	4,538	1.22x
recall@10	0.9475	0.9477	~equal
client p50	357 µs	431 µs	1.21x
client p99	477 µs	563 µs	1.18x
client p99.9	541 µs	621 µs	1.15x
server p50	264 µs	328 µs	1.24x
server p99	367 µs	437 µs	1.19x
wall clock	9.1 s	11.1 s	1.22x
queries sent	50,000	50,000
cpu	14.7 s	22.2 s	1.51x
cpu, % of wall	162%	201%
waiting for a core	3.4 ms	15.1 ms	4.38x
migrations	0	10,269
faults min/maj	0 / 0	463 / 0
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	0	0
disk read	0	0

W8-upload PQ: load

upload · collection=bench8 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=100,000

metric	strawmANN	Qdrant
wall clock	47.5 s	48.2 s	~equal
of which upload	0.3 s	1.0 s	3.04x
of which index wait	45.2 s	45.1 s	~equal
time to Green	45.2 s	45.1 s	~equal
cpu	342.9 s	271.2 s	0.79x
cpu, % of wall	722%	563%
waiting for a core	104.2 ms	1,586.7 ms	15.23x
migrations	0	8,238
faults min/maj	121,701 / 0	92,993 / 126,801
peak RSS	1.4 GiB	3.1 GiB	2.15x
disk written	195.4 MiB	814.3 MiB
disk read	0	11.6 MiB

W8 quantized: PQ 1.19x

ef=128 · closed-loop · collection=bench8 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	3,440	2,888	1.19x
recall@10	0.9927	0.9863	~equal
client p50	580 µs	692 µs	1.19x
client p99	726 µs	816 µs	1.12x
client p99.9	780 µs	897 µs	1.15x
server p50	483 µs	576 µs	1.19x
server p99	616 µs	679 µs	1.10x
wall clock	14.6 s	17.4 s	1.19x
queries sent	50,000	50,000
cpu	25.8 s	35.1 s	1.36x
cpu, % of wall	177%	202%
waiting for a core	1.4 ms	18.7 ms	13.20x
migrations	0	12,396
faults min/maj	0 / 0	724 / 0
peak RSS	1.4 GiB	3.1 GiB	2.15x
disk written	0	0
disk read	0	0

W14-1bit-upload TurboQuant, 1 bit: load

upload · collection=bench14t1 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=100,000

metric	strawmANN	Qdrant
wall clock	12.4 s	12.7 s	1.02x
of which upload	0.3 s	1.5 s	4.61x
of which index wait	10.0 s	9.0 s	0.90x
time to Green	10.0 s	9.0 s	0.90x
cpu	57.8 s	33.9 s	0.59x
cpu, % of wall	466%	267%
waiting for a core	25.7 ms	1,597.1 ms	62.08x
migrations	0	3,844
faults min/maj	67,661 / 0	99,537 / 179,408
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	195.4 MiB	898.4 MiB
disk read	0	16.1 MiB

W14-1bit quantized: TurboQuant, 1 bit + oversampling parity

ef=128 · closed-loop · collection=bench14t1 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	4,612	4,570	parity
recall@10	0.9895	0.9880	~equal
client p50	433 µs	429 µs	~equal
client p99	538 µs	554 µs	1.03x
client p99.9	604 µs	610 µs	~equal
server p50	330 µs	326 µs	~equal
server p99	425 µs	426 µs	~equal
wall clock	10.9 s	11.0 s	~equal
queries sent	50,000	50,000
cpu	18.2 s	22.1 s	1.21x
cpu, % of wall	167%	201%
waiting for a core	11.2 ms	13.4 ms	1.19x
migrations	0	10,473
faults min/maj	0 / 0	364 / 0
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	0	0
disk read	0	0

within the ±2.0% band this dataset's noise floor puts on W14-1bit: no measured difference, not a small one

W14-1p5bit-upload TurboQuant, 1.5 bits: load

upload · collection=bench14t15 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=100,000

metric	strawmANN	Qdrant
wall clock	12.4 s	13.1 s	1.06x
of which upload	0.3 s	1.9 s	6.05x
of which index wait	10.0 s	9.0 s	0.90x
time to Green	10.0 s	9.0 s	0.90x
cpu	58.7 s	35.9 s	0.61x
cpu, % of wall	473%	274%
waiting for a core	27.5 ms	1,853.0 ms	67.32x
migrations	0	3,132
faults min/maj	69,155 / 0	99,394 / 163,226
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	195.4 MiB	880.6 MiB
disk read	0	15.2 MiB

W14-1p5bit quantized: TurboQuant, 1.5 bits + oversampling 1.07x

ef=128 · closed-loop · collection=bench14t15 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	4,356	4,090	1.07x
recall@10	0.9922	0.9900	~equal
client p50	458 µs	484 µs	1.06x
client p99	576 µs	655 µs	1.14x
client p99.9	631 µs	726 µs	1.15x
server p50	357 µs	381 µs	1.07x
server p99	465 µs	536 µs	1.15x
wall clock	11.5 s	12.3 s	1.06x
queries sent	50,000	50,000
cpu	19.5 s	24.7 s	1.27x
cpu, % of wall	169%	202%
waiting for a core	3.3 ms	9.1 ms	2.74x
migrations	0	11,740
faults min/maj	0 / 0	511 / 0
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	0	0
disk read	0	0

W14-2bit-upload TurboQuant, 2 bits: load

upload · collection=bench14t2 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=100,000

metric	strawmANN	Qdrant
wall clock	12.4 s	13.0 s	1.05x
of which upload	0.3 s	1.8 s	5.58x
of which index wait	10.0 s	9.0 s	0.90x
time to Green	10.0 s	9.0 s	0.90x
cpu	58.1 s	36.5 s	0.63x
cpu, % of wall	468%	281%
waiting for a core	27.2 ms	2,027.8 ms	74.53x
migrations	0	4,012
faults min/maj	71,668 / 0	100,282 / 163,988
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	195.4 MiB	889.9 MiB
disk read	0	12.8 MiB

W14-2bit quantized: TurboQuant, 2 bits + oversampling 1.19x

ef=128 · closed-loop · collection=bench14t2 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	5,464	4,596	1.19x
recall@10	0.9935	0.9900	~equal
client p50	362 µs	425 µs	1.17x
client p99	476 µs	551 µs	1.16x
client p99.9	519 µs	607 µs	1.17x
server p50	270 µs	321 µs	1.19x
server p99	368 µs	424 µs	1.15x
wall clock	9.2 s	10.9 s	1.19x
queries sent	50,000	50,000
cpu	15.2 s	21.9 s	1.44x
cpu, % of wall	165%	200%
waiting for a core	2.5 ms	15.0 ms	6.08x
migrations	0	9,711
faults min/maj	0 / 0	154 / 0
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	0	0
disk read	0	0

W14-4bit-upload TurboQuant, 4 bits: load

upload · collection=bench14t4 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=100,000

metric	strawmANN	Qdrant
wall clock	12.4 s	13.3 s	1.07x
of which upload	0.3 s	2.2 s	6.73x
of which index wait	10.0 s	9.0 s	0.90x
time to Green	10.0 s	9.0 s	0.90x
cpu	58.8 s	37.2 s	0.63x
cpu, % of wall	475%	279%
waiting for a core	26.8 ms	1,927.8 ms	71.83x
migrations	0	4,040
faults min/maj	79,674 / 0	113,836 / 167,086
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	195.4 MiB	912.4 MiB
disk read	0	11.7 MiB

W14-4bit quantized: TurboQuant, 4 bits + oversampling 1.11x

ef=128 · closed-loop · collection=bench14t4 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	5,086	4,596	1.11x
recall@10	0.9937	0.9904	~equal
client p50	387 µs	429 µs	1.11x
client p99	516 µs	557 µs	1.08x
client p99.9	572 µs	623 µs	1.09x
server p50	293 µs	327 µs	1.11x
server p99	404 µs	436 µs	1.08x
wall clock	9.9 s	10.9 s	1.11x
queries sent	50,000	50,000
cpu	16.3 s	21.9 s	1.34x
cpu, % of wall	165%	201%
waiting for a core	3.9 ms	15.2 ms	3.89x
migrations	0	9,940
faults min/maj	0 / 0	481 / 0
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	0	0
disk read	0	0

W9 exact / brute force parity

exact · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=2,000

metric	strawmANN	Qdrant
queries/second	311	311	parity
client p50	25.51 ms	26.05 ms	1.02x
client p99	35.49 ms	31.80 ms	0.90x
client p99.9	39.53 ms	36.92 ms	0.93x
server p50	25.08 ms	25.39 ms	~equal
server p99	35.06 ms	31.03 ms	0.89x
wall clock	6.4 s	6.4 s	~equal
queries sent	2,000	2,000
cpu	45.4 s	49.8 s	1.10x
cpu, % of wall	704%	772%
waiting for a core	4.3 ms	7,100.6 ms	1,662.09x
migrations	0	5,751
faults min/maj	0 / 0	80 / 0
peak RSS	1.4 GiB	3.1 GiB	2.15x
disk written	0	0
disk read	0	0

within the ±12.2% band this dataset's noise floor puts on W9: no measured difference, not a small one brute force over the whole collection, no index involved

W10-ef32 recall control, ef=32 (latency only)

ef=32 · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=100,000

metric	strawmANN	Qdrant
queries/second	27,901	18,065	-
recall@10	0.9630	0.9513
client p50	279 µs	411 µs
client p99	403 µs	801 µs
client p99.9	499 µs	1.11 ms
server p50	185 µs	274 µs
server p99	293 µs	577 µs
wall clock	3.7 s	2.8 s
queries sent	100,000	50,000
cpu	21.4 s	16.1 s
cpu, % of wall	584%	575%
waiting for a core	3.3 ms	6,265.7 ms
migrations	0	65,474
faults min/maj	0 / 0	856 / 0
peak RSS	1.4 GiB	3.1 GiB
disk written	0	0
disk read	0	0

strawmANN: -n 100000 (2x the table's 50000) recall unequal: 0.9630 vs 0.9513; §7.4 compares at equal recall

W10-ef64 recall control, ef=64 (latency only) 1.42x

ef=64 · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	17,667	12,438	1.42x
recall@10	0.9852	0.9786	~equal
client p50	447 µs	595 µs	1.33x
client p99	644 µs	1.19 ms	1.85x
client p99.9	798 µs	1.55 ms	1.95x
server p50	353 µs	441 µs	1.25x
server p99	543 µs	983 µs	1.81x
wall clock	2.9 s	4.1 s	1.41x
queries sent	50,000	50,000
cpu	18.9 s	24.0 s	1.27x
cpu, % of wall	659%	592%
waiting for a core	2.2 ms	14,521.1 ms	6,656.05x
migrations	0	111,342
faults min/maj	0 / 0	524 / 0
peak RSS	1.4 GiB	3.1 GiB	2.15x
disk written	0	0
disk read	0	0

W10-ef128 recall control, ef=128 (latency only) 1.27x

ef=128 · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	10,487	8,268	1.27x
recall@10	0.9940	0.9906	~equal
client p50	757 µs	885 µs	1.17x
client p99	1.14 ms	1.84 ms	1.62x
client p99.9	1.29 ms	2.35 ms	1.82x
server p50	660 µs	713 µs	1.08x
server p99	1.04 ms	1.57 ms	1.52x
wall clock	4.8 s	6.1 s	1.27x
queries sent	50,000	50,000
cpu	33.3 s	36.6 s	1.10x
cpu, % of wall	693%	602%
waiting for a core	3.8 ms	26,270.7 ms	6,914.82x
migrations	0	117,455
faults min/maj	0 / 0	214 / 0
peak RSS	1.4 GiB	3.1 GiB	2.15x
disk written	0	0
disk read	0	0

W10-ef256 recall control, ef=256 (latency only) 1.12x

ef=256 · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	6,012	5,380	1.12x
recall@10	0.9976	0.9956	~equal
client p50	1.32 ms	1.36 ms	1.03x
client p99	2.02 ms	2.78 ms	1.38x
client p99.9	2.28 ms	3.48 ms	1.53x
server p50	1.22 ms	1.16 ms	0.95x
server p99	1.91 ms	2.42 ms	1.27x
wall clock	8.4 s	9.3 s	1.12x
queries sent	50,000	50,000
cpu	58.8 s	58.6 s	~equal
cpu, % of wall	704%	628%
waiting for a core	4.7 ms	38,470.1 ms	8,170.80x
migrations	0	143,422
faults min/maj	0 / 0	179 / 0
peak RSS	1.4 GiB	3.1 GiB	2.15x
disk written	0	0
disk read	0	0

W10-ef512 recall control, ef=512 (latency only) parity

ef=512 · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	3,423	3,373	parity
recall@10	0.9989	0.9979	~equal
client p50	2.33 ms	2.24 ms	0.96x
client p99	3.56 ms	4.18 ms	1.17x
client p99.9	4.06 ms	5.27 ms	1.30x
server p50	2.23 ms	2.00 ms	0.90x
server p99	3.45 ms	3.80 ms	1.10x
wall clock	14.6 s	14.9 s	~equal
queries sent	50,000	50,000
cpu	103.6 s	96.4 s	0.93x
cpu, % of wall	708%	649%
waiting for a core	8.4 ms	53,106.6 ms	6,359.74x
migrations	0	169,278
faults min/maj	0 / 0	293 / 0
peak RSS	1.4 GiB	3.1 GiB	2.15x
disk written	0	0
disk read	0	0

within the ±2.0% band this dataset's noise floor puts on W10-ef512: no measured difference, not a small one

W12-upload filtered search: load 100,000 with payloads

upload · collection=bench12 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=100,000

metric	strawmANN	Qdrant
wall clock	12.6 s	24.3 s	1.94x
of which upload	0.5 s	1.1 s	2.36x
of which index wait	10.0 s	21.1 s	2.10x
time to Green	10.0 s	21.1 s	2.10x
cpu	56.5 s	85.5 s	1.51x
cpu, % of wall	450%	351%
waiting for a core	24.9 ms	1,653.3 ms	66.34x
migrations	0	3,459
faults min/maj	63,572 / 0	112,773 / 154,555
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	195.4 MiB	837.4 MiB
disk read	0	22.1 MiB

W12-sel1 filtered search, one keyword (~1% of bench12)

ef=128 · closed-loop · collection=bench12 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	7,096	3,368	-
recall@10	1.0000	0.9999
client p50	271 µs	591 µs
client p99	449 µs	676 µs
client p99.9	698 µs	751 µs
server p50	184 µs	486 µs
server p99	335 µs	555 µs
wall clock	7.1 s	14.9 s
queries sent	50,000	50,000
cpu	11.2 s	29.7 s
cpu, % of wall	159%	200%
waiting for a core	10.8 ms	15.7 ms
migrations	0	9,230
faults min/maj	1 / 0	621 / 0
peak RSS	1.5 GiB	3.1 GiB
disk written	0	0
disk read	0	0

strawmANN pass 1 -16% against two passes that agree: one pass set the spread on this row, so a band built from it is not a noise band filtered search over a keyword payload index both engines were asked to build; whether each did is read back from the engine and refused beside the row when it did not. Recall is measured under the filter, against ground truth restricted to the matching set

W12-sel10 filtered search, any of 10 keywords (~10% of bench12)

ef=128 · closed-loop · collection=bench12 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	1,322	1,803	-
recall@10	1.0000	0.9854
client p50	1.51 ms	1.11 ms
client p99	1.64 ms	1.34 ms
client p99.9	1.70 ms	1.42 ms
server p50	1.41 ms	992 µs
server p99	1.53 ms	1.22 ms
wall clock	37.9 s	27.8 s
queries sent	50,000	50,000
cpu	71.2 s	56.2 s
cpu, % of wall	188%	202%
waiting for a core	7.6 ms	13.6 ms
migrations	0	17,354
faults min/maj	0 / 0	381 / 0
peak RSS	1.5 GiB	3.1 GiB
disk written	0	0
disk read	0	0

recall unequal: 1.0000 vs 0.9854; §7.4 compares at equal recall filtered search over a keyword payload index both engines were asked to build; whether each did is read back from the engine and refused beside the row when it did not. Recall is measured under the filter, against ground truth restricted to the matching set

W12-sel1-ef32 filtered recall control, one keyword, ef=32 (latency only) 1.26x

ef=32 · closed-loop · collection=bench12 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	7,291	5,774	1.26x
recall@10	1.0000	0.9934	~equal
client p50	269 µs	341 µs	1.27x
client p99	339 µs	413 µs	1.22x
client p99.9	560 µs	489 µs	0.87x
server p50	183 µs	239 µs	1.30x
server p99	241 µs	289 µs	1.20x
wall clock	6.9 s	8.7 s	1.26x
queries sent	50,000	50,000
cpu	11.0 s	17.2 s	1.56x
cpu, % of wall	160%	198%
waiting for a core	1.2 ms	12.5 ms	10.26x
migrations	0	9,810
faults min/maj	0 / 0	36 / 0
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	0	0
disk read	0	0

W12-sel1-ef64 filtered recall control, one keyword, ef=64 (latency only) 1.62x

ef=64 · closed-loop · collection=bench12 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	7,312	4,527	1.62x
recall@10	1.0000	0.9988	~equal
client p50	269 µs	438 µs	1.63x
client p99	335 µs	510 µs	1.52x
client p99.9	569 µs	559 µs	~equal
server p50	184 µs	336 µs	1.82x
server p99	242 µs	391 µs	1.61x
wall clock	6.9 s	11.1 s	1.61x
queries sent	50,000	50,000
cpu	11.0 s	22.2 s	2.01x
cpu, % of wall	160%	200%
waiting for a core	2.1 ms	13.0 ms	6.31x
migrations	0	9,451
faults min/maj	0 / 0	62 / 0
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	0	0
disk read	0	0

W12-sel1-ef128 filtered recall control, one keyword, ef=128 (latency only) 2.17x

ef=128 · closed-loop · collection=bench12 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	7,255	3,339	2.17x
recall@10	1.0000	0.9999	~equal
client p50	271 µs	596 µs	2.19x
client p99	331 µs	689 µs	2.08x
client p99.9	504 µs	761 µs	1.51x
server p50	185 µs	491 µs	2.65x
server p99	235 µs	564 µs	2.40x
wall clock	6.9 s	15.0 s	2.17x
queries sent	50,000	50,000
cpu	11.1 s	30.2 s	2.73x
cpu, % of wall	159%	201%
waiting for a core	1.9 ms	16.8 ms	8.85x
migrations	0	10,052
faults min/maj	0 / 0	251 / 0
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	0	0
disk read	0	0

2.17x is 2.72x less work per query x 0.80x cores busy during the row (1.60 against 2.00) x 1.00x clock: the middle term is occupancy, not search speed

W12-sel1-ef256 filtered recall control, one keyword, ef=256 (latency only)

ef=256 · closed-loop · collection=bench12 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	7,076	2,441	-
recall@10	1.0000	1.0000
client p50	273 µs	815 µs
client p99	372 µs	928 µs
client p99.9	629 µs	998 µs
server p50	186 µs	709 µs
server p99	270 µs	801 µs
wall clock	7.1 s	20.5 s
queries sent	50,000	50,000
cpu	11.3 s	41.3 s
cpu, % of wall	159%	201%
waiting for a core	2.2 ms	22.1 ms
migrations	0	10,027
faults min/maj	0 / 0	218 / 0
peak RSS	1.5 GiB	3.1 GiB
disk written	0	0
disk read	0	0

strawmANN pass 1 -39% against two passes that agree: one pass set the spread on this row, so a band built from it is not a noise band

W12-sel1-ef512 filtered recall control, one keyword, ef=512 (latency only) 4.04x

ef=512 · closed-loop · collection=bench12 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	6,833	1,693	4.04x
recall@10	1.0000	1.0000	~equal
client p50	284 µs	1.18 ms	4.14x
client p99	402 µs	1.33 ms	3.29x
client p99.9	629 µs	1.46 ms	2.32x
server p50	195 µs	1.07 ms	5.47x
server p99	283 µs	1.19 ms	4.22x
wall clock	7.4 s	29.6 s	4.02x
queries sent	50,000	50,000
cpu	11.7 s	59.5 s	5.07x
cpu, % of wall	159%	201%
waiting for a core	1.8 ms	41.6 ms	22.87x
migrations	0	11,588
faults min/maj	0 / 0	313 / 0
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	0	0
disk read	0	0

4.04x is 5.32x less work per query x 0.79x cores busy during the row (1.59 against 2.00) x 1.00x clock x 0.96x counted on-CPU share: the middle term is occupancy, not search speed

W12-sel10-ef32 filtered recall control, any of 10 keywords, ef=32 (latency only)

ef=32 · closed-loop · collection=bench12 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	3,212	3,980	-
recall@10	0.9808	0.8164
client p50	627 µs	495 µs
client p99	775 µs	638 µs
client p99.9	935 µs	716 µs
server p50	534 µs	387 µs
server p99	677 µs	522 µs
wall clock	15.6 s	12.6 s
queries sent	50,000	50,000
cpu	27.7 s	25.5 s
cpu, % of wall	178%	202%
waiting for a core	3.0 ms	10.8 ms
migrations	0	12,232
faults min/maj	0 / 0	195 / 0
peak RSS	1.5 GiB	3.1 GiB
disk written	0	0
disk read	0	0

recall unequal: 0.9808 vs 0.8164; §7.4 compares at equal recall

W12-sel10-ef64 filtered recall control, any of 10 keywords, ef=64 (latency only)

ef=64 · closed-loop · collection=bench12 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	2,077	2,780	-
recall@10	0.9932	0.9326
client p50	974 µs	714 µs
client p99	1.17 ms	886 µs
client p99.9	1.23 ms	979 µs
server p50	880 µs	603 µs
server p99	1.07 ms	765 µs
wall clock	24.1 s	18.0 s
queries sent	50,000	50,000
cpu	44.7 s	36.5 s
cpu, % of wall	185%	202%
waiting for a core	2.1 ms	13.1 ms
migrations	0	14,006
faults min/maj	0 / 0	292 / 0
peak RSS	1.5 GiB	3.1 GiB
disk written	0	0
disk read	0	0

recall unequal: 0.9932 vs 0.9326; §7.4 compares at equal recall

W12-sel10-ef128 filtered recall control, any of 10 keywords, ef=128 (latency only)

ef=128 · closed-loop · collection=bench12 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	1,319	1,800	-
recall@10	1.0000	0.9854
client p50	1.52 ms	1.11 ms
client p99	1.65 ms	1.35 ms
client p99.9	1.73 ms	1.45 ms
server p50	1.41 ms	994 µs
server p99	1.53 ms	1.23 ms
wall clock	37.9 s	27.8 s
queries sent	50,000	50,000
cpu	71.3 s	56.2 s
cpu, % of wall	188%	202%
waiting for a core	8.4 ms	16.4 ms
migrations	0	17,649
faults min/maj	0 / 0	260 / 0
peak RSS	1.5 GiB	3.1 GiB
disk written	0	0
disk read	0	0

recall unequal: 1.0000 vs 0.9854; §7.4 compares at equal recall

W12-sel10-ef256 filtered recall control, any of 10 keywords, ef=256 (latency only) 1.29x

ef=256 · closed-loop · collection=bench12 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	1,319	1,024	1.29x
recall@10	1.0000	0.9972	~equal
client p50	1.52 ms	1.95 ms	1.29x
client p99	1.64 ms	2.37 ms	1.44x
client p99.9	1.73 ms	2.50 ms	1.44x
server p50	1.41 ms	1.82 ms	1.29x
server p99	1.53 ms	2.23 ms	1.46x
wall clock	37.9 s	48.9 s	1.29x
queries sent	50,000	50,000
cpu	71.3 s	98.6 s	1.38x
cpu, % of wall	188%	202%
waiting for a core	5.8 ms	53.8 ms	9.24x
migrations	0	19,972
faults min/maj	0 / 0	756 / 0
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	0	0
disk read	0	0

W12-sel10-ef512 filtered recall control, any of 10 keywords, ef=512 (latency only) 2.16x

ef=512 · closed-loop · collection=bench12 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	1,320	612	2.16x
recall@10	1.0000	0.9993	~equal
client p50	1.51 ms	3.26 ms	2.15x
client p99	1.65 ms	3.90 ms	2.36x
client p99.9	1.72 ms	4.12 ms	2.39x
server p50	1.41 ms	3.08 ms	2.18x
server p99	1.53 ms	3.68 ms	2.41x
wall clock	37.9 s	81.8 s	2.16x
queries sent	50,000	50,000
cpu	71.2 s	162.6 s	2.28x
cpu, % of wall	188%	199%
waiting for a core	4.7 ms	91.6 ms	19.56x
migrations	0	21,393
faults min/maj	0 / 0	1,381 / 0
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	0	0
disk read	0	0

W13 scroll / pagination 1.09x

closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=200,000

metric	strawmANN	Qdrant
queries/second	50,282	46,273	1.09x
client p50	139 µs	159 µs	1.15x
client p99	235 µs	296 µs	1.26x
client p99.9	275 µs	440 µs	1.60x
server p50	11 µs	28 µs	2.58x
server p99	16 µs	93 µs	5.82x
wall clock	4.1 s	2.2 s	0.54x
queries sent	200,000	100,000
cpu	5.6 s	9.1 s	1.62x
cpu, % of wall	137%	409%
waiting for a core	2.5 ms	922.2 ms	369.09x
migrations	0	113,126
faults min/maj	0 / 0	622 / 0
peak RSS	1.5 GiB	3.1 GiB	2.03x
disk written	0	0
disk read	0	0

strawmANN: -n 200000 (4x the table's 50000) Qdrant: -n 100000 (2x the table's 50000) strawmANN: 0.13 ms of the client's 0.14 ms p50 is not server time (13x), so 92% of this row is the load generator and the socket Qdrant: 0.13 ms of the client's 0.16 ms p50 is not server time (6x), so 82% of this row is the load generator and the socket 1.09x is 3.29x less work per query x 0.33x cores busy during the row (1.38 against 4.24) x 1.00x clock x 1.01x counted on-CPU share: the middle term is occupancy, not search speed

W11-steady mixed read/write below the rebuild threshold: search bench2 while 5,000 synthetic points append

ef=128 · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=330,611

metric	strawmANN	Qdrant
queries/second	10,102	7,170	-
client p50	771 µs	932 µs
client p99	1.34 ms	2.26 ms
client p99.9	3.00 ms	3.83 ms
server p50	674 µs	744 µs
server p99	1.14 ms	1.87 ms
wall clock	33.3 s	44.4 s
queries sent	330,611	330,611
cpu	228.5 s	277.1 s
cpu, % of wall	685%	624%
waiting for a core	3,877.6 ms	247,440.9 ms
migrations	16	894,806
faults min/maj	23,682 / 0	376,849 / 480,766
peak RSS	1.5 GiB	3.3 GiB
disk written	9.9 MiB	3.7 GiB
disk read	0	95.2 MiB

strawmANN: qps over the 25.0 s of the append the search saw (10,152 over the whole search); append 200 points/s Qdrant: qps over the 25.0 s of the append the search saw (7,523 over the whole search); append 200 points/s search-during-write; no recall join

W11 mixed read/write: search bench2 while 20,000 synthetic points append (runs last)

ef=128 · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=863,509

metric	strawmANN	Qdrant
queries/second	9,526	7,151	-
client p50	809 µs	937 µs
client p99	1.98 ms	2.25 ms
client p99.9	3.91 ms	6.85 ms
server p50	711 µs	753 µs
server p99	1.43 ms	1.87 ms
wall clock	92.1 s	116.7 s
queries sent	863,509	863,509
cpu	634.6 s	720.3 s
cpu, % of wall	689%	617%
waiting for a core	18,357.3 ms	572,136.9 ms
migrations	60	2,221,296
faults min/maj	89,971 / 0	658,653 / 439,853
peak RSS	1.5 GiB	5.0 GiB
disk written	39.4 MiB	5.8 GiB
disk read	0	147.3 MiB

strawmANN: qps over the 66.7 s of the append the search saw (9,566 over the whole search); append 300 points/s Qdrant: qps over the 66.7 s of the append the search saw (7,528 over the whole search); append 300 points/s search-during-write; no recall join

Host discipline the §7.1 gate, per check, and ambient load per row

Ambient load per row

The §7.1 gate checks the machine once, at the start. Colour is the engine, as everywhere else; a hatched bar with a red edge is a row that had another process on the box while it ran, which the gate cannot see.

strawmANN gate pass

environment hash 373e8531fd3d1bf6

okAMD Ryzen AI 9 HX PRO 370 w/ Radeon 890M (24 logical)

okgovernor=performance

okboost disabled

oksmt=on

declared rather than required (§7.1 asks for SMT on or off, not

for off), and hashed, so these rows never share a chart with

SMT-off ones

cpu N and cpu N+12 share a core (12 pairs);

a --server-cpus or --client-cpus set holding both cpus of a

pair measures the engine on half the cores it names

okprofile=as-deployed (isolcpus=none nohz_full=none)

the scheduler is running as it ships, so these numbers describe a

deployment rather than the engine in isolation; they may not be

compared against an `isolated` run, and the environment hash

enforces that

okthp=madvise numa_nodes=1 numa_balancing=0

okperf_event_paranoid=-1

oktask_delayacct=1, so time blocked on a device is measurable

okperf at /usr/bin/perf, so `workloads.py run --perf` can attach

okcgroup io controller reaches the engines' own scope, so block-layer read/write operations are measurable

okzig=0.16.0

okmachine is quiescent (load 0.1 over 24 cores)

measurement profile: as-deployed

Qdrant gate pass

environment hash 373e8531fd3d1bf6

okAMD Ryzen AI 9 HX PRO 370 w/ Radeon 890M (24 logical)

okgovernor=performance

okboost disabled

oksmt=on

declared rather than required (§7.1 asks for SMT on or off, not

for off), and hashed, so these rows never share a chart with

SMT-off ones

cpu N and cpu N+12 share a core (12 pairs);

a --server-cpus or --client-cpus set holding both cpus of a

pair measures the engine on half the cores it names

okprofile=as-deployed (isolcpus=none nohz_full=none)

the scheduler is running as it ships, so these numbers describe a

deployment rather than the engine in isolation; they may not be

compared against an `isolated` run, and the environment hash

enforces that

okthp=madvise numa_nodes=1 numa_balancing=0

okperf_event_paranoid=-1

oktask_delayacct=1, so time blocked on a device is measurable

okperf at /usr/bin/perf, so `workloads.py run --perf` can attach

okcgroup io controller reaches the engines' own scope, so block-layer read/write operations are measurable

okzig=0.16.0

busy: Xorg(7%)

measurement profile: as-deployed

What this does not establish

No Qdrant developer has used this tool. Nothing here has been run, checked against something already known, or disagreed with by anyone outside the project: every number has one author and one reviewer, and they are the same person. A result you can refute is more useful to us than one you accept.

It has never been pointed at a real Qdrant regression. The harness detects a known ISA slowdown and correctly reports no difference on a control row. That is internal consistency, not evidence it would flag a regression in your tree or stay quiet through a refactor. bench/harness/qdrant_ab.py exists to run two of your commits through it blind, and that experiment has not been done.

Part of the measured throughput gap is kernel width, not architecture. Read from a dev checkout: Qdrant's distance path tops out at four 256-bit accumulators and has no AVX-512, while these kernels use 512-bit ones on a machine that has them. That is a real difference and it is not the same claim as "the design is faster".

The engines are not equivalent, by construction. §1 removes sharding, replication, consensus, snapshots, sparse vectors, multivectors and disk-resident operation — permanently. Payload storage and filtering were in that list until 2026-09-03 and are now built (M7), so W12 measures a filtered search with a keyword index on both engines. Anything still unbuilt answers UNIMPLEMENTED naming the construct and reports n/a, never a silent degradation. A strawman that was not faster would mean it was badly built; the question is by how much, and where the model was wrong.

Reproducing this

Everything below runs from a clone. The engine has no dependencies; the analysis path is a uv project pinned by bench/uv.lock.

scripts/doctor.py                    # what can this machine measure?
bench/setup.py check                 # §7.1 host gate: governor, boost, isolation, idle
conformance/datasets/datasets.py fetch

# one engine at a time, §7.1
zig build -Doptimize=ReleaseFast
# --connections matters: W4 opens ~32 sockets (-t 16 -c 2) and a server with
# fewer closes the excess before the HTTP/2 preface, which the client reports
# only as "transport error".
./zig-out/bin/strawmann --port 6334 --connections 64
bench/harness/workloads.py run http://localhost:6334 strawmann --sink --report

# the other half of every number: recall, and the licence to compare at all
cd conformance
cargo run --release -- relevance --engine http://localhost:6334 --label strawmann \
  --base $DATA/laion-small-clip/base.fbin --queries $DATA/laion-small-clip/queries.fbin \
  --ground-truth $DATA/laion-small-clip/gt/laion-small-clip.cosine.k100.gt.json --metric cosine \
  --ef 32 64 128 256 512 --json ../bench/results/strawmann/recall.json
cargo run --release -- differ --strawmann http://localhost:6334 --qdrant http://localhost:6434 \
  --base $DATA/laion-small-clip/base.fbin --queries $DATA/laion-small-clip/queries.fbin \
  --ground-truth $DATA/laion-small-clip/gt/laion-small-clip.cosine.k100.gt.json \
  --json ../bench/results/strawmann/conformance.json

uv run --project bench bench/harness/report.py strawmann qdrant

Disagreements are the point. docs/spec.md is what all of this is measured against, docs/bugs.md lists the measurement bugs found so far (each produced a plausible wrong number rather than a failure), and docs/validation.md records what has and has not been checked.

Generated by bench/harness/report.py from bench/results/. Self-contained: charts are embedded, nothing is fetched at open time. Regenerate with uv run --project bench bench/harness/report.py sm-laion-perf-1003 qd-laion-perf-1003.

Summary

Throughput against recall

Throughput

Throughput by workload

Recall

At matched recall

At matched recall, filtered to 10%

Recall by ef

What each quantization costs in recall

Latency

Ingest and index build

Where the load time goes

Memory and disk

Appendix

Dataset

Host

W10: throughput against ef

Per workload

Time spent waiting for a core

Peak memory per row

Time stalled on I/O

What one query cost, in cycles

Demand loads from DRAM per query

TLB walks per query

Instructions per cycle

Branch mispredictions per 1k instructions

W0-upload transport floor: load

W0 d=4 floor: graph traversal with the distance taken out 1.18x

W1 ingest throughput

W2 index build time

W3 search, fp32, single query parity

W4 search, saturating (closed loop) parity

W4-sat50 search, fixed rate at 50% of saturation (open loop) offered

W4-sat70 search, fixed rate at 70% of saturation (open loop) offered

W4-sat90 search, fixed rate at 90% of saturation (open loop) offered

W5 search batched (16 distinct dataset queries per request) 2.83x

W6-upload scalar quantization: load

W6 quantized: scalar 1.24x

W6-ef32 SQ8 recall control, ef=32 (latency only)

W6-ef64 SQ8 recall control, ef=64 (latency only) 1.41x

W6-ef128 SQ8 recall control, ef=128 (latency only) 1.21x

W6-ef256 SQ8 recall control, ef=256 (latency only) 1.13x

W6-ef512 SQ8 recall control, ef=512 (latency only)

W7-upload binary quantization: load

W7 quantized: binary + oversampling 1.25x

W7-2bit-upload binary quantization, 2 bits: load

W7-2bit quantized: binary quantization, 2 bits + oversampling 1.18x

W7-1p5bit-upload binary quantization, 1.5 bits: load

W7-1p5bit quantized: binary quantization, 1.5 bits + oversampling 1.22x

W8-upload PQ: load

W8 quantized: PQ 1.19x

W14-1bit-upload TurboQuant, 1 bit: load

W14-1bit quantized: TurboQuant, 1 bit + oversampling parity

W14-1p5bit-upload TurboQuant, 1.5 bits: load

W14-1p5bit quantized: TurboQuant, 1.5 bits + oversampling 1.07x

W14-2bit-upload TurboQuant, 2 bits: load

W14-2bit quantized: TurboQuant, 2 bits + oversampling 1.19x

W14-4bit-upload TurboQuant, 4 bits: load

W14-4bit quantized: TurboQuant, 4 bits + oversampling 1.11x

W9 exact / brute force parity

W10-ef32 recall control, ef=32 (latency only)

W10-ef64 recall control, ef=64 (latency only) 1.42x

W10-ef128 recall control, ef=128 (latency only) 1.27x

W10-ef256 recall control, ef=256 (latency only) 1.12x

W10-ef512 recall control, ef=512 (latency only) parity

W12-upload filtered search: load 100,000 with payloads

W12-sel1 filtered search, one keyword (~1% of bench12)

W12-sel10 filtered search, any of 10 keywords (~10% of bench12)

W12-sel1-ef32 filtered recall control, one keyword, ef=32 (latency only) 1.26x

W12-sel1-ef64 filtered recall control, one keyword, ef=64 (latency only) 1.62x

W12-sel1-ef128 filtered recall control, one keyword, ef=128 (latency only) 2.17x

W12-sel1-ef256 filtered recall control, one keyword, ef=256 (latency only)

W12-sel1-ef512 filtered recall control, one keyword, ef=512 (latency only) 4.04x

W12-sel10-ef32 filtered recall control, any of 10 keywords, ef=32 (latency only)

W12-sel10-ef64 filtered recall control, any of 10 keywords, ef=64 (latency only)

W12-sel10-ef128 filtered recall control, any of 10 keywords, ef=128 (latency only)

W12-sel10-ef256 filtered recall control, any of 10 keywords, ef=256 (latency only) 1.29x

W12-sel10-ef512 filtered recall control, any of 10 keywords, ef=512 (latency only) 2.16x

W13 scroll / pagination 1.09x

W11-steady mixed read/write below the rebuild threshold: search bench2 while 5,000 synthetic points append

Recall by `ef`