Benchmarks

tools/bench/openads_bench generates a synthetic 100 000-row DBF (ID N(8,0), TAG C(4), AMT N(8,2)) and times a fixed set of SQL workloads through the public ABI (AdsExecuteSQLDirect). Median of 5 repeats per workload, Release builds.

v0.4.x results — local in-process SQL (2026-05-06)

Workload (median ms)	Windows MSVC	Linux clang -O3	macOS AppleClang
stage 100 k-row DBF	63.5	57.9	34.0
SELECT COUNT(*)	297.7	42.0	103.9
WHERE TAG = 'AAAA'	303.7	48.3	108.4
SUM/AVG/MIN/MAX(AMT)	374.3	120.5	136.1
GROUP BY TAG	321.9	58.6	120.9
ORDER BY AMT LIMIT 10	668.0	165.4	260.5
DISTINCT TAG	598.4	95.2	213.4
BETWEEN 100 AND 500	314.1	63.7	114.4

Linux clang -O3 wins every SQL workload — roughly 7× faster than MSVC Release on the full-table count, 4× on the heavier ORDER BY. macOS Intel sits between the two.

Bench v2 — index-aware workloads (Windows MSVC, 100 k rows)

Workload (median ms)	ms
CREATE INDEX ID_IDX	38.0
WHERE ID = 50000 (post-index)	308.0
WHERE ID BETWEEN 10000 AND 20000	308.2
UNION ALL of two filtered selects	608.2
GROUP BY TAG HAVING COUNT(*) > 100	0.2

The indexed_eq row at ~308 ms ≈ seq_walk_where at ~315 ms exposes a known opportunity: the SQL planner currently does not push WHERE predicates into a matching CDX/NTX index. Closing that gap is a future milestone.

Bench v3 — AOF (Rushmore-style) results (rc12, 100 k rows)

AdsSetAOF parses + evaluates the cond, installs a per-record bitmap as a filter predicate that Skip / GoTop honour, and routes individual leaves through CDX / NTX index range scans whenever an open index has the leaf’s field as its key expression. AdsGetAOFOptLevel reports ADS_OPTIMIZED_FULL / PART / NONE based on per-leaf coverage. Sparse-bitmap navigation (M-AOF.5) lifts the visible-set walk from O(N) to O(M).

Cross-platform numbers, all on the same 100 000-row synthetic DBF, all medians of 5 repeats, all Release builds:

AOF workload	Win MSVC x64	Linux clang -O3	macOS AppleClang	OptLevel
AdsSetAOF("TAG='AAAA'"), no TAG index	593 ms	93 ms	210 ms	NONE
AdsSetAOF("TAG='AAAA'"), TAG indexed	323 ms	58 ms	119 ms	FULL
AdsSetAOF("TAG BETWEEN 'AAAA' AND 'CCCC'"), idx	24 ms	4.5 ms	9 ms	FULL

Speedups vs the unindexed full-scan baseline on the same host:

AOF workload	Win MSVC	Linux clang	macOS
AdsSetAOF("TAG='AAAA'"), TAG indexed	1.83×	1.61×	1.77×
AdsSetAOF("TAG BETWEEN 'AAAA' AND 'CCCC'")	24.4×	20.7×	23.4×

What drives the speedup:

1. AdsSetAOF itself becomes a range scan over the matching index instead of a full record decode + AST eval per row. That alone is the rc11 (M-AOF.4) gain.
2. Skip / GoTop walk only the matching records (sparse-bitmap navigation, M-AOF.5) instead of iterating every recno asking the predicate. That’s the rc12 gain — the textbook Rushmore “10-100×” speedup window for selective filters.

The eq-walk speedup at ~1.83× is bounded by the per-visible- record load_record_ decode cost (~80 µs × 3848 matching records ≈ 310 ms floor on Windows). Applications that don’t touch the matched data — for example COUNT(*) over the AOF set, or dbSeek-style point lookups — hit the full Rushmore window on top of the range-scan gain.

Bench v4 — wire-protocol xbrowse repaint (rc18)

The wire stack (rc18, M12.17 .. M12.20) ships four overlapping optimisations that collapse the cost of an xbrowse-style PgDn — W columns × H rows × per-row metadata — from ~300 round-trips into ~20 round-trips × ~5 ms with Nagle off ≈ ~100 ms total.

End-to-end LAN measurement, server openads_serverd on the same subnet, 100 columns × 25 visible rows over tcp://:

Stage	RTT	Notes
pre-M12.17 baseline (per-cell FieldGet)	~300	W cells per row + per-row metadata each round-trip.
M12.17 — FetchCurrentRow row cache	~80	W cells per row collapse to 1 RTT.
M12.18 — nav-ack carries row trailer	~20	GotoTop / Skip / GotoRecord acks bundle the row buffer + recno + deleted flag, so AdsGetField / AdsGetRecordNum / AdsIsRecordDeleted hit the cache populated by the prior nav.
M12.19 — record-count cache	~20	AdsGetRecordCount / AdsGetRelKeyPos (the scrollbar) serve from cache; invalidated only by writes that change row count.
M12.20 — TCP_NODELAY	~100 ms	Nagle off — strict ping-pong removes the 200 ms accumulation delay tax.

Net: ~30× end-to-end speedup vs pre-M12.17 on a typical LAN.

Run it on your hardware

cmake --build build/default --target openads_bench --config Release
./build/default/tools/bench/openads_bench --rows 100000 --repeats 5 --csv

The --csv flag emits one CSV row per workload so results pipe straight into a spreadsheet.

Methodology notes

• Each workload runs N times; min / median / max are reported.
• stage_dbf is the time to write the synthetic DBF; reported once (no repeats).
• All SQL workloads run after the same in-process AdsConnect60 so warm-up cost is amortised.
• run_query walks the first 10 rows of the result cursor so the time captures materialisation, not just plan compilation.