UniversalContainer/results/benchmarks/20260813-focused-resize/README.md

Focused resize benchmark — 2026-08-13

This run replaces the retired benchmark matrix with one size sweep and one fixed workload. Every candidate is AdaptiveSequence<uint32_t, true> and uses the same flat hash index; only storage residency differs.

Workload per mixed trace:

• 97,000 random indexed reads (97%);
• 1,000 find_one calls through the flat hash index (1%, 500 hit / 500 miss);
• 1,000 random inserts (1%);
• 1,000 random erases (1%).

Each cell is the median of seven paired repeats. Candidate order is rotated per repeat. Construction is outside the timer; automatic conversion on a resize boundary is inside it. The per-operation columns are separate batched traces, while mixed is the requested end-to-end workload.

Environment: AMD Ryzen 9 5900X (12 cores / 24 threads), Windows 11 Pro build 26200, MSVC 19.51.36248, x64 Release /O2 /MT, baseline ISA profile, no IPO/LTO.

Command:

.\out\bin\baseline\Release\uc_focused_bench.exe `
  --operations 100000 --repeats 7 `
  --output results\benchmarks\20260813-focused-resize\baseline.csv

Median throughput (operations/second)

n	container	mixed mode	mixed leaf	mixed ops/s	indexed read ops/s	hash find ops/s	insert ops/s	erase ops/s
256	forced vector + hash	vector	16	164,500,740	654,450,262	145,369,967	1,739,433	8,205,128
256	forced tiered + hash	tiered	17	41,291,601	54,182,922	81,479,671	2,459,420	5,203,252
256	adaptive + hash	vector	16	151,676,020	571,755,289	146,134,736	1,740,644	8,421,053
512	forced vector + hash	vector	23	103,412,616	657,030,223	129,584,035	1,358,880	5,014,691
512	forced tiered + hash	tiered	23	38,755,184	50,032,521	78,963,992	2,698,327	5,344,468
512	adaptive + hash	vector	23	98,328,417	575,373,993	129,098,890	1,300,221	5,256,674
1,024	forced vector + hash	vector	32	58,816,610	662,251,656	110,411,836	931,619	2,801,120
1,024	forced tiered + hash	tiered	33	35,664,610	46,172,315	71,118,697	3,144,654	4,748,338
1,024	adaptive + hash	vector	32	56,631,555	584,112,150	109,673,174	911,743	2,893,519
2,048	forced vector + hash	vector	46	31,979,533	654,878,847	81,426,594	570,125	973,615
2,048	forced tiered + hash	tiered	46	32,177,103	43,327,556	59,157,596	2,581,311	3,880,481
2,048	adaptive + hash	vector	46	31,330,284	580,383,053	81,406,708	565,835	995,619
4,096	forced vector + hash	vector	64	17,199,862	661,813,369	77,297,673	313,342	421,248
4,096	forced tiered + hash	tiered	65	32,526,672	45,460,745	55,303,617	1,793,079	3,138,732
4,096	adaptive + hash	vector → tiered	65	32,639,206	581,395,349	76,710,647	1,822,822	422,048
8,192	forced vector + hash	vector	91	8,791,518	657,894,737	72,385,089	164,395	198,736
8,192	forced tiered + hash	tiered	91	28,479,481	42,929,510	51,116,904	1,317,003	2,276,867
8,192	adaptive + hash	tiered	91	28,603,301	40,546,568	50,296,751	1,348,618	2,704,896
16,384	forced vector + hash	vector	128	4,489,580	649,350,649	79,478,620	86,972	96,230
16,384	forced tiered + hash	tiered	129	23,255,814	40,364,899	51,284,681	935,629	1,709,694
16,384	adaptive + hash	tiered	129	22,955,260	38,213,153	50,658,561	836,400	2,076,843
32,768	forced vector + hash	vector	182	2,246,076	610,873,549	79,687,625	43,018	47,784
32,768	forced tiered + hash	tiered	182	18,402,989	37,572,797	50,266,412	571,004	1,245,640
32,768	adaptive + hash	tiered	182	17,943,013	35,944,071	49,058,085	514,165	1,548,227
65,536	forced vector + hash	vector	256	1,089,099	490,436,488	63,678,044	21,507	22,895
65,536	forced tiered + hash	tiered	257	12,614,158	32,799,790	44,620,945	315,139	910,167
65,536	adaptive + hash	tiered	257	12,777,430	30,628,809	42,513,392	295,255	1,116,819

Crossover

At n=2,048, forced tiered had only a 0.618% median advantage and won four of seven paired repeats (forced vector won three), so this boundary is treated as noise rather than a stable crossover. At n=4,096, forced tiered won all seven paired repeats and its median throughput was 1.8911x that of forced vector. The adaptive row includes the one-time vector → tiered conversion and still reached 32.64M mixed ops/s, essentially matching already-tiered storage. Therefore the calibrated resize-boundary cutoff is 4,096 elements, the first decisive and stable size in this sweep.

At exactly 4,096 elements an adaptive container is still vector. The next insert changes logical capacity from 4,096 to 8,192 and performs the conversion; the final leaf is ceil(sqrt(4097)) = 65. In the separate read/find/erase batches no growth occurs, so those adaptive 4,096 measurements remain vector; the insert batch and mixed trace include the conversion.

All 189 raw rows have the required operation counts, and checksums match across the three candidates for every (n, repeat) pair. The CSV also records the mode and leaf used by each separate read/find/insert/erase batch. Raw data is in baseline.csv.