加载 cohere-1m 数据集...
加载完成! 1000000 x 768 训练集, 1000 测试集, 耗时 2.37s

======================================================================
实验 1: 参数敏感性分析
数据集: cohere-1m (1000000 x 768)
======================================================================

计算 brute-force 基准 QPS...
Brute-force 基准: QPS=18.4, latency=54.22ms/q

--- 1a: m 的影响 (ef_c=128, α=1.2) ---
m        Build(s)     vecs/s    Hot(MB)  ef=32  R  ef=32  Q  ef=64  R  ef=64  Q  ef=128 R  ef=128 Q  ef=256 R  ef=256 Q  ef=512 R  ef=512 Q  ef=1024R  ef=1024Q
------------------------------------------------------------------------------------------------------------------------------------------------------------------------
4            18.7      53498        247   17.89%  178872   32.45%  105092   46.09%   53922   55.75%   28911   62.58%   14974   66.66%    7763
8            24.7      40469        308   51.40%  127228   69.22%   59564   79.39%   43587   85.82%   23717   89.23%   12628   91.21%    6484
12           32.3      30918        370   69.57%  105241   82.70%   64960   89.80%   36311   93.56%   19829   95.48%   10347   96.52%    5414
16           41.0      24373        431   78.49%   89048   88.33%   56406   93.81%   31749   96.18%   17423   97.56%    9156   98.24%    4655
20           57.0      17552        492   82.78%   75883   90.98%   49003   95.70%   27238   97.54%   12843   98.71%    6978   99.04%    4121
24           65.6      15241        553   85.95%   69300   93.30%   44794   96.79%   24973   98.30%   13864   99.03%    7359   99.36%    3860
28           77.9      12841        614   87.92%   65046   94.23%   40495   97.54%   23065   98.74%   12158   99.38%    6652   99.62%    3442
32           94.8      10551        675   88.72%   58291   95.10%   35213   97.66%   19504   98.92%   11377   99.55%    6081   99.70%    3263
40          138.1       7242        797   89.74%   51768   95.99%   32118   98.05%   18058   99.16%    9298   99.74%    5311   99.79%    2918
48          174.0       5748        919   90.96%   39706   96.33%   27928   98.38%   15900   99.34%    8762   99.77%    4805   99.83%    2565
56          220.4       4536       1041   91.47%   41324   96.67%   25840   98.55%   14218   99.52%    7945   99.79%    4414   99.88%    2411
64          277.5       3604       1163   91.89%   36409   96.83%   23350   98.64%   13071   99.59%    7303   99.81%    4004   99.91%    2204

--- 1b: ef_construction 的影响 (m=32, α=1.2) ---
ef_c       Build(s)     vecs/s    Hot(MB)  ef=32  R  ef=32  Q  ef=64  R  ef=64  Q  ef=128 R  ef=128 Q  ef=256 R  ef=256 Q  ef=512 R  ef=512 Q  ef=1024R  ef=1024Q
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------
16             67.5      14814        675   79.14%   43530   87.68%   25136   92.34%   13911   94.99%    7203   97.10%    3722   98.35%    1925
32             68.8      14526        675   87.39%   53051   93.32%   31805   96.37%   17779   97.81%    9205   98.80%    4811   99.33%    2533
48             70.6      14171        675   87.87%   58682   94.49%   36485   96.87%   20393   98.19%   11225   98.95%    5988   99.33%    3137
64             75.3      13277        675   87.74%   58752   94.17%   36825   96.90%   20975   98.22%   11635   98.95%    6318   99.30%    3345
96             84.8      11794        675   88.36%   60000   95.00%   34871   97.58%   20682   98.83%   11414   99.46%    6054   99.61%    3213
128            92.9      10764        675   88.63%   58678   95.11%   36869   97.69%   21136   98.95%   11618   99.56%    6238   99.71%    3299
160           100.5       9951        675   88.98%   59375   95.19%   36587   97.83%   21184   99.07%   11608   99.58%    6173   99.78%    3293
200           110.6       9046        675   89.08%   60167   95.21%   37079   97.89%   21211   99.14%   11574   99.61%    6192   99.79%    3233
256           122.4       8171        675   88.95%   59248   95.27%   36295   97.95%   20703   99.17%   11432   99.65%    6073   99.83%    3206
384           152.3       6564        675   89.07%   58947   95.30%   37403   97.94%   20790   99.21%   11517   99.66%    6124   99.85%    3204
512           176.1       5680        675   88.95%   59898   95.25%   36320   97.97%   20797   99.24%   11410   99.71%    6167   99.88%    3243

--- 1c: α 的影响 (m=32, ef_c=128) ---
alpha      Build(s)     vecs/s    Hot(MB)  ef=32  R  ef=32  Q  ef=64  R  ef=64  Q  ef=128 R  ef=128 Q  ef=256 R  ef=256 Q  ef=512 R  ef=512 Q  ef=1024R  ef=1024Q
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------
1.00          225.0       4444        675   91.37%   54481   96.73%   31909   98.66%   16997   99.56%    9466   99.84%    5064   99.97%    2697
1.05          101.4       9865        675   89.29%   58380   95.47%   36547   97.89%   20922   99.13%   11303   99.59%    5981   99.78%    3207
1.10           93.8      10657        675   88.68%   58898   95.08%   36480   97.72%   20540   99.00%   11492   99.57%    5936   99.73%    3234
1.15           93.0      10751        675   88.71%   60057   95.10%   37001   97.63%   21167   98.94%   11491   99.54%    6166   99.70%    3224
1.20           92.5      10812        675   88.69%   59673   95.07%   37077   97.66%   21177   98.95%   11378   99.55%    6104   99.71%    3220
1.25           92.4      10822        675   88.50%   60368   95.07%   36985   97.61%   21167   98.93%   11408   99.54%    6219   99.70%    3242

--- 1d: ef_search 精细 Recall-QPS 曲线 (m=32, ef_c=128, α=1.2) ---
ef          R@10(%)       MT-QPS       1T-QPS    lat(ms)      vs BF
------------------------------------------------------------------
8            45.43%       133717        13919      0.07    7250.2x
12           62.29%        98310        10755      0.09    5330.4x
16           72.59%        85273         9037      0.11    4623.6x
20           79.21%        75405         7998      0.13    4088.5x
24           83.26%        65835         7234      0.14    3569.6x
32           88.59%        56239         5841      0.17    3049.3x
40           91.35%        46191         5076      0.20    2504.5x
48           93.27%        44035         4515      0.22    2387.6x
56           94.43%        39131         3856      0.26    2121.7x
64           95.13%        36729         3555      0.28    1991.5x
80           96.07%        30307         2997      0.33    1643.3x
96           96.99%        26116         2656      0.38    1416.0x
112          97.46%        23620         2359      0.42    1280.7x
128          97.69%        21021         2140      0.47    1139.8x
160          98.22%        17388         1770      0.56     942.8x
192          98.57%        15084         1492      0.67     817.8x
224          98.69%        13228         1329      0.75     717.2x
256          98.96%        11837         1182      0.85     641.8x
320          99.11%         9358          980      1.02     507.4x
384          99.38%         8233          829      1.21     446.4x
448          99.49%         7183          730      1.37     389.5x
512          99.55%         6032          644      1.55     327.1x
640          99.59%         5223          533      1.88     283.2x
768          99.63%         4407          451      2.22     239.0x
896          99.70%         3826          394      2.54     207.4x
1024         99.71%         3376          355      2.82     183.0x

--- 1e: 不同 m 的 Pareto 曲线 (ef_c=128, α=1.2) ---
m      ef          R@10(%)       MT-QPS      vs BF
--------------------------------------------------
8      32           50.96%       123364    6688.9x
8      64           68.62%        71714    3888.4x
8      128          79.36%        39278    2129.7x
8      256          85.75%        23010    1247.6x
8      512          89.22%        12401     672.4x
8      1024         91.08%         6320     342.7x

16     32           78.52%        85672    4645.2x
16     64           88.44%        52980    2872.6x
16     128          93.88%        30139    1634.1x
16     256          96.37%        16811     911.5x
16     512          97.59%         9017     488.9x
16     1024         98.26%         4698     254.8x

32     32           88.60%        59758    3240.1x
32     64           95.08%        35453    1922.3x
32     128          97.62%        20291    1100.2x
32     256          98.93%        11348     615.3x
32     512          99.55%         6145     333.2x
32     1024         99.70%         3239     175.6x

48     32           91.09%        44731    2425.3x
48     64           96.47%        27103    1469.5x
48     128          98.38%        15601     845.9x
48     256          99.34%         8771     475.6x
48     512          99.77%         4848     262.8x
48     1024         99.83%         2583     140.0x

64     32           91.89%        34056    1846.5x
64     64           96.85%        21399    1160.3x
64     128          98.66%        12669     686.9x
64     256          99.59%         7127     386.4x
64     512          99.81%         4029     218.5x
64     1024         99.91%         2234     121.1x


--- 1f: m × α 交叉实验 (ef_c=128, ef_search=64) ---
m\α        1.00     1.10     1.15     1.20     1.25
---------------------------------------------------
8        89.60%   69.16%   67.41%   68.10%   67.11%
16       95.10%   88.35%   88.28%   88.28%   88.19%
32       96.61%   95.10%   95.14%   95.10%   95.04%
48       97.18%   96.37%   96.39%   96.23%   96.34%
64       97.29%   96.83%   96.84%   96.83%   96.87%

======================================================================
实验 2: 单线程 vs 多线程查询扩展性
数据集: cohere-1m (1000000 x 768)
======================================================================
索引构建: 93.3s (10722 vecs/s), Hot 675 MB

--- 2a: 不同 ef 下的 单线程 vs 多线程 QPS ---
ef             1T-QPS       MT-QPS      R@10(%)        加速比
----------------------------------------------------------
64               3602        37685       95.10%      10.5x
128              2165        21414       97.69%       9.9x
256              1173        11943       98.94%      10.2x
512               636         6404       99.57%      10.1x
1024              345         3403       99.72%       9.9x

--- 2b: 线程数扩展性 (ef=64) ---
线程数                 QPS        相对1线程
--------------------------------------
1                  3665         1.0x
2                  7433         2.0x
4                 14555         4.0x
8                 26923         7.3x
16                37959        10.4x

全部实验完成!




======================================================================
加载数据...
  训练集: C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB\cohere_train.f32
  测试集: C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB\cohere_test.f32
  GT: C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB\cohere_groundtruth.i32
  维度: 768
  训练集: (1000000, 768)
  测试集: (1000, 768)
  GT: (1000, 10)
======================================================================

======================================================================
Brute-Force Baseline (numpy exact search)
======================================================================
  Brute-force QPS: 131  (latency: 7.64 ms/q, recall: 100.0%)

将运行: hnswlib, faiss_hnsw, faiss_ivfpq, usearch, diskann
硬件: 16 核

======================================================================
Baseline: hnswlib
======================================================================

--- hnswlib M=8, ef_c=64 ---
构建: 60.5s (16,527 vecs/s)
内存: 3,006 MB (向量: 2,930 MB + 邻接: 61 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    50.87%     10,667     54,151      0.09    81.5x   413.6x
    20    63.90%      7,297     35,665      0.14    55.7x   272.4x
    40    75.38%      5,283     24,111      0.19    40.3x   184.1x
    80    82.32%      3,303     14,839      0.30    25.2x   113.3x
   120    85.48%      2,326     10,838      0.43    17.8x    82.8x
   200    88.86%      1,520      7,101      0.66    11.6x    54.2x
   400    90.95%        831      3,814      1.20     6.3x    29.1x
   600    92.49%        564      2,623      1.77     4.3x    20.0x
   800    93.28%        434      2,016      2.31     3.3x    15.4x

--- hnswlib M=16, ef_c=64 ---
构建: 90.2s (11,084 vecs/s)
内存: 3,067 MB (向量: 2,930 MB + 邻接: 122 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    68.49%      7,698     36,212      0.13    58.8x   276.6x
    20    81.25%      5,417     25,877      0.18    41.4x   197.6x
    40    87.85%      3,552     16,728      0.28    27.1x   127.7x
    80    92.66%      2,193      9,936      0.46    16.8x    75.9x
   120    94.13%      1,515      7,077      0.66    11.6x    54.0x
   200    95.73%        993      4,480      1.01     7.6x    34.2x
   400    97.03%        529      2,390      1.89     4.0x    18.3x
   600    97.49%        368      1,639      2.72     2.8x    12.5x
   800    97.84%        280      1,279      3.57     2.1x     9.8x

--- hnswlib M=32, ef_c=64 ---
构建: 115.1s (8,688 vecs/s)
内存: 3,189 MB (向量: 2,930 MB + 邻接: 244 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    75.89%      5,961     28,266      0.17    45.5x   215.9x
    20    85.65%      4,297     19,974      0.23    32.8x   152.5x
    40    91.91%      2,839     12,797      0.35    21.7x    97.7x
    80    95.31%      1,655      7,466      0.60    12.6x    57.0x
   120    96.47%      1,161      5,257      0.86     8.9x    40.1x
   200    97.36%        745      3,323      1.34     5.7x    25.4x
   400    98.67%        391      1,765      2.56     3.0x    13.5x
   600    99.08%        273      1,218      3.66     2.1x     9.3x
   800    99.25%        213        949      4.70     1.6x     7.2x

--- hnswlib M=48, ef_c=64 ---
构建: 132.2s (7,567 vecs/s)
内存: 3,311 MB (向量: 2,930 MB + 邻接: 366 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    77.64%      5,346     25,361      0.19    40.8x   193.7x
    20    87.18%      3,862     17,844      0.26    29.5x   136.3x
    40    92.46%      2,490     11,473      0.40    19.0x    87.6x
    80    95.84%      1,473      6,706      0.68    11.3x    51.2x
   120    96.94%      1,034      4,655      0.97     7.9x    35.6x
   200    97.85%        664      2,969      1.51     5.1x    22.7x
   400    98.99%        351      1,578      2.85     2.7x    12.0x
   600    99.30%        251      1,102      3.98     1.9x     8.4x
   800    99.47%        192        845      5.20     1.5x     6.5x

--- hnswlib M=8, ef_c=128 ---
构建: 120.3s (8,311 vecs/s)
内存: 3,006 MB (向量: 2,930 MB + 邻接: 61 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    54.15%     11,439     51,289      0.09    87.4x   391.7x
    20    67.39%      6,996     34,630      0.14    53.4x   264.5x
    40    79.81%      5,189     23,673      0.19    39.6x   180.8x
    80    86.61%      3,151     14,358      0.32    24.1x   109.6x
   120    89.77%      2,297     10,465      0.44    17.5x    79.9x
   200    92.42%      1,463      6,719      0.68    11.2x    51.3x
   400    94.81%        775      3,585      1.29     5.9x    27.4x
   600    95.75%        546      2,466      1.83     4.2x    18.8x
   800    96.37%        414      1,862      2.41     3.2x    14.2x

--- hnswlib M=16, ef_c=128 ---
构建: 201.6s (4,961 vecs/s)
内存: 3,067 MB (向量: 2,930 MB + 邻接: 122 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    74.08%      7,045     33,548      0.14    53.8x   256.2x
    20    84.18%      5,278     23,711      0.19    40.3x   181.1x
    40    91.00%      3,369     15,211      0.30    25.7x   116.2x
    80    94.66%      1,930      8,940      0.52    14.7x    68.3x
   120    96.21%      1,412      6,357      0.71    10.8x    48.5x
   200    97.41%        904      4,003      1.11     6.9x    30.6x
   400    98.58%        481      2,115      2.08     3.7x    16.2x
   600    98.94%        332      1,476      3.02     2.5x    11.3x
   800    99.18%        256      1,133      3.91     2.0x     8.6x

--- hnswlib M=32, ef_c=128 ---
构建: 284.4s (3,516 vecs/s)
内存: 3,189 MB (向量: 2,930 MB + 邻接: 244 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    80.99%      5,008     24,840      0.20    38.2x   189.7x
    20    89.00%      3,872     17,503      0.26    29.6x   133.7x
    40    93.85%      2,408     10,552      0.42    18.4x    80.6x
    80    96.95%      1,439      6,302      0.69    11.0x    48.1x
   120    98.00%        983      4,418      1.02     7.5x    33.7x
   200    98.80%        630      2,798      1.59     4.8x    21.4x
   400    99.47%        330      1,479      3.03     2.5x    11.3x
   600    99.69%        236      1,042      4.24     1.8x     8.0x
   800    99.74%        181        805      5.52     1.4x     6.1x

--- hnswlib M=48, ef_c=128 ---
构建: 321.7s (3,109 vecs/s)
内存: 3,311 MB (向量: 2,930 MB + 邻接: 366 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    82.41%      4,506     21,936      0.22    34.4x   167.5x
    20    90.15%      3,205     15,549      0.31    24.5x   118.8x
    40    94.74%      2,165      9,730      0.46    16.5x    74.3x
    80    97.62%      1,213      5,505      0.82     9.3x    42.0x
   120    98.40%        874      3,798      1.14     6.7x    29.0x
   200    99.16%        559      2,422      1.79     4.3x    18.5x
   400    99.65%        295      1,303      3.38     2.3x    10.0x
   600    99.74%        207        908      4.82     1.6x     6.9x
   800    99.80%        161        707      6.20     1.2x     5.4x

--- hnswlib M=8, ef_c=168 ---
构建: 165.0s (6,060 vecs/s)
内存: 3,006 MB (向量: 2,930 MB + 邻接: 61 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    54.97%     10,629     51,384      0.09    81.2x   392.4x
    20    70.94%      7,622     35,072      0.13    58.2x   267.8x
    40    81.37%      5,139     23,344      0.19    39.2x   178.3x
    80    88.19%      3,192     14,438      0.31    24.4x   110.3x
   120    91.23%      2,279     10,254      0.44    17.4x    78.3x
   200    93.75%      1,454      6,617      0.69    11.1x    50.5x
   400    95.94%        756      3,550      1.32     5.8x    27.1x
   600    96.83%        526      2,436      1.90     4.0x    18.6x
   800    97.19%        396      1,865      2.52     3.0x    14.2x

--- hnswlib M=16, ef_c=168 ---
构建: 281.0s (3,559 vecs/s)
内存: 3,067 MB (向量: 2,930 MB + 邻接: 122 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    74.66%      6,690     33,435      0.15    51.1x   255.3x
    20    85.02%      4,885     22,960      0.20    37.3x   175.3x
    40    91.42%      3,170     14,989      0.32    24.2x   114.5x
    80    95.19%      1,954      8,678      0.51    14.9x    66.3x
   120    96.79%      1,355      6,121      0.74    10.4x    46.7x
   200    97.99%        859      3,855      1.16     6.6x    29.4x
   400    99.01%        454      2,059      2.20     3.5x    15.7x
   600    99.33%        317      1,413      3.16     2.4x    10.8x
   800    99.50%        239      1,098      4.18     1.8x     8.4x

--- hnswlib M=32, ef_c=168 ---
构建: 397.3s (2,517 vecs/s)
内存: 3,189 MB (向量: 2,930 MB + 邻接: 244 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    81.75%      4,886     24,641      0.20    37.3x   188.2x
    20    90.04%      3,373     16,943      0.30    25.8x   129.4x
    40    95.03%      2,274     10,417      0.44    17.4x    79.6x
    80    97.50%      1,327      5,933      0.75    10.1x    45.3x
   120    98.26%        925      4,171      1.08     7.1x    31.9x
   200    99.02%        590      2,631      1.70     4.5x    20.1x
   400    99.56%        315      1,417      3.17     2.4x    10.8x
   600    99.79%        219        978      4.56     1.7x     7.5x
   800    99.80%        173        764      5.79     1.3x     5.8x

--- hnswlib M=48, ef_c=168 ---
构建: 447.0s (2,237 vecs/s)
内存: 3,311 MB (向量: 2,930 MB + 邻接: 366 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    84.07%      4,046     21,261      0.25    30.9x   162.4x
    20    91.37%      3,068     14,924      0.33    23.4x   114.0x
    40    95.93%      2,062      9,178      0.49    15.7x    70.1x
    80    98.09%      1,137      5,165      0.88     8.7x    39.4x
   120    98.77%        808      3,602      1.24     6.2x    27.5x
   200    99.33%        516      2,255      1.94     3.9x    17.2x
   400    99.75%        273      1,225      3.67     2.1x     9.4x
   600    99.82%        193        855      5.17     1.5x     6.5x
   800    99.88%        150        668      6.65     1.1x     5.1x

======================================================================
Baseline: FAISS HNSW (IndexHNSWFlat)
======================================================================

--- FAISS HNSW M=8, ef_c=64 ---
构建: 64.3s (15,550 vecs/s)
内存: 3,006 MB (向量: 2,930 MB + 邻接: 61 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    51.60%     15,037     50,364      0.07   114.8x   384.6x
    20    66.18%      9,593     31,514      0.10    73.3x   240.7x
    40    76.42%      6,659     20,366      0.15    50.9x   155.5x
    80    82.96%      3,870     13,588      0.26    29.6x   103.8x
   120    85.71%      2,760      9,467      0.36    21.1x    72.3x
   200    88.39%      1,767      6,099      0.57    13.5x    46.6x
   400    90.80%        931      3,309      1.07     7.1x    25.3x
   600    91.94%        615      2,227      1.63     4.7x    17.0x
   800    92.65%        480      1,595      2.09     3.7x    12.2x

--- FAISS HNSW M=16, ef_c=64 ---
构建: 108.4s (9,224 vecs/s)
内存: 3,067 MB (向量: 2,930 MB + 邻接: 122 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    70.92%     10,923     33,220      0.09    83.4x   253.7x
    20    80.38%      8,093     21,868      0.12    61.8x   167.0x
    40    87.27%      4,989     14,910      0.20    38.1x   113.9x
    80    92.18%      2,847      8,717      0.35    21.7x    66.6x
   120    94.22%      1,978      6,294      0.51    15.1x    48.1x
   200    95.55%      1,308      3,970      0.76    10.0x    30.3x
   400    96.91%        687      2,086      1.45     5.3x    15.9x
   600    97.31%        466      1,399      2.15     3.6x    10.7x
   800    97.64%        358      1,075      2.79     2.7x     8.2x

--- FAISS HNSW M=32, ef_c=64 ---
构建: 148.1s (6,754 vecs/s)
内存: 3,189 MB (向量: 2,930 MB + 邻接: 244 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    76.37%      8,726     24,354      0.11    66.6x   186.0x
    20    85.80%      6,327     17,536      0.16    48.3x   133.9x
    40    92.03%      4,001     11,426      0.25    30.6x    87.3x
    80    95.52%      2,360      6,728      0.42    18.0x    51.4x
   120    96.56%      1,593      4,664      0.63    12.2x    35.6x
   200    97.62%      1,013      2,969      0.99     7.7x    22.7x
   400    98.62%        532      1,554      1.88     4.1x    11.9x
   600    98.97%        366      1,068      2.73     2.8x     8.2x
   800    99.16%        282        809      3.55     2.2x     6.2x

--- FAISS HNSW M=48, ef_c=64 ---
构建: 428.4s (2,334 vecs/s)
内存: 3,311 MB (向量: 2,930 MB + 邻接: 366 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    89.03%      5,175     13,636      0.19    39.5x   104.1x
    20    94.01%      3,523      9,152      0.28    26.9x    69.9x
    40    97.28%      2,035      5,657      0.49    15.5x    43.2x
    80    98.68%      1,176      3,146      0.85     9.0x    24.0x
   120    99.18%        804      2,192      1.24     6.1x    16.7x
   200    99.56%        524      1,374      1.91     4.0x    10.5x
   400    99.88%        285        747      3.51     2.2x     5.7x
   600    99.91%        198        517      5.04     1.5x     3.9x
   800    99.94%        151        404      6.64     1.2x     3.1x

--- FAISS HNSW M=8, ef_c=128 ---
构建: 133.8s (7,473 vecs/s)
内存: 3,006 MB (向量: 2,930 MB + 邻接: 61 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    57.24%     14,778     46,234      0.07   112.9x   353.1x
    20    72.57%      9,874     30,926      0.10    75.4x   236.2x
    40    82.29%      6,593     19,581      0.15    50.3x   149.5x
    80    88.19%      3,747     12,769      0.27    28.6x    97.5x
   120    90.61%      2,673      9,427      0.37    20.4x    72.0x
   200    92.63%      1,655      5,910      0.60    12.6x    45.1x
   400    94.44%        880      3,078      1.14     6.7x    23.5x
   600    95.20%        588      1,981      1.70     4.5x    15.1x
   800    95.75%        445      1,512      2.25     3.4x    11.5x

--- FAISS HNSW M=16, ef_c=128 ---
构建: 227.0s (4,405 vecs/s)
内存: 3,067 MB (向量: 2,930 MB + 邻接: 122 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    73.78%     10,599     30,109      0.09    80.9x   229.9x
    20    84.72%      7,525     20,897      0.13    57.5x   159.6x
    40    90.88%      4,533     13,008      0.22    34.6x    99.3x
    80    94.80%      2,650      7,679      0.38    20.2x    58.6x
   120    96.25%      1,860      5,350      0.54    14.2x    40.9x
   200    97.43%      1,155      3,400      0.87     8.8x    26.0x
   400    98.59%        609      1,808      1.64     4.7x    13.8x
   600    98.87%        421      1,224      2.38     3.2x     9.3x
   800    99.12%        323        944      3.10     2.5x     7.2x

--- FAISS HNSW M=32, ef_c=128 ---
构建: 326.5s (3,063 vecs/s)
内存: 3,189 MB (向量: 2,930 MB + 邻接: 244 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    79.56%      7,586     22,509      0.13    57.9x   171.9x
    20    88.68%      5,437     14,834      0.18    41.5x   113.3x
    40    94.36%      3,555      9,728      0.28    27.1x    74.3x
    80    97.45%      1,945      5,484      0.51    14.9x    41.9x
   120    98.32%      1,371      3,758      0.73    10.5x    28.7x
   200    98.91%        851      2,443      1.18     6.5x    18.7x
   400    99.46%        455      1,322      2.20     3.5x    10.1x
   600    99.64%        317        897      3.16     2.4x     6.9x
   800    99.68%        250        691      4.00     1.9x     5.3x

--- FAISS HNSW M=48, ef_c=128 ---
构建: 374.4s (2,671 vecs/s)
内存: 3,311 MB (向量: 2,930 MB + 邻接: 366 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    81.45%      7,070     19,359      0.14    54.0x   147.8x
    20    90.03%      4,946     13,925      0.20    37.8x   106.3x
    40    95.44%      3,005      8,564      0.33    22.9x    65.4x
    80    97.68%      1,698      4,870      0.59    13.0x    37.2x
   120    98.51%      1,216      3,435      0.82     9.3x    26.2x
   200    99.18%        738      2,170      1.35     5.6x    16.6x
   400    99.60%        413      1,167      2.42     3.2x     8.9x
   600    99.70%        290        800      3.44     2.2x     6.1x
   800    99.81%        221        613      4.53     1.7x     4.7x

--- FAISS HNSW M=8, ef_c=168 ---
构建: 195.4s (5,117 vecs/s)
内存: 3,006 MB (向量: 2,930 MB + 邻接: 61 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    59.37%     14,319     47,365      0.07   109.4x   361.7x
    20    73.66%      9,791     28,954      0.10    74.8x   221.1x
    40    83.32%      5,820     19,653      0.17    44.4x   150.1x
    80    88.95%      3,461     12,229      0.29    26.4x    93.4x
   120    91.21%      2,504      8,745      0.40    19.1x    66.8x
   200    93.36%      1,501      5,101      0.67    11.5x    39.0x
   400    95.14%        801      2,987      1.25     6.1x    22.8x
   600    96.10%        544      1,831      1.84     4.2x    14.0x
   800    96.60%        398      1,403      2.51     3.0x    10.7x

--- FAISS HNSW M=16, ef_c=168 ---
构建: 320.1s (3,124 vecs/s)
内存: 3,067 MB (向量: 2,930 MB + 邻接: 122 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    74.39%     10,367     30,131      0.10    79.2x   230.1x
    20    84.41%      7,348     19,736      0.14    56.1x   150.7x
    40    91.38%      4,479     12,502      0.22    34.2x    95.5x
    80    95.55%      2,602      7,447      0.38    19.9x    56.9x
   120    97.01%      1,758      5,188      0.57    13.4x    39.6x
   200    97.84%      1,140      3,275      0.88     8.7x    25.0x
   400    98.68%        588      1,766      1.70     4.5x    13.5x
   600    99.03%        404      1,197      2.48     3.1x     9.1x
   800    99.26%        314        908      3.18     2.4x     6.9x

--- FAISS HNSW M=32, ef_c=168 ---
构建: 449.2s (2,226 vecs/s)
内存: 3,189 MB (向量: 2,930 MB + 邻接: 244 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    82.42%      7,010     20,605      0.14    53.5x   157.4x
    20    90.02%      5,419     14,464      0.18    41.4x   110.5x
    40    94.97%      3,297      9,069      0.30    25.2x    69.3x
    80    97.72%      1,827      5,278      0.55    14.0x    40.3x
   120    98.31%      1,244      3,692      0.80     9.5x    28.2x
   200    99.13%        782      2,247      1.28     6.0x    17.2x
   400    99.63%        435      1,239      2.30     3.3x     9.5x
   600    99.77%        307        843      3.26     2.3x     6.4x
   800    99.81%        235        655      4.26     1.8x     5.0x

--- FAISS HNSW M=48, ef_c=168 ---
构建: 514.5s (1,944 vecs/s)
内存: 3,311 MB (向量: 2,930 MB + 邻接: 366 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    83.89%      6,958     18,217      0.14    53.1x   139.1x
    20    91.02%      4,846     13,142      0.21    37.0x   100.4x
    40    96.19%      2,907      8,154      0.34    22.2x    62.3x
    80    97.98%      1,593      4,467      0.63    12.2x    34.1x
   120    99.00%      1,158      3,245      0.86     8.8x    24.8x
   200    99.60%        712      2,010      1.40     5.4x    15.3x
   400    99.72%        395      1,094      2.53     3.0x     8.4x
   600    99.81%        277        747      3.61     2.1x     5.7x
   800    99.86%        212        578      4.71     1.6x     4.4x

======================================================================
Baseline: FAISS IVF-PQ
======================================================================

--- FAISS IVF-PQ nlist=256, m_pq=32 ---
构建(含训练): 18.1s
内存: 32 MB (PQ码: 31 MB + 簇心: 1 MB)
  nprobe      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
--------------------------------------------------------------------
       1    24.48%     10,129     28,240      0.10    77.4x   215.7x
       4    30.41%      3,571     15,568      0.28    27.3x   118.9x
       8    31.36%      2,175      8,887      0.46    16.6x    67.9x
      16    31.70%      1,198      4,173      0.83     9.2x    31.9x
      32    31.86%        636      3,033      1.57     4.9x    23.2x
      64    31.89%        324      1,655      3.09     2.5x    12.6x
     128    31.89%        164        873      6.09     1.3x     6.7x
     256    31.89%         89        470     11.28     0.7x     3.6x

--- FAISS IVF-PQ nlist=256, m_pq=48 ---
构建(含训练): 18.4s
内存: 47 MB (PQ码: 46 MB + 簇心: 1 MB)
  nprobe      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
--------------------------------------------------------------------
       1    30.49%      8,224     32,353      0.12    62.8x   247.1x
       4    39.93%      2,803     10,050      0.36    21.4x    76.8x
       8    41.39%      1,620      5,856      0.62    12.4x    44.7x
      16    42.16%        869      2,782      1.15     6.6x    21.2x
      32    42.40%        452      1,975      2.21     3.4x    15.1x
      64    42.45%        229      1,120      4.38     1.7x     8.6x
     128    42.46%        116        526      8.62     0.9x     4.0x
     256    42.46%         62        298     16.03     0.5x     2.3x

--- FAISS IVF-PQ nlist=256, m_pq=64 ---
构建(含训练): 9.4s
内存: 63 MB (PQ码: 61 MB + 簇心: 1 MB)
  nprobe      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
--------------------------------------------------------------------
       1    34.35%      6,258     25,210      0.16    47.8x   192.5x
       4    46.30%      2,327      7,972      0.43    17.8x    60.9x
       8    48.33%      1,308      4,945      0.76    10.0x    37.8x
      16    49.28%        684      2,698      1.46     5.2x    20.6x
      32    49.83%        348      1,604      2.87     2.7x    12.2x
      64    49.94%        176        780      5.67     1.3x     6.0x
     128    49.94%         89        410     11.22     0.7x     3.1x
     256    49.94%         48        218     20.99     0.4x     1.7x

--- FAISS IVF-PQ nlist=256, m_pq=96 ---
构建(含训练): 9.0s
内存: 93 MB (PQ码: 92 MB + 簇心: 1 MB)
  nprobe      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
--------------------------------------------------------------------
       1    39.12%      4,635     19,683      0.22    35.4x   150.3x
       4    54.47%      1,495      5,263      0.67    11.4x    40.2x
       8    57.99%        817      3,063      1.22     6.2x    23.4x
      16    59.84%        415      1,578      2.41     3.2x    12.1x
      32    60.77%        212        914      4.72     1.6x     7.0x
      64    61.05%        107        491      9.36     0.8x     3.7x
     128    61.11%         55        258     18.32     0.4x     2.0x
     256    61.11%         29        138     34.24     0.2x     1.1x

--- FAISS IVF-PQ nlist=1024, m_pq=32 ---
构建(含训练): 27.4s
内存: 34 MB (PQ码: 31 MB + 簇心: 3 MB)
  nprobe      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
--------------------------------------------------------------------
       1    23.81%     12,588     58,493      0.08    96.1x   446.7x
       4    31.24%      7,041     32,721      0.14    53.8x   249.9x
       8    32.43%      5,126     16,327      0.20    39.1x   124.7x
      16    33.00%      3,325     13,632      0.30    25.4x   104.1x
      32    33.20%      1,918      7,052      0.52    14.6x    53.9x
      64    33.23%      1,058      3,934      0.95     8.1x    30.0x
     128    33.26%        563      2,265      1.78     4.3x    17.3x
     256    33.26%        288      1,421      3.47     2.2x    10.9x

--- FAISS IVF-PQ nlist=1024, m_pq=48 ---
构建(含训练): 27.7s
内存: 50 MB (PQ码: 46 MB + 簇心: 3 MB)
  nprobe      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
--------------------------------------------------------------------
       1    28.57%     14,579     57,076      0.07   111.3x   435.9x
       4    39.48%      6,366     25,714      0.16    48.6x   196.4x
       8    41.71%      4,120     15,929      0.24    31.5x   121.7x
      16    42.90%      2,623      9,268      0.38    20.0x    70.8x
      32    43.48%      1,476      5,341      0.68    11.3x    40.8x
      64    43.66%        779      2,709      1.28     6.0x    20.7x
     128    43.73%        399      1,805      2.51     3.0x    13.8x
     256    43.73%        204        922      4.90     1.6x     7.0x

--- FAISS IVF-PQ nlist=1024, m_pq=64 ---
构建(含训练): 18.7s
内存: 65 MB (PQ码: 61 MB + 簇心: 3 MB)
  nprobe      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
--------------------------------------------------------------------
       1    31.86%     13,113     57,170      0.08   100.1x   436.6x
       4    45.39%      5,695     23,616      0.18    43.5x   180.4x
       8    48.37%      3,756     10,305      0.27    28.7x    78.7x
      16    50.05%      2,219      6,691      0.45    16.9x    51.1x
      32    50.89%      1,185      4,113      0.84     9.1x    31.4x
      64    51.21%        610      2,405      1.64     4.7x    18.4x
     128    51.40%        309      1,345      3.23     2.4x    10.3x
     256    51.41%        158        717      6.32     1.2x     5.5x

--- FAISS IVF-PQ nlist=1024, m_pq=96 ---
构建(含训练): 18.2s
内存: 95 MB (PQ码: 92 MB + 簇心: 3 MB)
  nprobe      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
--------------------------------------------------------------------
       1    35.44%     10,370     36,533      0.10    79.2x   279.0x
       4    52.28%      4,379     14,529      0.23    33.4x   111.0x
       8    56.79%      2,551      9,311      0.39    19.5x    71.1x
      16    59.58%      1,412      4,327      0.71    10.8x    33.0x
      32    61.40%        744      2,372      1.34     5.7x    18.1x
      64    62.06%        380      1,584      2.63     2.9x    12.1x
     128    62.36%        192        862      5.21     1.5x     6.6x
     256    62.47%         98        466     10.16     0.8x     3.6x

--- FAISS IVF-PQ nlist=4096, m_pq=32 ---
构建(含训练): 139.3s
内存: 43 MB (PQ码: 31 MB + 簇心: 12 MB)
  nprobe      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
--------------------------------------------------------------------
       1    30.25%     14,044     35,445      0.07   107.3x   270.7x
       4    35.64%      9,351     27,618      0.11    71.4x   210.9x
       8    36.73%      8,035     29,448      0.12    61.4x   224.9x
      16    37.40%      6,371     23,360      0.16    48.7x   178.4x
      32    37.55%      4,190     17,275      0.24    32.0x   131.9x
      64    37.62%      2,701     10,340      0.37    20.6x    79.0x
     128    37.60%      1,560      5,999      0.64    11.9x    45.8x
     256    37.61%        836      3,286      1.20     6.4x    25.1x

--- FAISS IVF-PQ nlist=4096, m_pq=48 ---
构建(含训练): 139.9s
内存: 59 MB (PQ码: 46 MB + 簇心: 12 MB)
  nprobe      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
--------------------------------------------------------------------
       1    34.39%     17,183     44,770      0.06   131.2x   341.9x
       4    42.40%      9,376     34,399      0.11    71.6x   262.7x
       8    44.30%      7,526     25,646      0.13    57.5x   195.9x
      16    45.40%      5,537     19,745      0.18    42.3x   150.8x
      32    45.85%      3,604      8,822      0.28    27.5x    67.4x
      64    46.08%      2,157      6,364      0.46    16.5x    48.6x
     128    46.14%      1,171      3,906      0.85     8.9x    29.8x
     256    46.18%        618      2,683      1.62     4.7x    20.5x

--- FAISS IVF-PQ nlist=4096, m_pq=64 ---
构建(含训练): 131.3s
内存: 74 MB (PQ码: 61 MB + 簇心: 12 MB)
  nprobe      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
--------------------------------------------------------------------
       1    37.57%     20,387     48,374      0.05   155.7x   369.4x
       4    47.53%     10,199     35,026      0.10    77.9x   267.5x
       8    50.44%      7,469     20,525      0.13    57.0x   156.8x
      16    52.14%      5,141     16,981      0.19    39.3x   129.7x
      32    52.88%      3,134     11,660      0.32    23.9x    89.0x
      64    53.39%      1,758      5,815      0.57    13.4x    44.4x
     128    53.59%        939      3,514      1.06     7.2x    26.8x
     256    53.69%        490      2,003      2.04     3.7x    15.3x

--- FAISS IVF-PQ nlist=4096, m_pq=96 ---
构建(含训练): 135.7s
内存: 104 MB (PQ码: 92 MB + 簇心: 12 MB)
  nprobe      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
--------------------------------------------------------------------
       1    40.48%     17,089     42,502      0.06   130.5x   324.6x
       4    53.22%      8,569     30,581      0.12    65.4x   233.5x
       8    57.45%      5,798     19,867      0.17    44.3x   151.7x
      16    59.98%      3,768     12,013      0.27    28.8x    91.7x
      32    61.60%      2,128      7,037      0.47    16.3x    53.7x
      64    62.69%      1,160      4,572      0.86     8.9x    34.9x
     128    63.26%        604      2,578      1.66     4.6x    19.7x
     256    63.48%        312      1,421      3.21     2.4x    10.9x

======================================================================
Baseline: USearch HNSW
======================================================================

--- USearch M=8, ef_c=64 ---
构建: 69.5s (14,390 vecs/s)
内存: 3,006 MB (向量: 2,930 MB + 邻接: 61 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    48.66%      6,766     43,383      0.15    51.7x   331.3x
    20    63.37%      4,703     30,261      0.21    35.9x   231.1x
    40    74.82%      3,070     20,233      0.33    23.4x   154.5x
    80    82.58%      1,834     12,495      0.55    14.0x    95.4x
   120    85.70%      1,281      8,076      0.78     9.8x    61.7x
   200    88.83%        786      5,691      1.27     6.0x    43.5x
   400    91.66%        435      3,084      2.30     3.3x    23.6x
   600    92.80%        302      2,127      3.31     2.3x    16.2x
   800    93.66%        231      1,640      4.32     1.8x    12.5x

--- USearch M=16, ef_c=64 ---
构建: 108.4s (9,221 vecs/s)
内存: 3,067 MB (向量: 2,930 MB + 邻接: 122 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    69.00%      4,761     30,500      0.21    36.4x   232.9x
    20    79.60%      3,553     21,766      0.28    27.1x   166.2x
    40    87.44%      2,213     13,944      0.45    16.9x   106.5x
    80    92.56%      1,281      8,254      0.78     9.8x    63.0x
   120    94.46%        888      5,898      1.13     6.8x    45.0x
   200    95.98%        557      3,732      1.79     4.3x    28.5x
   400    97.45%        295      1,948      3.39     2.3x    14.9x
   600    97.91%        201      1,361      4.98     1.5x    10.4x
   800    98.20%        157      1,035      6.39     1.2x     7.9x

--- USearch M=32, ef_c=64 ---
构建: 146.5s (6,824 vecs/s)
内存: 3,189 MB (向量: 2,930 MB + 邻接: 244 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    75.68%      3,786     23,753      0.26    28.9x   181.4x
    20    85.00%      2,612     16,947      0.38    19.9x   129.4x
    40    91.65%      1,638     10,387      0.61    12.5x    79.3x
    80    95.36%        963      5,869      1.04     7.4x    44.8x
   120    96.62%        690      4,386      1.45     5.3x    33.5x
   200    97.79%        418      2,712      2.40     3.2x    20.7x
   400    98.62%        225      1,457      4.45     1.7x    11.1x
   600    98.97%        154        989      6.49     1.2x     7.6x
   800    99.11%        120        771      8.34     0.9x     5.9x

--- USearch M=48, ef_c=64 ---
构建: 161.2s (6,203 vecs/s)
内存: 3,311 MB (向量: 2,930 MB + 邻接: 366 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    77.44%      3,363     20,674      0.30    25.7x   157.9x
    20    86.90%      2,365     14,877      0.42    18.1x   113.6x
    40    92.48%      1,554      8,791      0.64    11.9x    67.1x
    80    95.63%        875      5,573      1.14     6.7x    42.6x
   120    97.16%        604      3,845      1.66     4.6x    29.4x
   200    98.22%        389      2,419      2.57     3.0x    18.5x
   400    99.01%        201      1,288      4.98     1.5x     9.8x
   600    99.27%        140        898      7.15     1.1x     6.9x
   800    99.38%        109        695      9.21     0.8x     5.3x

--- USearch M=8, ef_c=128 ---
构建: 139.1s (7,188 vecs/s)
内存: 3,006 MB (向量: 2,930 MB + 邻接: 61 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    52.40%      6,766     43,182      0.15    51.7x   329.8x
    20    67.91%      4,805     29,304      0.21    36.7x   223.8x
    40    80.80%      2,998     19,250      0.33    22.9x   147.0x
    80    87.35%      1,780     11,217      0.56    13.6x    85.7x
   120    90.15%      1,281      8,486      0.78     9.8x    64.8x
   200    92.75%        819      5,578      1.22     6.3x    42.6x
   400    95.08%        420      2,839      2.38     3.2x    21.7x
   600    95.94%        289      1,984      3.47     2.2x    15.2x
   800    96.51%        222      1,510      4.50     1.7x    11.5x

--- USearch M=16, ef_c=128 ---
构建: 237.8s (4,205 vecs/s)
内存: 3,067 MB (向量: 2,930 MB + 邻接: 122 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    72.30%      4,101     28,460      0.24    31.3x   217.3x
    20    82.67%      3,256     19,425      0.31    24.9x   148.3x
    40    90.43%      2,004     12,601      0.50    15.3x    96.2x
    80    94.61%      1,156      7,236      0.87     8.8x    55.3x
   120    96.38%        818      5,149      1.22     6.2x    39.3x
   200    97.40%        518      3,269      1.93     4.0x    25.0x
   400    98.66%        267      1,750      3.74     2.0x    13.4x
   600    98.99%        185      1,218      5.41     1.4x     9.3x
   800    99.17%        142        919      7.02     1.1x     7.0x

--- USearch M=32, ef_c=128 ---
构建: 331.1s (3,020 vecs/s)
内存: 3,189 MB (向量: 2,930 MB + 邻接: 244 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    79.98%      3,347     20,858      0.30    25.6x   159.3x
    20    88.60%      2,394     13,990      0.42    18.3x   106.8x
    40    93.78%      1,470      9,098      0.68    11.2x    69.5x
    80    96.85%        825      5,189      1.21     6.3x    39.6x
   120    97.92%        568      3,633      1.76     4.3x    27.7x
   200    99.00%        370      2,320      2.70     2.8x    17.7x
   400    99.41%        196      1,216      5.09     1.5x     9.3x
   600    99.61%        134        849      7.44     1.0x     6.5x
   800    99.68%        104        665      9.60     0.8x     5.1x

--- USearch M=48, ef_c=128 ---
构建: 376.5s (2,656 vecs/s)
内存: 3,311 MB (向量: 2,930 MB + 邻接: 366 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    82.27%      2,944     18,569      0.34    22.5x   141.8x
    20    89.62%      2,093     12,475      0.48    16.0x    95.3x
    40    94.71%      1,309      7,984      0.76    10.0x    61.0x
    80    97.30%        724      4,579      1.38     5.5x    35.0x
   120    98.25%        518      3,157      1.93     4.0x    24.1x
   200    99.30%        324      1,978      3.08     2.5x    15.1x
   400    99.65%        172      1,076      5.80     1.3x     8.2x
   600    99.70%        119        747      8.39     0.9x     5.7x
   800    99.78%         93        583     10.75     0.7x     4.5x

--- USearch M=8, ef_c=168 ---
构建: 184.1s (5,432 vecs/s)
内存: 3,006 MB (向量: 2,930 MB + 邻接: 61 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    53.68%      6,556     44,031      0.15    50.1x   336.3x
    20    69.63%      4,622     28,337      0.22    35.3x   216.4x
    40    81.53%      3,028     18,943      0.33    23.1x   144.7x
    80    88.31%      1,754     11,842      0.57    13.4x    90.4x
   120    90.95%      1,241      8,458      0.81     9.5x    64.6x
   200    93.54%        797      5,256      1.25     6.1x    40.1x
   400    95.93%        420      2,872      2.38     3.2x    21.9x
   600    96.80%        284      1,990      3.52     2.2x    15.2x
   800    97.25%        216      1,524      4.63     1.7x    11.6x

--- USearch M=16, ef_c=168 ---
构建: 327.0s (3,058 vecs/s)
内存: 3,067 MB (向量: 2,930 MB + 邻接: 122 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    73.12%      4,346     26,169      0.23    33.2x   199.9x
    20    84.03%      3,093     19,168      0.32    23.6x   146.4x
    40    91.20%      1,907     12,421      0.52    14.6x    94.9x
    80    95.28%      1,115      7,163      0.90     8.5x    54.7x
   120    96.94%        778      5,026      1.29     5.9x    38.4x
   200    97.95%        484      3,164      2.07     3.7x    24.2x
   400    98.80%        260      1,685      3.85     2.0x    12.9x
   600    99.16%        180      1,154      5.54     1.4x     8.8x
   800    99.31%        138        908      7.27     1.1x     6.9x

--- USearch M=32, ef_c=168 ---
构建: 458.2s (2,183 vecs/s)
内存: 3,189 MB (向量: 2,930 MB + 邻接: 244 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    81.07%      3,229     20,174      0.31    24.7x   154.1x
    20    89.71%      2,225     13,633      0.45    17.0x   104.1x
    40    94.80%      1,392      8,496      0.72    10.6x    64.9x
    80    97.36%        778      4,952      1.29     5.9x    37.8x
   120    98.37%        537      3,383      1.86     4.1x    25.8x
   200    99.11%        343      2,172      2.92     2.6x    16.6x
   400    99.60%        185      1,156      5.40     1.4x     8.8x
   600    99.72%        125        797      7.97     1.0x     6.1x
   800    99.77%         95        622     10.48     0.7x     4.8x

--- USearch M=48, ef_c=168 ---
构建: 518.9s (1,927 vecs/s)
内存: 3,311 MB (向量: 2,930 MB + 邻接: 366 MB)
    ef      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
------------------------------------------------------------------
    10    83.23%      2,798     17,432      0.36    21.4x   133.1x
    20    90.68%      1,977     12,195      0.51    15.1x    93.1x
    40    95.29%      1,213      7,515      0.82     9.3x    57.4x
    80    97.65%        681      4,273      1.47     5.2x    32.6x
   120    98.60%        475      2,952      2.10     3.6x    22.5x
   200    99.28%        301      1,893      3.32     2.3x    14.5x
   400    99.70%        163      1,017      6.14     1.2x     7.8x
   600    99.76%        113        702      8.87     0.9x     5.4x
   800    99.83%         88        548     11.33     0.7x     4.2x
[跳过] diskannpy 未安装 (pip install diskannpy)

======================================================================
Baseline: FAISS IVF-PQ
======================================================================

>>> 模式 A: 裸 IVF-PQ（无 OPQ / 无 Refine）<<<

--- 裸 IVF-PQ nlist=1024, m_pq=96 ---
构建(含训练): 17.7s
内存: 92 MB (仅 PQ 码)
  nprobe      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
--------------------------------------------------------------------
       1    35.44%     10,708     46,300      0.09    95.0x   410.6x
       4    52.28%      4,437     15,088      0.23    39.3x   133.8x
       8    56.79%      2,586      8,255      0.39    22.9x    73.2x
      16    59.58%      1,425      5,101      0.70    12.6x    45.2x
      32    61.40%        746      2,544      1.34     6.6x    22.6x
      64    62.06%        384      1,548      2.60     3.4x    13.7x
     128    62.36%        194        816      5.15     1.7x     7.2x
     256    62.47%         99        452     10.13     0.9x     4.0x

>>> 模式 B: OPQ + IVF-PQ + Refine（生产级配置）<<<

--- OPQ+IVF-PQ+Refine nlist=1024, m_pq=64 ---
构建(含 OPQ 训练): 94.3s
内存: 2,991 MB (PQ码: 61 MB + f32精排: 2,930 MB)
  nprobe  k_fac      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
---------------------------------------------------------------------------
       4      1    55.58%      7,552     23,732      0.13    67.0x   210.5x
       4      4    70.52%      6,905     21,465      0.14    61.2x   190.4x
       4     10    71.77%      5,592     19,106      0.18    49.6x   169.4x
       4     20    71.87%      4,438     16,059      0.23    39.4x   142.4x
      16      1    63.32%      2,280      6,959      0.44    20.2x    61.7x
      16      4    85.32%      2,201      7,357      0.45    19.5x    65.2x
      16     10    88.28%      2,066      7,245      0.48    18.3x    64.3x
      16     20    88.88%      1,836      6,529      0.54    16.3x    57.9x
      32      1    64.91%      1,149      2,974      0.87    10.2x    26.4x
      32      4    89.22%      1,112      3,438      0.90     9.9x    30.5x
      32     10    93.02%      1,112      3,335      0.90     9.9x    29.6x
      32     20    93.91%      1,060      3,732      0.94     9.4x    33.1x
      64      1    65.89%        611      2,236      1.64     5.4x    19.8x
      64      4    91.50%        603      2,363      1.66     5.3x    21.0x
      64     10    95.87%        593      1,967      1.69     5.3x    17.4x
      64     20    96.90%        575      2,248      1.74     5.1x    19.9x
     128      1    66.25%        312      1,406      3.21     2.8x    12.5x
     128      4    92.75%        310      1,304      3.22     2.8x    11.6x
     128     10    97.34%        305      1,258      3.28     2.7x    11.2x
     128     20    98.54%        302      1,105      3.31     2.7x     9.8x
     256      1    66.38%        160        686      6.27     1.4x     6.1x
     256      4    93.21%        157        720      6.39     1.4x     6.4x
     256     10    97.93%        158        708      6.32     1.4x     6.3x
     256     20    99.22%        156        725      6.41     1.4x     6.4x

--- OPQ+IVF-PQ+Refine nlist=1024, m_pq=96 ---
构建(含 OPQ 训练): 97.1s
内存: 3,021 MB (PQ码: 92 MB + f32精排: 2,930 MB)
  nprobe  k_fac      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
---------------------------------------------------------------------------
       4      1    61.41%      4,883     15,348      0.20    43.3x   136.1x
       4      4    71.61%      4,688     14,818      0.21    41.6x   131.4x
       4     10    71.85%      3,995     13,028      0.25    35.4x   115.5x
       4     20    71.85%      3,426     13,068      0.29    30.4x   115.9x
      16      1    71.23%      1,442      4,628      0.69    12.8x    41.0x
      16      4    88.39%      1,400      4,923      0.71    12.4x    43.7x
      16     10    89.18%      1,353      4,442      0.74    12.0x    39.4x
      16     20    89.25%      1,253      4,562      0.80    11.1x    40.5x
      32      1    73.66%        746      2,701      1.34     6.6x    24.0x
      32      4    93.27%        730      2,423      1.37     6.5x    21.5x
      32     10    94.24%        722      2,460      1.39     6.4x    21.8x
      32     20    94.34%        685      2,608      1.46     6.1x    23.1x
      64      1    74.98%        379      1,529      2.64     3.4x    13.6x
      64      4    95.94%        377      1,601      2.65     3.3x    14.2x
      64     10    97.19%        370      1,661      2.70     3.3x    14.7x
      64     20    97.34%        364      1,691      2.75     3.2x    15.0x
     128      1    75.70%        194        896      5.15     1.7x     7.9x
     128      4    97.56%        193        891      5.17     1.7x     7.9x
     128     10    98.96%        191        829      5.25     1.7x     7.4x
     128     20    99.13%        190        846      5.27     1.7x     7.5x
     256      1    75.92%         99        453     10.15     0.9x     4.0x
     256      4    98.18%         98        456     10.16     0.9x     4.0x
     256     10    99.61%         98        447     10.15     0.9x     4.0x
     256     20    99.78%         96        441     10.42     0.9x     3.9x

--- OPQ+IVF-PQ+Refine nlist=1024, m_pq=128 ---
构建(含 OPQ 训练): 167.5s
内存: 3,052 MB (PQ码: 122 MB + f32精排: 2,930 MB)
  nprobe  k_fac      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
---------------------------------------------------------------------------
       4      1    65.15%      2,948     11,136      0.34    26.1x    98.8x
       4      4    71.92%      2,896     11,482      0.35    25.7x   101.8x
       4     10    71.94%      2,639     10,037      0.38    23.4x    89.0x
       4     20    71.94%      2,283      8,713      0.44    20.2x    77.3x
      16      1    76.91%        950      3,649      1.05     8.4x    32.4x
      16      4    89.13%        934      3,615      1.07     8.3x    32.1x
      16     10    89.23%        915      3,485      1.09     8.1x    30.9x
      16     20    89.23%        873      3,277      1.14     7.7x    29.1x
      32      1    79.93%        506      2,216      1.98     4.5x    19.7x
      32      4    94.04%        503      2,016      1.99     4.5x    17.9x
      32     10    94.23%        494      2,151      2.02     4.4x    19.1x
      32     20    94.23%        480      2,054      2.08     4.3x    18.2x
      64      1    81.72%        261      1,186      3.83     2.3x    10.5x
      64      4    97.00%        256      1,200      3.91     2.3x    10.6x
      64     10    97.25%        257      1,193      3.89     2.3x    10.6x
      64     20    97.25%        255      1,046      3.92     2.3x     9.3x
     128      1    82.57%        133        637      7.52     1.2x     5.7x
     128      4    98.86%        132        641      7.57     1.2x     5.7x
     128     10    99.10%        133        641      7.52     1.2x     5.7x
     128     20    99.11%        132        620      7.59     1.2x     5.5x
     256      1    82.89%         68        331     14.66     0.6x     2.9x
     256      4    99.56%         68        344     14.63     0.6x     3.1x
     256     10    99.81%         68        345     14.80     0.6x     3.1x
     256     20    99.82%         68        337     14.78     0.6x     3.0x

--- OPQ+IVF-PQ+Refine nlist=1024, m_pq=192 ---
构建(含 OPQ 训练): 129.2s
内存: 3,113 MB (PQ码: 183 MB + f32精排: 2,930 MB)
  nprobe  k_fac      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
---------------------------------------------------------------------------
       4      1    67.99%      2,284      6,972      0.44    20.3x    61.8x
       4      4    71.93%      2,200      6,238      0.45    19.5x    55.3x
       4     10    71.93%      2,086      7,532      0.48    18.5x    66.8x
       4     20    71.93%      1,876      7,191      0.53    16.6x    63.8x
      16      1    81.75%        616      2,829      1.62     5.5x    25.1x
      16      4    89.14%        605      2,258      1.65     5.4x    20.0x
      16     10    89.15%        597      2,751      1.67     5.3x    24.4x
      16     20    89.15%        579      3,004      1.73     5.1x    26.6x
      32      1    85.36%        315      1,376      3.18     2.8x    12.2x
      32      4    94.30%        311      1,382      3.22     2.8x    12.3x
      32     10    94.31%        306      1,518      3.27     2.7x    13.5x
      32     20    94.31%        304      1,663      3.29     2.7x    14.7x
      64      1    87.22%        160        852      6.26     1.4x     7.6x
      64      4    97.30%        159        811      6.31     1.4x     7.2x
      64     10    97.32%        158        879      6.32     1.4x     7.8x
      64     20    97.32%        157        812      6.39     1.4x     7.2x
     128      1    88.21%         81        438     12.36     0.7x     3.9x
     128      4    99.04%         80        453     12.45     0.7x     4.0x
     128     10    99.08%         80        451     12.47     0.7x     4.0x
     128     20    99.08%         80        435     12.50     0.7x     3.9x
     256      1    88.58%         41        233     24.32     0.4x     2.1x
     256      4    99.76%         41        223     24.38     0.4x     2.0x
     256     10    99.81%         41        231     24.40     0.4x     2.1x
     256     20    99.81%         41        225     24.47     0.4x     2.0x

--- OPQ+IVF-PQ+Refine nlist=4096, m_pq=64 ---
构建(含 OPQ 训练): 206.3s
内存: 2,991 MB (PQ码: 61 MB + f32精排: 2,930 MB)
  nprobe  k_fac      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
---------------------------------------------------------------------------
       4      1    55.51%     14,468     35,987      0.07   128.3x   319.2x
       4      4    69.15%     12,793     30,322      0.08   113.5x   268.9x
       4     10    69.74%      9,486     23,177      0.11    84.1x   205.6x
       4     20    69.74%      6,527     21,754      0.15    57.9x   192.9x
      16      1    63.52%      5,995     16,674      0.17    53.2x   147.9x
      16      4    84.27%      5,372     16,112      0.19    47.6x   142.9x
      16     10    86.08%      4,768     13,690      0.21    42.3x   121.4x
      16     20    86.37%      3,865      9,884      0.26    34.3x    87.7x
      32      1    65.25%      3,318      9,990      0.30    29.4x    88.6x
      32      4    88.23%      3,174      9,279      0.32    28.1x    82.3x
      32     10    90.68%      2,934      8,367      0.34    26.0x    74.2x
      32     20    91.16%      2,584      6,919      0.39    22.9x    61.4x
      64      1    66.51%      1,787      5,627      0.56    15.8x    49.9x
      64      4    91.24%      1,750      5,882      0.57    15.5x    52.2x
      64     10    94.20%      1,620      5,399      0.62    14.4x    47.9x
      64     20    94.91%      1,555      5,566      0.64    13.8x    49.4x
     128      1    67.03%        929      3,307      1.08     8.2x    29.3x
     128      4    93.12%        925      3,353      1.08     8.2x    29.7x
     128     10    96.45%        892      2,996      1.12     7.9x    26.6x
     128     20    97.25%        855      3,200      1.17     7.6x    28.4x
     256      1    67.35%        495      2,027      2.02     4.4x    18.0x
     256      4    94.18%        491      2,111      2.04     4.4x    18.7x
     256     10    97.70%        483      1,929      2.07     4.3x    17.1x
     256     20    98.65%        466      1,911      2.14     4.1x    17.0x

--- OPQ+IVF-PQ+Refine nlist=4096, m_pq=96 ---
构建(含 OPQ 训练): 207.2s
内存: 3,021 MB (PQ码: 92 MB + f32精排: 2,930 MB)
  nprobe  k_fac      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
---------------------------------------------------------------------------
       4      1    60.62%     11,544     24,564      0.09   102.4x   217.9x
       4      4    69.66%     10,105     28,648      0.10    89.6x   254.1x
       4     10    69.71%      7,843     24,117      0.13    69.6x   213.9x
       4     20    69.71%      5,828     19,909      0.17    51.7x   176.6x
      16      1    70.75%      4,168     13,926      0.24    37.0x   123.5x
      16      4    86.10%      3,868     11,525      0.26    34.3x   102.2x
      16     10    86.36%      3,508     10,718      0.29    31.1x    95.1x
      16     20    86.37%      2,963     10,126      0.34    26.3x    89.8x
      32      1    73.30%      2,261      7,298      0.44    20.1x    64.7x
      32      4    90.77%      2,151      6,173      0.46    19.1x    54.7x
      32     10    91.26%      2,031      5,941      0.49    18.0x    52.7x
      32     20    91.28%      1,846      5,923      0.54    16.4x    52.5x
      64      1    74.95%      1,185      3,763      0.84    10.5x    33.4x
      64      4    94.23%      1,166      3,928      0.86    10.3x    34.8x
      64     10    95.00%      1,124      3,303      0.89    10.0x    29.3x
      64     20    95.02%      1,050      3,722      0.95     9.3x    33.0x
     128      1    75.88%        610      2,192      1.64     5.4x    19.4x
     128      4    96.52%        606      2,318      1.65     5.4x    20.6x
     128     10    97.49%        593      2,715      1.69     5.3x    24.1x
     128     20    97.54%        578      2,174      1.73     5.1x    19.3x
     256      1    76.38%        317      1,279      3.16     2.8x    11.3x
     256      4    97.75%        316      1,482      3.17     2.8x    13.1x
     256     10    98.95%        313      1,501      3.19     2.8x    13.3x
     256     20    99.03%        307      1,275      3.26     2.7x    11.3x

--- OPQ+IVF-PQ+Refine nlist=4096, m_pq=128 ---
构建(含 OPQ 训练): 273.4s
内存: 3,052 MB (PQ码: 122 MB + f32精排: 2,930 MB)
  nprobe  k_fac      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
---------------------------------------------------------------------------
       4      1    64.17%      5,669     19,469      0.18    50.3x   172.7x
       4      4    69.85%      5,495     17,619      0.18    48.7x   156.3x
       4     10    69.85%      4,610     15,986      0.22    40.9x   141.8x
       4     20    69.85%      3,882     13,435      0.26    34.4x   119.2x
      16      1    76.14%      2,564      9,071      0.39    22.7x    80.4x
      16      4    86.42%      2,488      8,029      0.40    22.1x    71.2x
      16     10    86.46%      2,292      7,892      0.44    20.3x    70.0x
      16     20    86.46%      2,070      7,784      0.48    18.4x    69.0x
      32      1    79.19%      1,478      5,456      0.68    13.1x    48.4x
      32      4    91.25%      1,443      5,155      0.69    12.8x    45.7x
      32     10    91.33%      1,388      5,040      0.72    12.3x    44.7x
      32     20    91.33%      1,289      5,048      0.78    11.4x    44.8x
      64      1    81.23%        803      3,007      1.24     7.1x    26.7x
      64      4    94.96%        789      3,127      1.27     7.0x    27.7x
      64     10    95.08%        774      2,965      1.29     6.9x    26.3x
      64     20    95.08%        740      2,819      1.35     6.6x    25.0x
     128      1    82.63%        423      1,900      2.36     3.8x    16.9x
     128      4    97.41%        420      1,913      2.38     3.7x    17.0x
     128     10    97.59%        415      2,065      2.41     3.7x    18.3x
     128     20    97.59%        404      1,431      2.48     3.6x    12.7x
     256      1    83.22%        221      1,063      4.53     2.0x     9.4x
     256      4    98.81%        218      1,083      4.59     1.9x     9.6x
     256     10    99.03%        218      1,114      4.59     1.9x     9.9x
     256     20    99.03%        216      1,065      4.64     1.9x     9.4x

--- OPQ+IVF-PQ+Refine nlist=4096, m_pq=192 ---
构建(含 OPQ 训练): 241.5s
内存: 3,113 MB (PQ码: 183 MB + f32精排: 2,930 MB)
  nprobe  k_fac      R@10     1T-QPS     MT-QPS   lat(ms)    1T/BF    MT/BF
---------------------------------------------------------------------------
       4      1    66.07%      6,312     19,560      0.16    56.0x   173.5x
       4      4    69.64%      5,928     18,007      0.17    52.6x   159.7x
       4     10    69.64%      4,991     16,868      0.20    44.3x   149.6x
       4     20    69.64%      4,107     15,231      0.24    36.4x   135.1x
      16      1    79.75%      1,927      6,902      0.52    17.1x    61.2x
      16      4    86.20%      1,848      6,253      0.54    16.4x    55.5x
      16     10    86.21%      1,746      6,342      0.57    15.5x    56.2x
      16     20    86.21%      1,637      6,280      0.61    14.5x    55.7x
      32      1    83.39%        991      4,749      1.01     8.8x    42.1x
      32      4    91.30%        988      3,968      1.01     8.8x    35.2x
      32     10    91.32%        952      3,789      1.05     8.4x    33.6x
      32     20    91.32%        908      3,522      1.10     8.1x    31.2x
      64      1    86.08%        515      2,269      1.94     4.6x    20.1x
      64      4    95.06%        512      2,321      1.95     4.5x    20.6x
      64     10    95.09%        501      2,130      2.00     4.4x    18.9x
      64     20    95.09%        485      2,008      2.06     4.3x    17.8x
     128      1    87.67%        263      1,282      3.80     2.3x    11.4x
     128      4    97.55%        264      1,348      3.78     2.3x    12.0x
     128     10    97.58%        261      1,323      3.83     2.3x    11.7x
     128     20    97.58%        257      1,344      3.90     2.3x    11.9x
     256      1    88.38%        135        742      7.39     1.2x     6.6x
     256      4    98.95%        136        755      7.36     1.2x     6.7x
     256     10    99.01%        135        735      7.42     1.2x     6.5x
     256     20    99.01%        134        764      7.44     1.2x     6.8x

======================================================================
全部完成！
======================================================================




######################
# Running Job 1 of 8 #
######################

Graph Index Full-Precision Build

               tag: graph-index-build
              file: c:/Users/Administrator/OneDrive/桌面/workspace/TriviumDB/diskann_data/cohere_1m_q1000\data.fbin
         data_type: float32
        max degree: 32
           L-build: 64
             alpha: 1
start point strategy: Medoid
    backedge ratio: 1
 Insert Batch Size: 128
 Batch Parallelism: 32
Intra Batch Candidates: none
start_point_strategy: Medoid
     build threads: 8
         Save Path: None

Building [00:01:50] ████████████████████████████████████████████████████████████████████████████████████████████████ 100

Index Build Time: 110.514803s
Vectors Inserted: 1000000
Kind: multi insert
Insert Latencies:
  average: 14093.968002047868us
      p90: 15741us
      p99: 17614us

 Ls,   KNN,    Avg cmps,   Avg hops,     QPS - mean(max),           Avg Latency,           p99 Latency,   Recall,   Threads
===========================================================================================================================
 10,    10,     490.334,     19.476,     7556.5 (7847.1),     132.0us (141.5us),       257.7us (303us),   0.7251,         1
 20,    10,     694.594,     29.734,     5478.4 (5495.6),     182.0us (183.0us),       313.7us (324us),   0.8202,         1
 40,    10,    1075.094,     49.453,     3351.1 (3397.7),     297.9us (303.2us),       497.3us (508us),   0.8902,         1
 80,    10,    1796.136,     87.986,     2019.7 (2032.8),     494.5us (498.8us),       712.3us (737us),   0.9363,         1
120,    10,    2507.405,    126.964,     1412.1 (1445.4),     708.0us (732.7us),     1149.3us (1375us),   0.9492,         1
200,    10,    3880.839,    205.894,       860.6 (864.2),   1161.3us (1166.7us),     2079.3us (2139us),   0.9613,         1
400,    10,     7137.34,    404.617,       453.3 (455.0),   2205.3us (2218.2us),     3712.7us (3868us),   0.9736,         1
600,    10,   10212.651,     604.01,       307.1 (309.6),   3255.6us (3304.1us),     5076.3us (5299us),   0.9797,         1
800,    10,   13169.471,    803.608,       233.7 (235.8),   4277.9us (4300.8us),     6621.3us (6875us),   0.9831,         1
 10,    10,     490.334,     19.476,   31460.2 (33812.3),     222.9us (231.4us),       480.0us (558us),   0.7251,         8
 20,    10,     694.594,     29.734,   20798.7 (21032.7),     331.1us (335.8us),       676.0us (690us),   0.8202,         8
 40,    10,    1075.094,     49.453,   13905.0 (13929.1),     528.0us (530.5us),     1007.7us (1017us),   0.8902,         8
 80,    10,    1796.136,     87.986,     7921.5 (8180.7),     920.2us (926.9us),     1672.3us (1716us),   0.9363,         8
120,    10,    2507.405,    126.964,     5604.2 (5773.5),   1318.7us (1345.8us),     2331.3us (2427us),   0.9492,         8
200,    10,    3880.839,    205.894,     3570.7 (3616.3),   2116.3us (2154.7us),     3649.0us (3794us),   0.9613,         8
400,    10,     7137.34,    404.617,     1892.6 (1944.6),   3955.9us (3973.3us),     6506.7us (6666us),   0.9736,         8
600,    10,   10212.651,     604.01,     1307.5 (1316.9),   5855.6us (5876.6us),     9601.0us (9713us),   0.9797,         8
800,    10,   13169.471,    803.608,     1007.8 (1025.3),   7565.9us (7602.0us),   12235.7us (12575us),   0.9831,         8



######################
# Running Job 2 of 8 #
######################

Graph Index Full-Precision Build

               tag: graph-index-build
              file: c:/Users/Administrator/OneDrive/桌面/workspace/TriviumDB/diskann_data/cohere_1m_q1000\data.fbin
         data_type: float32
        max degree: 32
           L-build: 64
             alpha: 1.2
start point strategy: Medoid
    backedge ratio: 1
 Insert Batch Size: 128
 Batch Parallelism: 32
Intra Batch Candidates: none
start_point_strategy: Medoid
     build threads: 8
         Save Path: None

Building [00:01:55] ████████████████████████████████████████████████████████████████████████████████████████████████ 100

Index Build Time: 115.790306s
Vectors Inserted: 1000000
Kind: multi insert
Insert Latencies:
  average: 14767.031997952132us
      p90: 16103us
      p99: 18349us

 Ls,   KNN,    Avg cmps,   Avg hops,     QPS - mean(max),           Avg Latency,           p99 Latency,   Recall,   Threads
===========================================================================================================================
 10,    10,     523.101,     18.943,     7302.8 (7319.8),     136.4us (137.0us),       252.0us (257us),   0.8155,         1
 20,    10,     723.699,     28.843,     5172.7 (5256.3),     192.8us (197.0us),       345.3us (364us),   0.8838,         1
 40,    10,    1103.692,     48.192,     3333.7 (3380.0),     299.4us (302.2us),       470.3us (476us),   0.9341,         1
 80,    10,    1852.851,     86.852,     1945.1 (1964.2),     513.5us (517.5us),       740.3us (747us),   0.9605,         1
120,    10,    2598.074,    126.186,     1266.2 (1281.2),     789.2us (795.7us),     1462.0us (1502us),   0.9708,         1
200,    10,    4035.557,    205.217,       818.3 (823.5),   1221.5us (1234.4us),     2099.3us (2171us),   0.9805,         1
400,    10,    7412.615,    404.312,       431.0 (433.2),   2319.7us (2336.7us),     3834.3us (3947us),   0.9872,         1
600,    10,   10565.808,     603.86,       296.2 (297.5),   3375.5us (3398.5us),     5535.3us (5686us),   0.9909,         1
800,    10,   13563.186,    803.468,       227.4 (227.8),   4396.1us (4409.2us),     7143.7us (7336us),   0.9937,         1
 10,    10,     523.101,     18.943,   29400.7 (31227.6),     241.9us (249.2us),       512.3us (546us),   0.8155,         8
 20,    10,     723.699,     28.843,   21943.9 (22316.4),     345.1us (346.4us),       660.7us (690us),   0.8838,         8
 40,    10,    1103.692,     48.192,   12716.7 (13191.2),     553.1us (554.4us),     1031.3us (1066us),   0.9341,         8
 80,    10,    1852.851,     86.852,     8000.7 (8069.9),     964.6us (972.1us),     1702.0us (1731us),   0.9605,         8
120,    10,    2598.074,    126.186,     5306.1 (5380.5),   1379.8us (1393.1us),     2561.3us (2648us),   0.9708,         8
200,    10,    4035.557,    205.217,     3401.3 (3467.4),   2166.2us (2167.2us),     4038.3us (4134us),   0.9805,         8
400,    10,    7412.615,    404.312,     1823.3 (1846.7),   4149.1us (4186.9us),     7306.0us (7505us),   0.9872,         8
600,    10,   10565.808,     603.86,     1242.1 (1270.2),   6177.7us (6244.3us),   10611.7us (10709us),   0.9909,         8
800,    10,   13563.186,    803.468,       920.2 (975.0),   8460.9us (9531.3us),   14681.0us (16638us),   0.9937,         8



######################
# Running Job 3 of 8 #
######################

Graph Index Full-Precision Build

               tag: graph-index-build
              file: c:/Users/Administrator/OneDrive/桌面/workspace/TriviumDB/diskann_data/cohere_1m_q1000\data.fbin
         data_type: float32
        max degree: 32
           L-build: 128
             alpha: 1
start point strategy: Medoid
    backedge ratio: 1
 Insert Batch Size: 128
 Batch Parallelism: 32
Intra Batch Candidates: none
start_point_strategy: Medoid
     build threads: 8
         Save Path: None

Building [00:02:58] ████████████████████████████████████████████████████████████████████████████████████████████████ 100

Index Build Time: 178.46328s
Vectors Inserted: 1000000
Kind: multi insert
Insert Latencies:
  average: 22761.0961218482us
      p90: 24586us
      p99: 28379us

 Ls,   KNN,    Avg cmps,   Avg hops,     QPS - mean(max),           Avg Latency,           p99 Latency,   Recall,   Threads
===========================================================================================================================
 10,    10,     466.437,      19.05,     8305.2 (8324.5),     119.8us (120.0us),       225.0us (226us),    0.762,         1
 20,    10,     642.689,     29.137,     5766.0 (5817.0),     172.9us (175.4us),       311.3us (321us),   0.8592,         1
 40,    10,     972.954,     48.384,     3714.9 (3741.0),     268.6us (271.4us),       446.3us (450us),   0.9209,         1
 80,    10,    1638.072,     86.967,     2163.2 (2186.5),     461.7us (464.4us),       681.7us (683us),   0.9566,         1
120,    10,    2303.196,    126.203,     1477.7 (1515.9),     676.7us (705.2us),     1092.7us (1285us),   0.9677,         1
200,    10,    3595.215,    205.222,       874.0 (886.5),   1143.8us (1169.4us),     2045.0us (2192us),   0.9785,         1
400,    10,    6641.168,    404.234,       479.0 (481.5),   2087.0us (2093.5us),     3573.7us (3649us),   0.9871,         1
600,    10,    9488.109,    603.703,       325.7 (327.2),   3069.7us (3097.9us),     5147.3us (5220us),   0.9907,         1
800,    10,   12196.865,    803.384,       250.0 (256.3),   4000.8us (4100.0us),     6680.3us (7114us),   0.9922,         1
 10,    10,     466.437,      19.05,   35329.9 (36145.4),     211.1us (213.0us),       428.7us (444us),    0.762,         8
 20,    10,     642.689,     29.137,   23951.1 (25055.7),     305.6us (307.9us),       633.7us (671us),   0.8592,         8
 40,    10,     972.954,     48.384,   14864.5 (15277.4),     493.0us (498.6us),      963.0us (1001us),   0.9209,         8
 80,    10,    1638.072,     86.967,     7600.4 (7754.0),     889.3us (891.9us),     1729.0us (1756us),   0.9566,         8
120,    10,    2303.196,    126.203,     5655.8 (5816.7),   1242.3us (1248.4us),     2406.0us (2427us),   0.9677,         8
200,    10,    3595.215,    205.222,     3479.4 (3611.5),   2046.7us (2104.9us),     3803.0us (3819us),   0.9785,         8
400,    10,    6641.168,    404.234,     1852.3 (1889.4),   3938.8us (3963.0us),     7046.0us (7095us),   0.9871,         8
600,    10,    9488.109,    603.703,     1307.5 (1331.9),   5814.0us (5841.4us),   10309.7us (10343us),   0.9907,         8
800,    10,   12196.865,    803.384,      980.4 (1001.3),   7488.3us (7587.2us),   13297.7us (13388us),   0.9922,         8



######################
# Running Job 4 of 8 #
######################

Graph Index Full-Precision Build

               tag: graph-index-build
              file: c:/Users/Administrator/OneDrive/桌面/workspace/TriviumDB/diskann_data/cohere_1m_q1000\data.fbin
         data_type: float32
        max degree: 32
           L-build: 128
             alpha: 1.2
start point strategy: Medoid
    backedge ratio: 1
 Insert Batch Size: 128
 Batch Parallelism: 32
Intra Batch Candidates: none
start_point_strategy: Medoid
     build threads: 8
         Save Path: None

Building [00:03:53] ████████████████████████████████████████████████████████████████████████████████████████████████ 100

Index Build Time: 233.587821s
Vectors Inserted: 1000000
Kind: multi insert
Insert Latencies:
  average: 29791.937284013824us
      p90: 32197us
      p99: 54758us

 Ls,   KNN,    Avg cmps,   Avg hops,     QPS - mean(max),           Avg Latency,           p99 Latency,   Recall,   Threads
===========================================================================================================================
 10,    10,     559.251,     19.009,     6890.2 (6897.2),     144.5us (144.8us),       281.0us (295us),   0.8298,         1
 20,    10,     783.264,     28.793,     4801.6 (4839.3),     207.7us (209.0us),       345.3us (351us),   0.8961,         1
 40,    10,    1211.325,     48.166,     3022.0 (3062.0),     330.3us (334.5us),       538.3us (545us),   0.9473,         1
 80,    10,    2037.448,     86.929,     1766.3 (1779.7),     565.5us (568.7us),       800.3us (810us),   0.9743,         1
120,    10,    2849.461,    126.141,     1219.4 (1259.0),     819.9us (835.6us),     1203.0us (1309us),   0.9822,         1
200,    10,    4411.495,    205.161,       769.2 (775.4),   1299.5us (1307.3us),     1858.0us (1923us),   0.9899,         1
400,    10,    8064.479,    404.303,       410.8 (413.1),   2433.7us (2448.6us),     3408.0us (3552us),   0.9951,         1
600,    10,   11456.938,    603.749,       280.7 (282.4),   3562.2us (3579.7us),     4792.7us (4873us),   0.9962,         1
800,    10,   14672.656,    803.382,       215.8 (217.7),   4632.9us (4669.7us),     6192.3us (6261us),    0.997,         1
 10,    10,     559.251,     19.009,   19540.5 (20290.1),     372.4us (376.8us),       867.3us (892us),   0.8298,         8
 20,    10,     783.264,     28.793,   14997.2 (17523.0),     491.8us (541.4us),     1013.0us (1132us),   0.8961,         8
 40,    10,    1211.325,     48.166,   11398.0 (11549.5),     615.3us (623.3us),     1195.0us (1236us),   0.9473,         8
 80,    10,    2037.448,     86.929,     6797.4 (7164.8),   1077.9us (1107.9us),     1908.0us (1993us),   0.9743,         8
120,    10,    2849.461,    126.141,     4843.6 (4944.4),   1509.1us (1525.0us),     2731.3us (2749us),   0.9822,         8
200,    10,    4411.495,    205.161,     3053.2 (3138.7),   2404.1us (2421.6us),     4313.0us (4366us),   0.9899,         8
400,    10,    8064.479,    404.303,     1691.8 (1701.3),   4545.6us (4564.0us),     7855.7us (8048us),   0.9951,         8
600,    10,   11456.938,    603.749,     1109.0 (1151.5),   6861.2us (6995.7us),   11601.7us (11819us),   0.9962,         8
800,    10,   14672.656,    803.382,       875.4 (895.9),   8787.1us (8838.1us),   14527.7us (14684us),    0.997,         8



######################
# Running Job 5 of 8 #
######################

Graph Index Full-Precision Build

               tag: graph-index-build
              file: c:/Users/Administrator/OneDrive/桌面/workspace/TriviumDB/diskann_data/cohere_1m_q1000\data.fbin
         data_type: float32
        max degree: 64
           L-build: 64
             alpha: 1
start point strategy: Medoid
    backedge ratio: 1
 Insert Batch Size: 128
 Batch Parallelism: 32
Intra Batch Candidates: none
start_point_strategy: Medoid
     build threads: 8
         Save Path: None

Building [00:02:40] ████████████████████████████████████████████████████████████████████████████████████████████████ 100

Index Build Time: 160.409522s
Vectors Inserted: 1000000
Kind: multi insert
Insert Latencies:
  average: 20458.25943939588us
      p90: 22372us
      p99: 39577us

 Ls,   KNN,    Avg cmps,   Avg hops,     QPS - mean(max),             Avg Latency,           p99 Latency,   Recall,   Threads
=============================================================================================================================
 10,    10,     706.479,     17.729,     5668.6 (5744.1),       175.8us (177.7us),       343.0us (350us),   0.7713,         1
 20,    10,    1027.464,     27.776,     3808.4 (3841.0),       262.0us (263.3us),       469.3us (474us),   0.8628,         1
 40,    10,    1612.794,     47.008,     2348.9 (2387.4),       425.1us (430.3us),       714.0us (733us),   0.9235,         1
 80,    10,    2729.446,     85.651,     1338.4 (1375.5),       747.5us (786.2us),     1299.7us (1613us),   0.9555,         1
120,    10,    3813.208,    125.052,       901.9 (906.9),     1108.1us (1113.4us),     1859.7us (1863us),   0.9671,         1
200,    10,    5860.298,    204.168,       581.1 (587.1),     1720.3us (1729.2us),     2961.7us (3104us),   0.9787,         1
400,    10,   10668.758,    403.374,       312.8 (315.0),     3196.1us (3215.1us),     5249.0us (5361us),   0.9888,         1
600,    10,   15162.642,    602.941,       210.1 (211.2),     4759.5us (4773.2us),     7755.3us (8057us),    0.992,         1
800,    10,   19457.951,    802.692,       165.0 (165.2),     6059.2us (6063.1us),     9291.3us (9333us),    0.994,         1
 10,    10,     706.479,     17.729,   21679.5 (23569.3),       326.9us (331.5us),       725.0us (771us),   0.7713,         8
 20,    10,    1027.464,     27.776,   15720.9 (15960.4),       484.5us (486.0us),       932.7us (997us),   0.8628,         8
 40,    10,    1612.794,     47.008,     9119.3 (9415.9),       793.1us (794.9us),     1587.7us (1641us),   0.9235,         8
 80,    10,    2729.446,     85.651,     5244.2 (5377.6),     1417.1us (1421.3us),     2642.3us (2729us),   0.9555,         8
120,    10,    3813.208,    125.052,     3845.8 (3863.6),     1980.4us (1995.7us),     3743.0us (3848us),   0.9671,         8
200,    10,    5860.298,    204.168,     2385.5 (2437.0),     3149.9us (3160.8us),     5887.3us (6009us),   0.9787,         8
400,    10,   10668.758,    403.374,     1279.4 (1303.7),     5920.4us (5931.2us),   10688.0us (10826us),   0.9888,         8
600,    10,   15162.642,    602.941,       881.0 (898.8),     8698.6us (8873.4us),   15080.3us (15213us),    0.992,         8
800,    10,   19457.951,    802.692,       693.1 (699.9),   11187.6us (11204.8us),   19493.0us (20292us),    0.994,         8



######################
# Running Job 6 of 8 #
######################

Graph Index Full-Precision Build

               tag: graph-index-build
              file: c:/Users/Administrator/OneDrive/桌面/workspace/TriviumDB/diskann_data/cohere_1m_q1000\data.fbin
         data_type: float32
        max degree: 64
           L-build: 64
             alpha: 1.2
start point strategy: Medoid
    backedge ratio: 1
 Insert Batch Size: 128
 Batch Parallelism: 32
Intra Batch Candidates: none
start_point_strategy: Medoid
     build threads: 8
         Save Path: None

Building [00:03:00] ████████████████████████████████████████████████████████████████████████████████████████████████ 100

Index Build Time: 180.34874s
Vectors Inserted: 1000000
Kind: multi insert
Insert Latencies:
  average: 23000.910149750416us
      p90: 26723us
      p99: 41795us

 Ls,   KNN,    Avg cmps,   Avg hops,     QPS - mean(max),             Avg Latency,           p99 Latency,   Recall,   Threads
=============================================================================================================================
 10,    10,     750.412,     16.637,     5049.3 (5081.3),       197.4us (198.1us),       365.7us (376us),   0.8819,         1
 20,    10,    1047.588,     26.301,     3475.4 (3503.1),       287.1us (289.1us),       510.3us (518us),   0.9313,         1
 40,    10,      1644.5,     45.608,     2174.9 (2180.3),       459.2us (459.9us),       774.3us (784us),   0.9644,         1
 80,    10,     2836.83,     84.703,     1219.5 (1260.1),       820.2us (854.5us),     1413.3us (1668us),   0.9794,         1
120,    10,    4008.365,      124.3,       826.9 (839.8),     1208.9us (1223.3us),     2184.3us (2243us),   0.9863,         1
200,    10,    6240.241,    203.743,       526.4 (533.8),     1899.1us (1920.0us),     3377.0us (3594us),   0.9928,         1
400,    10,   11384.648,      403.1,       277.3 (280.0),     3606.1us (3650.1us),     5972.3us (6012us),   0.9965,         1
600,    10,   16113.781,    602.752,       192.7 (193.2),     5187.7us (5206.8us),     8429.0us (8610us),   0.9973,         1
800,    10,   20561.992,    802.481,       149.2 (149.8),     6703.5us (6738.6us),   10718.0us (11039us),   0.9978,         1
 10,    10,     750.412,     16.637,   21483.0 (21784.6),       341.0us (347.0us),       708.0us (725us),   0.8819,         8
 20,    10,    1047.588,     26.301,   14039.5 (14624.2),       500.5us (503.9us),     1039.7us (1061us),   0.9313,         8
 40,    10,      1644.5,     45.608,     9137.2 (9263.2),       820.1us (824.2us),     1667.0us (1729us),   0.9644,         8
 80,    10,     2836.83,     84.703,     4718.6 (4922.9),     1487.6us (1495.1us),     3022.3us (3066us),   0.9794,         8
120,    10,    4008.365,      124.3,     3501.9 (3552.5),     2118.6us (2127.2us),     4313.3us (4357us),   0.9863,         8
200,    10,    6240.241,    203.743,     2193.5 (2248.4),     3460.7us (3471.0us),     6470.0us (6631us),   0.9928,         8
400,    10,   11384.648,      403.1,     1126.8 (1177.7),     6766.5us (6933.9us),   12545.0us (12777us),   0.9965,         8
600,    10,   16113.781,    602.752,       765.0 (786.9),    9949.3us (10409.7us),   18358.0us (19311us),   0.9973,         8
800,    10,   20561.992,    802.481,       616.2 (619.0),   12406.7us (12560.6us),   22239.3us (23149us),   0.9978,         8



######################
# Running Job 7 of 8 #
######################

Graph Index Full-Precision Build

               tag: graph-index-build
              file: c:/Users/Administrator/OneDrive/桌面/workspace/TriviumDB/diskann_data/cohere_1m_q1000\data.fbin
         data_type: float32
        max degree: 64
           L-build: 128
             alpha: 1
start point strategy: Medoid
    backedge ratio: 1
 Insert Batch Size: 128
 Batch Parallelism: 32
Intra Batch Candidates: none
start_point_strategy: Medoid
     build threads: 8
         Save Path: None

Building [00:04:15] ████████████████████████████████████████████████████████████████████████████████████████████████ 100

Index Build Time: 255.090396s
Vectors Inserted: 1000000
Kind: multi insert
Insert Latencies:
  average: 32532.761935236143us
      p90: 38741us
      p99: 60051us

 Ls,   KNN,    Avg cmps,   Avg hops,     QPS - mean(max),           Avg Latency,           p99 Latency,   Recall,   Threads
===========================================================================================================================
 10,    10,      627.15,     17.609,     6046.0 (6074.5),     164.8us (166.0us),       310.0us (316us),   0.8271,         1
 20,    10,     851.749,     27.396,     4338.6 (4393.1),     230.0us (235.3us),       418.3us (436us),   0.9008,         1
 40,    10,    1301.258,     46.611,     2719.9 (2738.2),     367.1us (370.6us),       651.7us (661us),   0.9486,         1
 80,    10,    2214.899,     85.363,     1593.8 (1610.5),     626.8us (632.6us),     1057.3us (1080us),   0.9728,         1
120,    10,    3129.007,    124.797,     1055.1 (1073.5),     947.3us (957.7us),     1846.0us (1892us),   0.9807,         1
200,    10,    4886.522,      204.2,       656.6 (662.5),   1522.3us (1529.4us),     2632.0us (2693us),   0.9899,         1
400,    10,    8985.743,    403.454,       339.2 (344.4),   2948.4us (3018.9us),     5366.0us (5915us),    0.995,         1
600,    10,    12779.51,    603.052,       229.4 (236.5),   4360.6us (4427.9us),     7694.0us (8138us),   0.9967,         1
800,    10,   16356.182,    802.805,       179.2 (185.0),   5582.9us (5680.3us),     9271.7us (9490us),   0.9976,         1
 10,    10,      627.15,     17.609,   24466.7 (24985.6),     285.1us (288.7us),       637.0us (646us),   0.8271,         8
 20,    10,     851.749,     27.396,   15521.6 (15999.0),     428.2us (435.5us),       928.3us (959us),   0.9008,         8
 40,    10,    1301.258,     46.611,   10615.9 (10793.2),     656.7us (660.6us),     1406.3us (1431us),   0.9486,         8
 80,    10,    2214.899,     85.363,     5807.2 (6021.2),   1169.9us (1186.3us),     2551.7us (2596us),   0.9728,         8
120,    10,    3129.007,    124.797,     4274.8 (4385.8),   1681.2us (1716.2us),     3614.7us (3713us),   0.9807,         8
200,    10,    4886.522,      204.2,     2571.7 (2587.9),   2731.1us (2749.2us),     5639.0us (5877us),   0.9899,         8
400,    10,    8985.743,    403.454,     1396.6 (1462.7),   5121.3us (5128.2us),   10395.3us (10617us),    0.995,         8
600,    10,    12779.51,    603.052,       967.8 (979.6),   7568.7us (7592.5us),   14907.3us (15003us),   0.9967,         8
800,    10,   16356.182,    802.805,       738.8 (756.7),   9785.9us (9980.0us),   18649.3us (19189us),   0.9976,         8



######################
# Running Job 8 of 8 #
######################

Graph Index Full-Precision Build

               tag: graph-index-build
              file: c:/Users/Administrator/OneDrive/桌面/workspace/TriviumDB/diskann_data/cohere_1m_q1000\data.fbin
         data_type: float32
        max degree: 64
           L-build: 128
             alpha: 1.2
start point strategy: Medoid
    backedge ratio: 1
 Insert Batch Size: 128
 Batch Parallelism: 32
Intra Batch Candidates: none
start_point_strategy: Medoid
     build threads: 8
         Save Path: None

Building [00:06:36] ████████████████████████████████████████████████████████████████████████████████████████████████ 100

Index Build Time: 396.191472s
Vectors Inserted: 1000000
Kind: multi insert
Insert Latencies:
  average: 50546.805452451044us
      p90: 71736us
      p99: 90857us

 Ls,   KNN,    Avg cmps,   Avg hops,     QPS - mean(max),             Avg Latency,           p99 Latency,   Recall,   Threads
=============================================================================================================================
 10,    10,     867.257,     16.684,     4334.3 (4365.6),       230.1us (231.2us),       397.7us (406us),    0.904,         1
 20,    10,    1233.106,     26.283,     3046.3 (3091.1),       327.7us (331.5us),       537.7us (556us),   0.9484,         1
 40,    10,    1953.179,     45.556,     1871.3 (1872.6),       533.7us (534.2us),       825.3us (835us),   0.9748,         1
 80,    10,    3365.151,     84.727,     1028.1 (1069.1),       972.8us (998.3us),     1681.7us (1906us),    0.987,         1
120,    10,    4722.734,     124.32,       704.0 (720.1),     1420.3us (1458.1us),     2622.7us (2863us),   0.9935,         1
200,    10,     7296.33,    203.717,       450.8 (451.4),     2217.7us (2220.4us),     3768.7us (3829us),   0.9967,         1
400,    10,   13154.999,    403.118,       241.9 (244.9),     4134.0us (4216.1us),     6697.7us (7031us),   0.9984,         1
600,    10,    18488.92,    602.786,       169.9 (171.5),     5886.1us (5976.3us),     9328.7us (9758us),   0.9992,         1
800,    10,    23476.48,    802.547,       130.9 (131.2),     7637.1us (7650.2us),   11385.0us (11432us),   0.9996,         1
 10,    10,     867.257,     16.684,   18303.3 (18515.8),       404.1us (406.7us),       800.7us (814us),    0.904,         8
 20,    10,    1233.106,     26.283,   11837.5 (12385.4),       605.5us (617.1us),     1198.3us (1247us),   0.9484,         8
 40,    10,    1953.179,     45.556,     7204.1 (7416.0),       989.6us (994.2us),     1960.0us (2028us),   0.9748,         8
 80,    10,    3365.151,     84.727,     4064.6 (4159.2),     1807.5us (1822.1us),     3471.3us (3530us),    0.987,         8
120,    10,    4722.734,     124.32,     2922.7 (2985.8),     2553.4us (2568.0us),     4807.0us (4894us),   0.9935,         8
200,    10,     7296.33,    203.717,     1878.5 (1900.9),     4052.2us (4062.7us),     7539.3us (7770us),   0.9967,         8
400,    10,   13154.999,    403.118,     1030.3 (1041.5),     7569.2us (7590.5us),   13818.7us (14098us),   0.9984,         8
600,    10,    18488.92,    602.786,       695.4 (702.0),   11045.0us (11241.1us),   19485.3us (19757us),   0.9992,         8
800,    10,    23476.48,    802.547,       552.5 (558.1),   14026.3us (14084.0us),   24168.7us (24851us),   0.9996,         8





======================================================================
  QuIVer 批量 benchmark
  待跑数据集: 10 个
    - glove-100: GloVe-1.18M (100-d)
    - sift-128: SIFT-1M (128-d)
    - minilm-384: MiniLM-1M (384-d)
    - wolt-clip-512: Wolt CLIP-1M (512-d)
    - random-1m: Synthetic-LR-1M (768-d)
    - sphere-1m: Random-Sphere-1M (768-d)
    - bge-m3-1024: BGE-M3-1M (1024-d)
    - gist-960: GIST-1M (960-d)
    - dbpedia-1536: DBpedia-OpenAI-1M (1536-d)
    - dbpedia-3072: DBpedia-OpenAI-3072-1M (3072-d)
  模式: all
  起始子实验: 1d
  日志目录: C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB\research\bench_logs
  开始时间: 2026-05-08 03:49:39
======================================================================

  [1/10] 开始 glove-100...

======================================================================
  数据集: GloVe-1.18M (glove-100, 100-d)
  模式: all, 起始: 1d
  日志: C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB\research\bench_logs\glove-100.log
  开始时间: 2026-05-08 03:49:39
======================================================================
   Compiling triviumdb v0.7.0 (C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB)
    Finished `bench` profile [optimized] target(s) in 13.40s
     Running benches\bench_sensitivity.rs (target\release\deps\bench_sensitivity-d2b34d7a5832ef72.exe)
加载 glove-100 数据集...
加载完成! 1183514 x 100 训练集, 10000 测试集, 耗时 0.19s

======================================================================
实验 1: 参数敏感性分析
数据集: glove-100 (1183514 x 100)
======================================================================

计算 brute-force 基准 QPS...
Brute-force 基准: QPS=127.1, latency=7.87ms/q

--- 1d: ef_search 精细 Recall-QPS 曲线 (m=32, ef_c=128, α=1.2) ---
ef          R@10(%)       MT-QPS       1T-QPS    lat(ms)      vs BF
------------------------------------------------------------------
8             3.17%       463276        37405      0.03    3644.9x
12            6.34%       249177        23493      0.04    1960.4x
16            9.27%       184360        18161      0.06    1450.5x
20           12.01%       174522        15439      0.06    1373.1x
24           14.49%       146497        12707      0.08    1152.6x
32           19.30%       111762         9834      0.10     879.3x
40           23.10%        89479         7163      0.14     704.0x
48           26.39%        77006         6681      0.15     605.9x
56           29.38%        65643         5844      0.17     516.5x
64           32.08%        59372         5115      0.20     467.1x
80           36.14%        47486         4183      0.24     373.6x
96           39.59%        41823         3601      0.28     329.0x
112          42.38%        37056         3153      0.32     291.5x
128          44.86%        32873         2794      0.36     258.6x
160          48.33%        26680         2313      0.43     209.9x
192          51.03%        22916         1993      0.50     180.3x
224          53.22%        20128         1728      0.58     158.4x
256          55.24%        17818         1561      0.64     140.2x
320          58.07%        14574         1271      0.79     114.7x
384          60.17%        12340         1068      0.94      97.1x
448          62.03%        10723          956      1.05      84.4x
512          63.57%         9438          844      1.19      74.3x
640          65.94%         7583          685      1.46      59.7x
768          67.84%         6364          571      1.75      50.1x
896          69.38%         5492          503      1.99      43.2x
1024         70.67%         4846          445      2.25      38.1x

--- 1e: 不同 m 的 Pareto 曲线 (ef_c=128, α=1.2) ---
m      ef          R@10(%)       MT-QPS      vs BF
--------------------------------------------------
8      32            4.84%       347188    2731.6x
8      64           11.54%       156395    1230.5x
8      128          21.27%        77125     606.8x
8      256          33.13%        39982     314.6x
8      512          44.33%        21007     165.3x
8      1024         53.55%        10960      86.2x

16     32           11.52%       173299    1363.5x
16     64           22.27%        88797     698.6x
16     128          35.23%        48784     383.8x
16     256          48.29%        26229     206.4x
16     512          58.91%        14034     110.4x
16     1024         66.77%         7417      58.4x

32     32           19.43%       101549     799.0x
32     64           32.11%        53172     418.3x
32     128          44.88%        30022     236.2x
32     256          55.22%        16995     133.7x
32     512          63.61%         9123      71.8x
32     1024         70.68%         4446      35.0x

48     32           24.11%        75857     596.8x
48     64           36.17%        40755     320.6x
48     128          47.35%        22881     180.0x
48     256          56.75%        12665      99.6x
48     512          64.48%         6889      54.2x
48     1024         71.31%         3535      27.8x

64     32           25.97%        48139     378.7x
64     64           37.77%        28277     222.5x
64     128          48.18%        15981     125.7x
64     256          57.13%         8623      67.8x
64     512          64.64%         4695      36.9x
64     1024         71.43%         2451      19.3x


======================================================================
实验 2: 单线程 vs 多线程查询扩展性
数据集: glove-100 (1183514 x 100)
======================================================================
索引构建: 81.9s (14456 vecs/s), Hot 618 MB

--- 2a: 不同 ef 下的 单线程 vs 多线程 QPS ---
ef             1T-QPS       MT-QPS      R@10(%)        加速比
----------------------------------------------------------
64               5116        60851       32.20%      11.9x
128              2810        33253       44.77%      11.8x
256              1516        17953       55.15%      11.8x
512               845         9444       63.58%      11.2x
1024              444         4853       70.64%      10.9x

--- 2b: 线程数扩展性 (ef=64) ---
线程数                 QPS        相对1线程
--------------------------------------
1                  5207         1.0x
2                 10571         2.0x
4                 21584         4.1x
8                 39457         7.6x
16                55331        10.6x

全部实验完成!

  ✅ glove-100 完成! 耗时 1041s (17.4min)

  [2/10] 开始 sift-128...

======================================================================
  数据集: SIFT-1M (sift-128, 128-d)
  模式: all, 起始: 1d
  日志: C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB\research\bench_logs\sift-128.log
  开始时间: 2026-05-08 04:07:01
======================================================================
    Finished `bench` profile [optimized] target(s) in 0.24s
     Running benches\bench_sensitivity.rs (target\release\deps\bench_sensitivity-d2b34d7a5832ef72.exe)
加载 sift-128 数据集...
加载完成! 1000000 x 128 训练集, 10000 测试集, 耗时 0.38s

======================================================================
实验 1: 参数敏感性分析
数据集: sift-128 (1000000 x 128)
======================================================================

计算 brute-force 基准 QPS...
Brute-force 基准: QPS=118.4, latency=8.45ms/q

--- 1d: ef_search 精细 Recall-QPS 曲线 (m=32, ef_c=128, α=1.2) ---
ef          R@10(%)       MT-QPS       1T-QPS    lat(ms)      vs BF
------------------------------------------------------------------
8             3.82%       404926        34758      0.03    3420.8x
12            5.40%       303582        25932      0.04    2564.7x
16            6.68%       241755        21057      0.05    2042.3x
20            7.70%       207188        17934      0.06    1750.3x
24            8.64%       177775        15763      0.06    1501.8x
32           10.26%       149782        12720      0.08    1265.4x
40           11.58%       124602        10505      0.10    1052.6x
48           12.73%       107278         9226      0.11     906.3x
56           13.80%        96152         8325      0.12     812.3x
64           14.85%        87212         7368      0.14     736.8x
80           16.51%        72789         6126      0.16     614.9x
96           17.98%        62527         5345      0.19     528.2x
112          19.25%        55473         4701      0.21     468.6x
128          20.40%        49679         4226      0.24     419.7x
160          22.42%        41294         3519      0.28     348.8x
192          24.20%        34968         2990      0.33     295.4x
224          25.76%        30644         2676      0.37     258.9x
256          27.18%        27442         2382      0.42     231.8x
320          29.48%        21815         1962      0.51     184.3x
384          31.40%        19094         1690      0.59     161.3x
448          33.14%        16408         1485      0.67     138.6x
512          34.73%        14788         1333      0.75     124.9x
640          37.40%        12023         1085      0.92     101.6x
768          39.59%        10109          935      1.07      85.4x
896          41.43%         8804          816      1.23      74.4x
1024         43.01%         7835          734      1.36      66.2x

--- 1e: 不同 m 的 Pareto 曲线 (ef_c=128, α=1.2) ---
m      ef          R@10(%)       MT-QPS      vs BF
--------------------------------------------------
8      32            4.49%       310512    2623.2x
8      64            7.55%       163204    1378.7x
8      128          11.33%        89407     755.3x
8      256          15.37%        50892     429.9x
8      512          19.90%        26613     224.8x
8      1024         24.75%        12960     109.5x

16     32            7.49%       187832    1586.8x
16     64           11.55%       103866     877.5x
16     128          16.49%        62451     527.6x
16     256          22.08%        35448     299.5x
16     512          28.31%        19262     162.7x
16     1024         34.98%        10347      87.4x

32     32           10.32%       137576    1162.2x
32     64           14.85%        75970     641.8x
32     128          20.43%        44800     378.5x
32     256          27.25%        25068     211.8x
32     512          34.73%        14175     119.7x
32     1024         43.00%         7593      64.1x

48     32           11.51%       104955     886.7x
48     64           16.39%        58133     491.1x
48     128          22.43%        34553     291.9x
48     256          29.70%        20254     171.1x
48     512          37.87%        11377      96.1x
48     1024         46.77%         6233      52.7x

64     32           12.30%        92569     782.0x
64     64           17.40%        51198     432.5x
64     128          23.64%        29838     252.1x
64     256          31.19%        17453     147.4x
64     512          39.73%         9804      82.8x
64     1024         49.05%         5429      45.9x


======================================================================
实验 2: 单线程 vs 多线程查询扩展性
数据集: sift-128 (1000000 x 128)
======================================================================
索引构建: 55.6s (17995 vecs/s), Hot 522 MB

--- 2a: 不同 ef 下的 单线程 vs 多线程 QPS ---
ef             1T-QPS       MT-QPS      R@10(%)        加速比
----------------------------------------------------------
64               7371        86472       14.80%      11.7x
128              4063        49807       20.46%      12.3x
256              2360        27189       27.21%      11.5x
512              1297        14412       34.76%      11.1x
1024              729         7843       42.98%      10.8x

--- 2b: 线程数扩展性 (ef=64) ---
线程数                 QPS        相对1线程
--------------------------------------
1                  7318         1.0x
2                 13921         1.9x
4                 29608         4.0x
8                 54401         7.4x
16                77889        10.6x

全部实验完成!

  ✅ sift-128 完成! 耗时 721s (12.0min)

  [3/10] 开始 minilm-384...

======================================================================
  数据集: MiniLM-1M (minilm-384, 384-d)
  模式: all, 起始: 1d
  日志: C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB\research\bench_logs\minilm-384.log
  开始时间: 2026-05-08 04:19:02
======================================================================
    Finished `bench` profile [optimized] target(s) in 0.22s
     Running benches\bench_sensitivity.rs (target\release\deps\bench_sensitivity-d2b34d7a5832ef72.exe)
加载 minilm-384 数据集...
加载完成! 1000000 x 384 训练集, 1000 测试集, 耗时 1.12s

======================================================================
实验 1: 参数敏感性分析
数据集: minilm-384 (1000000 x 384)
======================================================================

计算 brute-force 基准 QPS...
Brute-force 基准: QPS=37.6, latency=26.62ms/q

--- 1d: ef_search 精细 Recall-QPS 曲线 (m=32, ef_c=128, α=1.2) ---
ef          R@10(%)       MT-QPS       1T-QPS    lat(ms)      vs BF
------------------------------------------------------------------
8            36.15%       164064        16835      0.06    4368.0x
12           50.80%       111834        12879      0.08    2977.5x
16           59.87%        83995        11148      0.09    2236.3x
20           65.91%        82744        10132      0.10    2203.0x
24           70.94%        75796         8515      0.12    2018.0x
32           76.80%        65303         7324      0.14    1738.6x
40           81.27%        56667         6010      0.17    1508.7x
48           84.21%        49121         5314      0.19    1307.8x
56           86.48%        44378         4869      0.21    1181.5x
64           88.09%        41106         4377      0.23    1094.4x
80           90.44%        34009         3642      0.27     905.4x
96           91.70%        28882         3086      0.32     769.0x
112          92.86%        27053         2754      0.36     720.3x
128          93.82%        23506         2528      0.40     625.8x
160          95.22%        20002         2120      0.47     532.5x
192          95.73%        16849         1824      0.55     448.6x
224          96.36%        15177         1614      0.62     404.1x
256          96.90%        13468         1441      0.69     358.6x
320          97.49%        11193         1208      0.83     298.0x
384          97.77%         9513         1021      0.98     253.3x
448          98.23%         8327          864      1.16     221.7x
512          98.37%         7371          803      1.24     196.2x
640          98.79%         5876          652      1.53     156.5x
768          99.00%         5124          558      1.79     136.4x
896          99.06%         4444          483      2.07     118.3x
1024         99.17%         3800          426      2.35     101.2x

--- 1e: 不同 m 的 Pareto 曲线 (ef_c=128, α=1.2) ---
m      ef          R@10(%)       MT-QPS      vs BF
--------------------------------------------------
8      32           44.06%       147872    3936.9x
8      64           61.28%        96767    2576.3x
8      128          74.56%        50440    1342.9x
8      256          83.33%        27400     729.5x
8      512          88.31%        14847     395.3x
8      1024         91.34%         7406     197.2x

16     32           65.99%       107266    2855.9x
16     64           80.17%        47680    1269.4x
16     128          88.58%        35185     936.8x
16     256          93.29%        19971     531.7x
16     512          96.35%        10641     283.3x
16     1024         97.44%         5550     147.8x

32     32           76.87%        68634    1827.3x
32     64           88.14%        38078    1013.8x
32     128          93.75%        21195     564.3x
32     256          96.92%        12517     333.3x
32     512          98.39%         6892     183.5x
32     1024         99.20%         3683      98.0x

48     32           81.93%        50931    1356.0x
48     64           90.88%        30239     805.1x
48     128          95.46%        17929     477.3x
48     256          97.74%        10129     269.7x
48     512          98.92%         5483     146.0x
48     1024         99.46%         2909      77.4x

64     32           83.88%        42238    1124.6x
64     64           92.19%        25082     667.8x
64     128          96.21%        14608     388.9x
64     256          97.96%         7961     211.9x
64     512          99.14%         4480     119.3x
64     1024         99.63%         2417      64.3x


======================================================================
实验 2: 单线程 vs 多线程查询扩展性
数据集: minilm-384 (1000000 x 384)
======================================================================
索引构建: 82.8s (12084 vecs/s), Hot 583 MB

--- 2a: 不同 ef 下的 单线程 vs 多线程 QPS ---
ef             1T-QPS       MT-QPS      R@10(%)        加速比
----------------------------------------------------------
64               4212        42045       88.09%      10.0x
128              2552        24162       93.73%       9.5x
256              1441        13571       96.90%       9.4x
512               790         7394       98.36%       9.4x
1024              431         3886       99.19%       9.0x

--- 2b: 线程数扩展性 (ef=64) ---
线程数                 QPS        相对1线程
--------------------------------------
1                  4375         1.0x
2                  8476         1.9x
4                 16792         3.8x
8                 30624         7.0x
16                41622         9.5x

全部实验完成!

  ✅ minilm-384 完成! 耗时 765s (12.8min)

  [4/10] 开始 wolt-clip-512...

======================================================================
  数据集: Wolt CLIP-1M (wolt-clip-512, 512-d)
  模式: all, 起始: 1d
  日志: C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB\research\bench_logs\wolt-clip-512.log
  开始时间: 2026-05-08 04:31:47
======================================================================
    Finished `bench` profile [optimized] target(s) in 0.24s
     Running benches\bench_sensitivity.rs (target\release\deps\bench_sensitivity-d2b34d7a5832ef72.exe)
加载 wolt-clip-512 数据集...
加载完成! 1000000 x 512 训练集, 1000 测试集, 耗时 0.74s

======================================================================
实验 1: 参数敏感性分析
数据集: wolt-clip-512 (1000000 x 512)
======================================================================

计算 brute-force 基准 QPS...
Brute-force 基准: QPS=27.3, latency=36.57ms/q

--- 1d: ef_search 精细 Recall-QPS 曲线 (m=32, ef_c=128, α=1.2) ---
ef          R@10(%)       MT-QPS       1T-QPS    lat(ms)      vs BF
------------------------------------------------------------------
8            32.93%       259168        29001      0.03    9476.8x
12           42.94%       189948        23855      0.04    6945.7x
16           48.89%       157235        19706      0.05    5749.5x
20           53.42%       126366        16695      0.06    4620.7x
24           56.72%       122886        13725      0.07    4493.5x
32           61.40%       102144        12314      0.08    3735.0x
40           64.71%        86417        10067      0.10    3159.9x
48           67.82%        77363         8879      0.11    2828.9x
56           69.39%        69601         8107      0.12    2545.1x
64           70.68%        62367         7152      0.14    2280.5x
80           73.35%        55178         6226      0.16    2017.7x
96           74.89%        48251         5410      0.18    1764.4x
112          76.36%        42839         4834      0.21    1566.5x
128          77.40%        39064         4180      0.24    1428.4x
160          78.83%        32763         3685      0.27    1198.0x
192          80.10%        28058         3132      0.32    1026.0x
224          80.50%        24688         2724      0.37     902.8x
256          81.66%        22560         2510      0.40     824.9x
320          82.72%        18396         2054      0.49     672.7x
384          83.36%        15764         1768      0.57     576.4x
448          84.21%        13707         1573      0.64     501.2x
512          84.28%        12128         1426      0.70     443.5x
640          85.09%        10082         1165      0.86     368.7x
768          85.34%         8459         1004      1.00     309.3x
896          85.36%         7364          870      1.15     269.3x
1024         85.68%         6606          749      1.33     241.6x

--- 1e: 不同 m 的 Pareto 曲线 (ef_c=128, α=1.2) ---
m      ef          R@10(%)       MT-QPS      vs BF
--------------------------------------------------
8      32           33.03%       205326    7508.0x
8      64           44.31%       115302    4216.2x
8      128          51.52%        64086    2343.4x
8      256          58.21%        35498    1298.0x
8      512          62.97%        18835     688.7x
8      1024         66.55%         9405     343.9x

16     32           49.71%       155790    5696.7x
16     64           61.67%        93799    3429.9x
16     128          68.74%        49591    1813.4x
16     256          73.89%        27833    1017.8x
16     512          77.58%        14905     545.0x
16     1024         80.12%         7749     283.3x

32     32           61.32%       106943    3910.5x
32     64           71.00%        69267    2532.9x
32     128          77.52%        37671    1377.5x
32     256          81.47%        21355     780.9x
32     512          84.29%        11490     420.1x
32     1024         85.70%         6197     226.6x

48     32           66.66%        84651    3095.4x
48     64           74.99%        44629    1631.9x
48     128          80.77%        31546    1153.5x
48     256          84.28%        18097     661.7x
48     512          86.23%        10071     368.3x
48     1024         87.63%         5423     198.3x

64     32           68.45%        77354    2828.6x
64     64           76.19%        50311    1839.7x
64     128          81.76%        28258    1033.3x
64     256          85.42%        15855     579.8x
64     512          87.41%         9020     329.8x
64     1024         88.10%         4889     178.8x


======================================================================
实验 2: 单线程 vs 多线程查询扩展性
数据集: wolt-clip-512 (1000000 x 512)
======================================================================
索引构建: 58.4s (17120 vecs/s), Hot 614 MB

--- 2a: 不同 ef 下的 单线程 vs 多线程 QPS ---
ef             1T-QPS       MT-QPS      R@10(%)        加速比
----------------------------------------------------------
64               6417        68136       71.02%      10.6x
128              4268        39104       77.40%       9.2x
256              2424        22521       81.71%       9.3x
512              1382        12439       84.24%       9.0x
1024              774         6533       85.45%       8.4x

--- 2b: 线程数扩展性 (ef=64) ---
线程数                 QPS        相对1线程
--------------------------------------
1                  7397         1.0x
2                 14037         1.9x
4                 27371         3.7x
8                 50692         6.9x
16                67820         9.2x

全部实验完成!

  ✅ wolt-clip-512 完成! 耗时 560s (9.3min)

  [5/10] 开始 random-1m...

======================================================================
  数据集: Synthetic-LR-1M (random-1m, 768-d)
  模式: all, 起始: 1d
  日志: C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB\research\bench_logs\random-1m.log
  开始时间: 2026-05-08 04:41:07
======================================================================
    Finished `bench` profile [optimized] target(s) in 0.21s
     Running benches\bench_sensitivity.rs (target\release\deps\bench_sensitivity-d2b34d7a5832ef72.exe)
加载 random-1m 数据集...
加载完成! 1000000 x 768 训练集, 1000 测试集, 耗时 2.39s

======================================================================
实验 1: 参数敏感性分析
数据集: random-1m (1000000 x 768)
======================================================================

计算 brute-force 基准 QPS...
Brute-force 基准: QPS=18.8, latency=53.31ms/q

--- 1d: ef_search 精细 Recall-QPS 曲线 (m=32, ef_c=128, α=1.2) ---
ef          R@10(%)       MT-QPS       1T-QPS    lat(ms)      vs BF
------------------------------------------------------------------
8            10.41%       128631        15530      0.06    6856.9x
12           14.82%        89795        12589      0.08    4786.7x
16           18.43%        83521         9752      0.10    4452.2x
20           21.41%        66465         8284      0.12    3543.1x
24           24.00%        59239         7502      0.13    3157.8x
32           28.84%        47964         5584      0.18    2556.8x
40           32.71%        40834         4819      0.21    2176.7x
48           36.19%        34320         4194      0.24    1829.5x
56           39.22%        31155         3755      0.27    1660.8x
64           41.76%        28091         3440      0.29    1497.4x
80           46.65%        22843         2668      0.37    1217.7x
96           50.73%        19773         2380      0.42    1054.0x
112          54.06%        17442         2057      0.49     929.8x
128          56.82%        15562         1826      0.55     829.5x
160          61.41%        12582         1515      0.66     670.7x
192          64.77%        10800         1302      0.77     575.7x
224          67.99%         9400         1143      0.88     501.1x
256          70.33%         8350         1021      0.98     445.1x
320          74.28%         6810          822      1.22     363.0x
384          77.41%         5691          712      1.40     303.4x
448          79.45%         4978          621      1.61     265.4x
512          81.30%         4442          548      1.83     236.8x
640          84.35%         3610          449      2.23     192.5x
768          86.46%         2995          375      2.66     159.7x
896          88.36%         2552          329      3.04     136.0x
1024         89.57%         2326          295      3.38     124.0x

--- 1e: 不同 m 的 Pareto 曲线 (ef_c=128, α=1.2) ---
m      ef          R@10(%)       MT-QPS      vs BF
--------------------------------------------------
8      32            8.62%       121177    6459.6x
8      64           14.64%        66702    3555.7x
8      128          21.90%        37401    1993.8x
8      256          30.27%        19848    1058.1x
8      512          40.04%        10189     543.1x
8      1024         50.30%         5226     278.6x

16     32           17.89%        83979    4476.6x
16     64           27.54%        47368    2525.1x
16     128          38.90%        24128    1286.2x
16     256          50.90%        12777     681.1x
16     512          63.22%         6508     346.9x
16     1024         74.17%         3522     187.7x

32     32           29.07%        49196    2622.5x
32     64           42.13%        28110    1498.5x
32     128          56.92%        15609     832.1x
32     256          70.28%         8370     446.2x
32     512          81.13%         4325     230.6x
32     1024         89.49%         2335     124.5x

48     32           35.79%        36833    1963.4x
48     64           50.94%        20234    1078.6x
48     128          66.35%        10999     586.3x
48     256          79.38%         5835     311.0x
48     512          88.63%         3096     165.1x
48     1024         94.67%         1692      90.2x

64     32           39.57%        30214    1610.6x
64     64           55.28%        15456     823.9x
64     128          71.70%         8975     478.5x
64     256          84.02%         4841     258.0x
64     512          92.03%         2589     138.0x
64     1024         96.77%         1398      74.5x


======================================================================
实验 2: 单线程 vs 多线程查询扩展性
数据集: random-1m (1000000 x 768)
======================================================================
索引构建: 125.4s (7973 vecs/s), Hot 675 MB

--- 2a: 不同 ef 下的 单线程 vs 多线程 QPS ---
ef             1T-QPS       MT-QPS      R@10(%)        加速比
----------------------------------------------------------
64               3189        22528       41.92%       7.1x
128              1802        16005       56.92%       8.9x
256              1031         8614       70.42%       8.4x
512               543         4543       81.44%       8.4x
1024              292         2354       89.73%       8.1x

--- 2b: 线程数扩展性 (ef=64) ---
线程数                 QPS        相对1线程
--------------------------------------
1                  3438         1.0x
2                  6718         2.0x
4                 13073         3.8x
8                 23368         6.8x
16                29073         8.5x

全部实验完成!

  ✅ random-1m 完成! 耗时 1161s (19.4min)

  [6/10] 开始 sphere-1m...

======================================================================
  数据集: Random-Sphere-1M (sphere-1m, 768-d)
  模式: all, 起始: 1d
  日志: C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB\research\bench_logs\sphere-1m.log
  开始时间: 2026-05-08 05:00:28
======================================================================
    Finished `bench` profile [optimized] target(s) in 0.24s
     Running benches\bench_sensitivity.rs (target\release\deps\bench_sensitivity-d2b34d7a5832ef72.exe)
加载 sphere-1m 数据集...
加载完成! 1000000 x 768 训练集, 1000 测试集, 耗时 2.60s

======================================================================
实验 1: 参数敏感性分析
数据集: sphere-1m (1000000 x 768)
======================================================================

计算 brute-force 基准 QPS...
Brute-force 基准: QPS=18.7, latency=53.37ms/q

--- 1d: ef_search 精细 Recall-QPS 曲线 (m=32, ef_c=128, α=1.2) ---
ef          R@10(%)       MT-QPS       1T-QPS    lat(ms)      vs BF
------------------------------------------------------------------
8             0.06%       164233        19711      0.05    8764.8x
12            0.08%       101605        13635      0.07    5422.5x
16            0.12%        80725        10710      0.09    4308.1x
20            0.17%        63983         9006      0.11    3414.7x
24            0.16%        55918         7540      0.13    2984.2x
32            0.21%        45336         5542      0.18    2419.5x
40            0.23%        37112         4611      0.22    1980.6x
48            0.30%        31370         3903      0.26    1674.2x
56            0.34%        27472         3473      0.29    1466.1x
64            0.40%        24314         2791      0.36    1297.6x
80            0.56%        19938         2471      0.40    1064.0x
96            0.68%        16935         2088      0.48     903.8x
112           0.77%        14686         1819      0.55     783.8x
128           0.85%        12890         1585      0.63     687.9x
160           1.03%        10516         1278      0.78     561.2x
192           1.24%         8790         1062      0.94     469.1x
224           1.44%         7678          933      1.07     409.7x
256           1.66%         6697          813      1.23     357.4x
320           2.13%         5497          655      1.53     293.4x
384           2.53%         4530          552      1.81     241.8x
448           2.89%         3919          482      2.07     209.2x
512           3.32%         3527          421      2.38     188.2x
640           4.20%         2766          338      2.96     147.6x
768           4.94%         2379          285      3.51     127.0x
896           5.72%         2052          245      4.09     109.5x
1024          6.55%         1795          213      4.69      95.8x

--- 1e: 不同 m 的 Pareto 曲线 (ef_c=128, α=1.2) ---
m      ef          R@10(%)       MT-QPS      vs BF
--------------------------------------------------
8      32            0.07%       136180    7267.7x
8      64            0.16%        68590    3660.5x
8      128           0.32%        35181    1877.5x
8      256           0.49%        17939     957.4x
8      512           1.10%         9108     486.1x
8      1024          2.11%         4550     242.8x

16     32            0.17%        80403    4291.0x
16     64            0.32%        40670    2170.5x
16     128           0.64%        21313    1137.5x
16     256           1.04%        11034     588.9x
16     512           1.98%         5476     292.3x
16     1024          3.91%         2959     157.9x

32     32            0.32%        46182    2464.6x
32     64            0.56%        23896    1275.3x
32     128           0.99%        12473     665.7x
32     256           1.94%         6588     351.6x
32     512           3.55%         3375     180.1x
32     1024          6.46%         1758      93.8x

48     32            0.32%        31643    1688.7x
48     64            0.55%        16958     905.0x
48     128           1.13%         9014     481.0x
48     256           2.00%         4686     250.1x
48     512           4.07%         2433     129.9x
48     1024          7.90%         1269      67.7x

64     32            0.31%        25012    1334.9x
64     64            0.66%        13231     706.1x
64     128           1.34%         7041     375.8x
64     256           2.75%         3618     193.1x
64     512           4.97%         1923     102.6x
64     1024          9.27%         1005      53.6x


======================================================================
实验 2: 单线程 vs 多线程查询扩展性
数据集: sphere-1m (1000000 x 768)
======================================================================
索引构建: 153.0s (6536 vecs/s), Hot 675 MB

--- 2a: 不同 ef 下的 单线程 vs 多线程 QPS ---
ef             1T-QPS       MT-QPS      R@10(%)        加速比
----------------------------------------------------------
64               2827        23924        0.40%       8.5x
128              1598        13097        0.74%       8.2x
256               792         6673        1.57%       8.4x
512               415         3541        3.30%       8.5x
1024              217         1812        6.25%       8.3x

--- 2b: 线程数扩展性 (ef=64) ---
线程数                 QPS        相对1线程
--------------------------------------
1                  3064         1.0x
2                  5831         1.9x
4                 11544         3.8x
8                 19065         6.2x
16                24822         8.1x

全部实验完成!

  ✅ sphere-1m 完成! 耗时 1381s (23.0min)

  [7/10] 开始 bge-m3-1024...

======================================================================
  数据集: BGE-M3-1M (bge-m3-1024, 1024-d)
  模式: all, 起始: 1d
  日志: C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB\research\bench_logs\bge-m3-1024.log
  开始时间: 2026-05-08 05:23:30
======================================================================
    Finished `bench` profile [optimized] target(s) in 0.28s
     Running benches\bench_sensitivity.rs (target\release\deps\bench_sensitivity-d2b34d7a5832ef72.exe)
加载 bge-m3-1024 数据集...
加载完成! 1000000 x 1024 训练集, 1000 测试集, 耗时 3.98s

======================================================================
实验 1: 参数敏感性分析
数据集: bge-m3-1024 (1000000 x 1024)
======================================================================

计算 brute-force 基准 QPS...
Brute-force 基准: QPS=14.0, latency=71.47ms/q

--- 1d: ef_search 精细 Recall-QPS 曲线 (m=32, ef_c=128, α=1.2) ---
ef          R@10(%)       MT-QPS       1T-QPS    lat(ms)      vs BF
------------------------------------------------------------------
8            42.07%       133844        18698      0.05    9565.6x
12           59.14%       115089        14841      0.07    8225.2x
16           69.29%        95493        12184      0.08    6824.7x
20           75.39%        82624        10640      0.09    5905.0x
24           80.15%        77762         8807      0.11    5557.5x
32           85.32%        63965         7604      0.13    4571.5x
40           88.43%        54399         6612      0.15    3887.8x
48           91.12%        49347         5937      0.17    3526.7x
56           92.86%        45351         5314      0.19    3241.1x
64           93.81%        41182         4930      0.20    2943.2x
80           94.88%        36563         4191      0.24    2613.1x
96           96.12%        31882         3692      0.27    2278.5x
112          96.70%        28048         3345      0.30    2004.6x
128          97.06%        25911         3028      0.33    1851.8x
160          98.01%        22023         2638      0.38    1574.0x
192          98.42%        18991         2275      0.44    1357.2x
224          98.65%        16751         2021      0.49    1197.2x
256          98.72%        15200         1847      0.54    1086.3x
320          99.06%        12407         1539      0.65     886.7x
384          99.16%        10352         1324      0.76     739.9x
448          99.33%         9239         1167      0.86     660.3x
512          99.35%         8325         1059      0.94     595.0x
640          99.47%         6689          853      1.17     478.0x
768          99.49%         5573          700      1.43     398.3x
896          99.56%         4774          632      1.58     341.2x
1024         99.57%         4284          559      1.79     306.2x

--- 1e: 不同 m 的 Pareto 曲线 (ef_c=128, α=1.2) ---
m      ef          R@10(%)       MT-QPS      vs BF
--------------------------------------------------
8      32           30.67%       134735    9629.2x
8      64           46.63%        77115    5511.3x
8      128          64.26%        43054    3077.0x
8      256          76.91%        23028    1645.8x
8      512          82.73%        12695     907.3x
8      1024         85.45%         6468     462.3x

16     32           62.27%        94012    6718.9x
16     64           79.00%        58079    4150.8x
16     128          90.29%        32057    2291.1x
16     256          94.77%        19170    1370.1x
16     512          96.41%        10409     743.9x
16     1024         96.91%         5358     383.0x

32     32           85.53%        65568    4686.0x
32     64           93.82%        37198    2658.5x
32     128          97.07%        25433    1817.7x
32     256          98.77%        14752    1054.3x
32     512          99.38%         8105     579.2x
32     1024         99.58%         4176     298.5x

48     32           91.71%        55788    3987.0x
48     64           96.52%        34748    2483.4x
48     128          98.40%        21447    1532.8x
48     256          99.52%        11988     856.7x
48     512          99.76%         6691     478.2x
48     1024         99.91%         3452     246.7x

64     32           94.71%        46817    3345.9x
64     64           98.15%        30274    2163.6x
64     128          99.14%        18125    1295.4x
64     256          99.82%        10390     742.6x
64     512          99.88%         5681     406.0x
64     1024         99.96%         2966     212.0x


======================================================================
实验 2: 单线程 vs 多线程查询扩展性
数据集: bge-m3-1024 (1000000 x 1024)
======================================================================
索引构建: 78.3s (12765 vecs/s), Hot 736 MB

--- 2a: 不同 ef 下的 单线程 vs 多线程 QPS ---
ef             1T-QPS       MT-QPS      R@10(%)        加速比
----------------------------------------------------------
64               4930        42499       93.82%       8.6x
128              3091        26652       97.08%       8.6x
256              1831        15194       98.66%       8.3x
512              1041         8327       99.38%       8.0x
1024              558         4287       99.59%       7.7x

--- 2b: 线程数扩展性 (ef=64) ---
线程数                 QPS        相对1线程
--------------------------------------
1                  4884         1.0x
2                  9849         2.0x
4                 18329         3.8x
8                 32041         6.6x
16                43059         8.8x

全部实验完成!

  ✅ bge-m3-1024 完成! 耗时 767s (12.8min)

  [8/10] 开始 gist-960...

======================================================================
  数据集: GIST-1M (gist-960, 960-d)
  模式: all, 起始: 1d
  日志: C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB\research\bench_logs\gist-960.log
  开始时间: 2026-05-08 05:36:17
======================================================================
    Finished `bench` profile [optimized] target(s) in 0.21s
     Running benches\bench_sensitivity.rs (target\release\deps\bench_sensitivity-d2b34d7a5832ef72.exe)
加载 gist-960 数据集...
加载完成! 1000000 x 960 训练集, 1000 测试集, 耗时 3.20s

======================================================================
实验 1: 参数敏感性分析
数据集: gist-960 (1000000 x 960)
======================================================================

计算 brute-force 基准 QPS...
Brute-force 基准: QPS=14.9, latency=66.99ms/q

--- 1d: ef_search 精细 Recall-QPS 曲线 (m=32, ef_c=128, α=1.2) ---
ef          R@10(%)       MT-QPS       1T-QPS    lat(ms)      vs BF
------------------------------------------------------------------
8             0.59%       429387        50823      0.02   28766.5x
12            0.83%       284779        34223      0.03   19078.5x
16            1.00%       247807        31521      0.03   16601.6x
20            1.17%       237801        26469      0.04   15931.3x
24            1.25%       216882        23967      0.04   14529.8x
32            1.38%       163191        19820      0.05   10932.8x
40            1.50%       120767        16945      0.06    8090.7x
48            1.76%       124544        14738      0.07    8343.7x
56            1.93%       112140        12611      0.08    7512.8x
64            2.01%       103832        11657      0.09    6956.2x
80            2.28%        79431         9845      0.10    5321.4x
96            2.43%        74789         8651      0.12    5010.4x
112           2.57%        66244         7648      0.13    4438.0x
128           2.70%        59094         6629      0.15    3958.9x
160           2.96%        47108         5550      0.18    3156.0x
192           3.12%        41234         4670      0.21    2762.5x
224           3.29%        33895         4233      0.24    2270.8x
256           3.38%        31644         3783      0.26    2120.0x
320           3.54%        25808         3110      0.32    1729.0x
384           3.68%        21683         2629      0.38    1452.6x
448           3.80%        18574         2259      0.44    1244.3x
512           3.90%        16459         1996      0.50    1102.7x
640           4.04%        13175         1577      0.63     882.6x
768           4.11%        11030         1335      0.75     739.0x
896           4.19%         9221         1141      0.88     617.8x
1024          4.22%         8176          996      1.00     547.7x

--- 1e: 不同 m 的 Pareto 曲线 (ef_c=128, α=1.2) ---
m      ef          R@10(%)       MT-QPS      vs BF
--------------------------------------------------
8      32            0.43%       453679   30393.9x
8      64            0.55%       268889   18014.0x
8      128           0.70%       147857    9905.6x
8      256           0.77%        74642    5000.6x
8      512           0.84%        40184    2692.1x
8      1024          0.84%        21104    1413.8x

16     32            0.76%       261835   17541.4x
16     64            1.15%       191858   12853.3x
16     128           1.42%        93230    6245.8x
16     256           1.73%        46847    3138.5x
16     512           1.90%        23665    1585.4x
16     1024          1.99%        12122     812.1x

32     32            1.48%       187112   12535.4x
32     64            2.06%       103046    6903.5x
32     128           2.69%        56615    3792.9x
32     256           3.39%        27530    1844.4x
32     512           3.88%        15433    1033.9x
32     1024          4.26%         8013     536.8x

48     32            1.95%       137349    9201.6x
48     64            2.50%        78832    5281.3x
48     128           3.32%        46984    3147.7x
48     256           4.21%        24610    1648.8x
48     512           4.95%        12638     846.6x
48     1024          5.63%         6463     433.0x

64     32            2.25%       102411    6860.9x
64     64            3.06%        67693    4535.0x
64     128           4.03%        40001    2679.8x
64     256           5.33%        20546    1376.5x
64     512           6.30%        10939     732.8x
64     1024          7.23%         5719     383.2x


======================================================================
实验 2: 单线程 vs 多线程查询扩展性
数据集: gist-960 (1000000 x 960)
======================================================================
索引构建: 75.2s (13302 vecs/s), Hot 720 MB

--- 2a: 不同 ef 下的 单线程 vs 多线程 QPS ---
ef             1T-QPS       MT-QPS      R@10(%)        加速比
----------------------------------------------------------
64              11051       110428        2.00%      10.0x
128              6720        60402        2.68%       9.0x
256              3808        32889        3.37%       8.6x
512              1990        16599        3.85%       8.3x
1024              998         8476        4.23%       8.5x

--- 2b: 线程数扩展性 (ef=64) ---
线程数                 QPS        相对1线程
--------------------------------------
1                 12085         1.0x
2                 20915         1.7x
4                 44462         3.7x
8                 74590         6.2x
16               108171         9.0x

全部实验完成!

  ✅ gist-960 完成! 耗时 749s (12.5min)

  [9/10] 开始 dbpedia-1536...

======================================================================
  数据集: DBpedia-OpenAI-1M (dbpedia-1536, 1536-d)
  模式: all, 起始: 1d
  日志: C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB\research\bench_logs\dbpedia-1536.log
  开始时间: 2026-05-08 05:48:46
======================================================================
    Finished `bench` profile [optimized] target(s) in 0.20s
     Running benches\bench_sensitivity.rs (target\release\deps\bench_sensitivity-d2b34d7a5832ef72.exe)
加载 dbpedia-1536 数据集...
加载完成! 990000 x 1536 训练集, 10000 测试集, 耗时 5.75s

======================================================================
实验 1: 参数敏感性分析
数据集: dbpedia-1536 (990000 x 1536)
======================================================================

计算 brute-force 基准 QPS...
Brute-force 基准: QPS=8.1, latency=124.16ms/q

--- 1d: ef_search 精细 Recall-QPS 曲线 (m=32, ef_c=128, α=1.2) ---
ef          R@10(%)       MT-QPS       1T-QPS    lat(ms)      vs BF
------------------------------------------------------------------
8            42.96%        89903        12466      0.08   11162.0x
12           62.32%        69035         9350      0.11    8571.1x
16           73.24%        56016         7900      0.13    6954.7x
20           79.91%        48207         6904      0.14    5985.2x
24           84.07%        42985         6144      0.16    5336.9x
32           89.12%        34347         5103      0.20    4264.3x
40           91.70%        30595         4251      0.24    3798.6x
48           93.43%        26809         3847      0.26    3328.4x
56           94.52%        23895         3492      0.29    2966.7x
64           95.34%        22035         3111      0.32    2735.8x
80           96.43%        18628         2676      0.37    2312.7x
96           97.04%        16408         2343      0.43    2037.1x
112          97.46%        14364         2078      0.48    1783.4x
128          97.74%        12996         1893      0.53    1613.5x
160          98.18%        10817         1583      0.63    1343.0x
192          98.44%         9318         1362      0.73    1156.9x
224          98.66%         8211         1194      0.84    1019.4x
256          98.83%         7309         1077      0.93     907.5x
320          98.99%         6017          882      1.13     747.1x
384          99.14%         5140          762      1.31     638.2x
448          99.25%         4464          668      1.50     554.2x
512          99.33%         3969          543      1.84     492.8x
640          99.43%         2675          476      2.10     332.1x
768          99.51%         2753          412      2.42     341.8x
896          99.57%         2385          358      2.79     296.1x
1024         99.61%         2104          318      3.15     261.2x

--- 1e: 不同 m 的 Pareto 曲线 (ef_c=128, α=1.2) ---
m      ef          R@10(%)       MT-QPS      vs BF
--------------------------------------------------
8      32           51.09%        80320    9972.1x
8      64           70.39%        43497    5400.4x
8      128          82.50%        25115    3118.2x
8      256          88.81%        13785    1711.5x
8      512          92.03%         7474     927.9x
8      1024         93.90%         3918     486.4x

16     32           76.43%        54667    6787.2x
16     64           88.29%        32272    4006.8x
16     128          93.79%        18490    2295.7x
16     256          96.45%        10504    1304.2x
16     512          97.79%         5680     705.2x
16     1024         98.55%         2979     369.9x

32     32           89.18%        35068    4353.9x
32     64           95.33%        22210    2757.4x
32     128          97.72%        13044    1619.4x
32     256          98.84%         7307     907.2x
32     512          99.32%         3960     491.6x
32     1024         99.61%         2101     260.8x

48     32           92.75%        25846    3208.9x
48     64           97.14%        14621    1815.3x
48     128          98.67%         8808    1093.6x
48     256          99.37%         5015     622.6x
48     512          99.66%         2666     331.0x
48     1024         99.83%         1437     178.4x

64     32           94.27%        22122    2746.5x
64     64           97.80%        12148    1508.2x
64     128          99.06%         7265     902.0x
64     256          99.60%         4084     507.0x
64     512          99.79%         2223     276.0x
64     1024         99.89%         1193     148.1x


======================================================================
实验 2: 单线程 vs 多线程查询扩展性
数据集: dbpedia-1536 (990000 x 1536)
======================================================================
索引构建: 141.6s (6992 vecs/s), Hot 849 MB

--- 2a: 不同 ef 下的 单线程 vs 多线程 QPS ---
ef             1T-QPS       MT-QPS      R@10(%)        加速比
----------------------------------------------------------
64               3082        21893       95.33%       7.1x
128              1885        12898       97.72%       6.8x
256              1073         7251       98.83%       6.8x
512               594         3939       99.32%       6.6x
1024              318         2092       99.61%       6.6x

--- 2b: 线程数扩展性 (ef=64) ---
线程数                 QPS        相对1线程
--------------------------------------
1                  3174         1.0x
2                  6227         2.0x
4                 12024         3.8x
8                 19699         6.2x
16                21581         6.8x

全部实验完成!

  ✅ dbpedia-1536 完成! 耗时 2827s (47.1min)


加载 dbpedia-3072 数据集...
加载完成! 990000 x 3072 训练集, 10000 测试集, 耗时 13.61s

======================================================================
实验 1: 参数敏感性分析
数据集: dbpedia-3072 (990000 x 3072)
======================================================================

计算 brute-force 基准 QPS...
PS C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB> ^C
PS C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB> cargo bench --bench bench_sensitivity
   Compiling triviumdb v0.7.0 (C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB)
    Finished `bench` profile [optimized] target(s) in 14.92s
     Running benches\bench_sensitivity.rs (target\release\deps\bench_sensitivity-d2b34d7a5832ef72.exe)
加载 dbpedia-3072 数据集...
加载完成! 990000 x 3072 训练集, 10000 测试集, 耗时 12.92s

======================================================================
实验 1: 参数敏感性分析
数据集: dbpedia-3072 (990000 x 3072)
======================================================================

计算 brute-force 基准 QPS (1000q)...
Brute-force 基准: QPS=4.6, latency=217.24ms/q

--- 1d: ef_search 精细 Recall-QPS 曲线 (m=32, ef_c=128, α=1.2) ---
ef          R@10(%)       MT-QPS       1T-QPS    lat(ms)      vs BF
------------------------------------------------------------------
8            50.64%        32526         8376      0.12    7066.1x
12           69.97%        35417         6679      0.15    7694.0x
16           79.12%        30491         5670      0.18    6624.0x
20           83.80%        26291         5034      0.20    5711.5x
24           87.19%        23738         4494      0.22    5156.9x
32           90.74%        20301         3567      0.28    4410.1x
40           92.82%        17589         3246      0.31    3821.0x
48           94.15%        15642         2891      0.35    3398.0x
56           95.02%        14083         2594      0.39    3059.5x
64           95.65%        12910         2352      0.43    2804.6x
80           96.58%        10883         1993      0.50    2364.3x
96           97.18%         9440         1757      0.57    2050.8x
112          97.61%         8266         1532      0.65    1795.6x
128          97.94%         7428         1390      0.72    1613.7x
160          98.26%         6265         1150      0.87    1361.0x
192          98.52%         5342         1007      0.99    1160.4x
224          98.71%         4691          889      1.13    1019.2x
256          98.84%         4171          781      1.28     906.0x
320          99.02%         3413          655      1.53     741.4x
384          99.15%         2902          559      1.79     630.4x
448          99.27%         2551          487      2.05     554.1x
512          99.34%         2259          431      2.32     490.7x
640          99.44%         1845          353      2.84     400.8x
768          99.51%         1567          299      3.35     340.4x
896          99.57%         1355          249      4.01     294.4x
1024         99.61%         1194          218      4.59     259.3x

--- 1e: 不同 m 的 Pareto 曲线 (ef_c=128, α=1.2) ---
m      ef          R@10(%)       MT-QPS      vs BF
--------------------------------------------------
8      32           55.25%        41341    8981.0x
8      64           71.53%        23609    5128.9x
8      128          82.27%        13827    3003.9x
8      256          88.46%         7700    1672.9x
8      512          91.97%         4092     888.9x
8      1024         93.82%         2126     461.9x

16     32           78.45%        28865    6270.7x
16     64           88.53%        18057    3922.8x
16     128          93.87%        10566    2295.4x
16     256          96.47%         5862    1273.4x
16     512          97.79%         3072     667.4x
16     1024         98.50%         1523     330.9x

32     32           90.77%        16354    3552.8x
32     64           95.67%        10386    2256.4x
32     128          97.94%         5665    1230.6x
32     256          98.84%         3199     694.9x
32     512          99.34%         1713     372.2x
32     1024         99.61%          930     202.0x

48     32           94.28%        12670    2752.4x
48     64           97.49%         7268    1578.9x
48     128          98.81%         4296     933.2x
48     256          99.38%         2437     529.4x
48     512          99.66%         1306     283.7x
48     1024         99.81%          709     154.0x

64     32           95.77%        10808    2347.9x
64     64           98.32%         5834    1267.3x
64     128          99.22%         3591     780.2x
64     256          99.60%         1969     427.8x
64     512          99.79%         1080     234.7x
64     1024         99.89%          599     130.2x


======================================================================
实验 2: 单线程 vs 多线程查询扩展性
数据集: dbpedia-3072 (990000 x 3072)
======================================================================
索引构建: 261.7s (3783 vecs/s), Hot 1212 MB

--- 2a: 不同 ef 下的 单线程 vs 多线程 QPS ---
ef             1T-QPS       MT-QPS      R@10(%)        加速比
----------------------------------------------------------
64               2274        12806       95.69%       5.6x
128              1383         7486       97.93%       5.4x
256               783         4131       98.85%       5.3x
512               429         2260       99.34%       5.3x
1024              227         1204       99.61%       5.3x

--- 2b: 线程数扩展性 (ef=64) ---
线程数                 QPS        相对1线程
--------------------------------------
1                  2344         1.0x
2                  4638         2.0x
4                  7772         3.3x
8                 10958         4.7x
16                12653         5.4x

全部实验完成!


======================================================================
  DiskANN Rust PQ Benchmark
  数据集: Cohere-1M (768-d, cosine)
  硬件: Ryzen 7 7840HS, 32GB DDR5-5600
  日期: 2026-05-12
======================================================================

Baseline: DiskANN Rust PQ (graph-index-build-pq)
  配置: R=32, L_build=128, alpha=1.2, PQ chunks=96
  multi_insert: batch_size=128, batch_parallelism=32
  seed: 13076402859301299683

>>> 模式 A: 纯 PQ 搜索 (use_fp_for_search=false) <<<

PQ 训练时间: 26.753s
图构建时间: 257.856s (总计 285s)
插入延迟: avg=33001us, p90=36428us, p99=45446us
内存: PQ codes (~96 MB) + 邻接表 + float32 vectors (~3 GB)

  Ls    R@10     1T-QPS     8T-QPS    1T-lat(us)   p99-lat(us)
-------------------------------------------------------------
  10   69.51%      4,086     23,752       244.1        476
  20   87.22%      2,767     15,702       361.0        646
  40   94.27%      1,866      9,939       535.3        896
  80   97.14%      1,048      6,212       953.8      1,746
 120   98.09%        779      4,337     1,283.1      1,837
 200   99.05%        502      2,779     1,993.0      2,723
 400   99.63%        254      1,477     3,945.4      5,726
 600   99.74%        182      1,085     5,510.8      8,552
 800   99.79%        139        842     7,183.2     10,112


>>> 模式 B: PQ 构建 + float32 rerank (use_fp_for_search=true) <<<

PQ 训练时间: 26.416s
图构建时间: 261.176s (总计 288s)
插入延迟: avg=33426us, p90=35725us, p99=40779us
内存: PQ codes (~96 MB) + 邻接表 + float32 vectors (~3 GB)

  Ls    R@10     1T-QPS     8T-QPS    1T-lat(us)   p99-lat(us)
-------------------------------------------------------------
  10   79.78%      6,319     25,006       157.7        318
  20   88.73%      4,326     16,032       230.6        481
  40   94.93%      2,732     10,535       365.5        593
  80   97.32%      1,617      6,093       617.9        891
 120   98.33%      1,156      4,382       864.3      1,171
 200   98.99%        736      2,810     1,357.3      1,794
 400   99.59%        394      1,505     2,539.1      3,459
 600   99.81%        270      1,022     3,700.9      5,087
 800   99.87%        207        801     4,829.5      6,742


注: 纯 PQ 模式 recall 天花板 ~99.8%，PQ+FP rerank 后 recall 更高。
    PQ+FP 的 1T-QPS 高于纯 PQ，因为 FP rerank 使得 beam search
    在较小的 Ls 即可达到相同 recall，从而整体更快。


======================================================================
  DiskANN Rust SSD (Disk-Index) Benchmark
  数据集: Cohere-1M (768-d, cosine)
  硬件: Ryzen 7 7840HS, 32GB DDR5-5600, NVMe SSD
  日期: 2026-05-12
======================================================================

Baseline: DiskANN Rust SSD (disk-index)
  构建配置: R=64, L_build=128, PQ chunks=96, build_ram_limit=8GB
  搜索配置: beam_width=4, num_nodes_to_cache=50000, 8 线程
  构建时间: 759.866s (~12.7 分钟)

冷热内存分布:
  Hot RAM:
    PQ compressed codes:  91.6 MB
    PQ pivots:             0.8 MB
    总计:                 92.3 MB
  Cold Storage (磁盘):
    disk.index:         3906.3 MB (3.81 GB, sector-aligned 图+向量)
    build_pq_pivots:       0.8 MB

工作原理:
  - PQ 压缩码 (~92 MB) 常驻 RAM，用于 beam search 中的近似距离计算
  - 每一跳需要从 SSD 读取 sector (4KB aligned)，sector 包含:
    - 该节点的邻接表 (邻居 ID 列表)
    - 该节点的 float32 完整向量
  - beam_width 控制每跳并行读取的 sector 数量
  - 与 QuIVer 的核心区别: DiskANN SSD 每一跳都有磁盘 I/O，
    而 QuIVer 导航全程零 I/O (仅 final rerank 访问 float32)

搜索结果 (8 线程, beam_width=4):

  Ls    R@10       QPS    avg-lat(us)  p95-lat(us)  p99.9-lat(us)  avg-IOs  IO-time(us)  CPU-time(us)  Cache-Hit%
-----------------------------------------------------------------------------------------------------------------
  10   76.11%     8,264       958.7       1,470         2,272       27.4       783.0       156.1         0.0%
  20   90.83%     5,518     1,442.3       2,131         4,145       36.4     1,223.3       197.9         0.0%
  40   96.21%     3,392     2,349.2       3,005         3,828       54.7     2,032.9       292.2         0.0%
  80   98.54%     1,858     4,296.8       5,114         8,342       93.0     3,786.1       483.6         0.0%
 120   99.31%     1,285     6,201.3       7,009         8,337      131.8     5,502.3       669.6         0.0%
 200   99.76%       771    10,350.2      11,641        22,905      210.5     9,264.0     1,053.2         0.0%
 400   99.93%       287    27,862.0      30,290       292,383      408.8    25,755.2     2,069.2         0.0%

注: Cache Hit 0% 表示 num_nodes_to_cache=50000 在 Load 模式下未生效。
    所有查询的图遍历数据都通过磁盘 I/O 获取。
    IO time 占总延迟的 ~80%，CPU time 仅 ~15%，确认瓶颈在 SSD IOPS。

对比 QuIVer (同硬件, 8T):
  Recall  QuIVer-8T   DiskANN-SSD   QuIVer/SSD
  ~95%     36,729       3,392        10.8x
  ~97%     21,021       ~2,600        8.1x
  ~99%     11,837       1,285         9.2x
  ~99.5%    6,032         771         7.8x


======================================================================
  FAISS IVF+RaBitQ Benchmark
  数据集: Cohere-1M (768-d, cosine/IP)
  硬件: Ryzen 7 7840HS, 32GB DDR5-5600, 16 核
  FAISS 版本: 1.13.2
  日期: 2026-05-12
======================================================================

注: RaBitQ 标准版 (非 FastScan) 在 768-d 下 QPS 极低 (10-40 1T-QPS)，
    不具备实际对比价值，仅记录少量数据作为参考。
    FastScan 版 (RaBitQfs) 使用 SIMD 优化，是 RaBitQ 的实用配置。

----------------------------------------------------------------------
  模式 A: 裸 IVF+RaBitQfs (无 Refine)
  recall 封顶 ~80%（1-bit 量化信息损失）
----------------------------------------------------------------------

>>> IVF1024+RaBitQfs <<<
构建: 28.0s
内存: ~110 MB (RaBitQ码: 107 MB + 质心: 3 MB)

  nprobe    R@10     1T-QPS     MT-QPS    1T-lat(us)
-----------------------------------------------------
       8   70.31%      4,800      7,579        208.3
      16   74.38%      2,491      7,897        401.4
      32   76.79%      1,441      3,755        693.9
      64   78.20%        762      2,157      1,311.7
     128   78.97%        405      1,462      2,469.2
     256   79.22%        198        875      5,040.8
     512   79.17%         93        466     10,802.3

>>> IVF4096+RaBitQfs <<<
构建: 250.3s
内存: ~119 MB (RaBitQ码: 107 MB + 质心: 12 MB)

  nprobe    R@10     1T-QPS     MT-QPS    1T-lat(us)
-----------------------------------------------------
       8   69.80%      5,187     17,911        192.8
      16   74.05%      3,500      9,925        285.7
      32   76.75%      1,878      5,595        532.5
      64   78.70%        978      3,316      1,022.5
     128   79.73%        504      2,150      1,986.1
     256   80.24%        252      1,130      3,961.0
     512   80.53%        128        567      7,842.6

注: 裸 RaBitQ (1-bit 量化, 无 rerank) recall 封顶 ~80%。
    这是因为 1-bit 量化在 768-d cosine 嵌入上信息损失过大。
    要达到 95%+ recall 必须加 Refine (SQ8 或 FP32 rerank)。


----------------------------------------------------------------------
  模式 B: IVF+RaBitQfs+Refine(SQ8) — RaBitQ 最强实用配置
  FastScan 粗筛 + SQ8 精排
----------------------------------------------------------------------

>>> IVF1024+RaBitQfs+Refine(SQ8) <<<
构建: 23.4s

  nprobe  k_fac    R@10     1T-QPS     MT-QPS    1T-lat(us)
------------------------------------------------------------
      16      1   74.38%      2,602      7,937        384.3
      16      2   86.36%      2,794      7,808        358.0
      16      4   88.38%      2,725      5,906        367.0
      16     10   88.68%      2,471      6,922        404.7
      16     20   88.69%      2,116      6,518        472.7
      32      1   76.79%      1,518      4,824        658.7
      32      2   90.64%      1,526      4,458        655.3
      32      4   93.13%      1,485      4,184        673.5
      32     10   93.58%      1,394      3,765        717.6
      32     20   93.62%      1,288      3,388        776.3
      64      1   78.20%        784      2,516      1,275.3
      64      2   92.95%        775      2,447      1,289.8
      64      4   95.93%        780      2,525      1,282.8
      64     10   96.51%        723      1,940      1,382.6
      64     20   96.59%        703      2,345      1,423.0
     128      1   78.97%        391      1,566      2,557.7
     128      2   94.29%        401      2,126      2,493.3
     128      4   97.55%        399      1,728      2,503.6
     128     10   98.21%        390      1,494      2,567.2
     128     20   98.29%        377      1,442      2,649.3
     256      1   79.22%        202        966      4,953.7
     256      2   94.81%        204        946      4,904.9
     256      4   98.24%        201        958      4,963.8
     256     10   98.89%        198        925      5,044.1
     256     20   98.99%        193        906      5,175.0

>>> IVF4096+RaBitQfs+Refine(SQ8) <<<
构建: 205.3s

  nprobe  k_fac    R@10     1T-QPS     MT-QPS    1T-lat(us)
------------------------------------------------------------
      32      1   76.75%      1,884      5,868        530.8
      32      2   88.88%      1,847      5,277        541.6
      32      4   90.48%      1,736      5,285        575.9
      32     10   90.71%      1,710      4,977        584.9
      32     20   90.72%      1,517      5,569        659.3
      64      1   78.70%        986      3,351      1,014.5
      64      2   92.06%        976      3,257      1,024.6
      64      4   94.04%        963      3,335      1,038.5
      64     10   94.45%        929      3,074      1,076.7
      64     20   94.48%        882      3,047      1,133.2
     128      1   79.73%        506      1,859      1,978.0
     128      2   93.98%        499      1,917      2,005.3
     128      4   96.26%        500      1,855      1,999.6
     128     10   96.75%        489      1,849      2,046.2
     128     20   96.82%        476      1,962      2,101.9
     256      1   80.24%        257      1,078      3,887.8
     256      2   95.03%        256      1,044      3,905.8
     256      4   97.60%        257      1,034      3,895.7
     256     10   98.16%        253      1,061      3,954.4
     256     20   98.25%        249        941      4,021.4
     512      1   80.53%        122        515      8,223.7
     512      2   95.49%        124        515      8,054.8
     512      4   98.18%        129        538      7,763.8
     512     10   98.77%        128        524      7,828.7
     512     20   98.87%        128        527      7,803.3


----------------------------------------------------------------------
  模式 C: IVF1024+RaBitQfs+Refine(Flat) — FP32 精排
----------------------------------------------------------------------

构建: 23.0s

  nprobe  k_fac    R@10     1T-QPS     MT-QPS    1T-lat(us)
------------------------------------------------------------
      16      1   74.38%      2,829      8,002        353.5
      16      2   86.66%      2,695      6,724        371.1
      16      4   88.83%      2,702      7,377        370.1
      16     10   89.17%      2,378      4,604        420.5
      32      1   76.79%      1,441      3,797        693.8
      32      2   91.03%      1,519      4,342        658.5
      32      4   93.68%      1,483      4,103        674.3
      32     10   94.19%      1,407      4,048        710.5
      64      1   78.20%        794      2,822      1,258.7
      64      2   93.40%        792      2,536      1,263.4
      64      4   96.56%        767      2,942      1,304.5
      64     10   97.21%        756      2,468      1,322.9
     128      1   78.97%        401      1,505      2,492.1
     128      2   94.78%        402      1,619      2,490.5
     128      4   98.23%        397      1,367      2,516.6
     128     10   98.99%        393      1,489      2,545.2


----------------------------------------------------------------------
  RaBitQ Pareto 最优点 & 对比 QuIVer
----------------------------------------------------------------------

RaBitQ 最强配置: IVF1024+RaBitQfs+Refine(SQ8)
Pareto 最优点 (每个 recall 级别取最高 MT-QPS):

  Recall  配置                         1T-QPS   MT-QPS
  ~93%    nprobe=32, k_fac=4             1,485    4,184
  ~95%    nprobe=64, k_fac=4               780    2,525
  ~97%    nprobe=128, k_fac=4              399    1,728
  ~98%    nprobe=128, k_fac=10             390    1,494
  ~99%    nprobe=256, k_fac=20             193      906

对比 QuIVer (同硬件):
  Recall  QuIVer-MT   RaBitQfs+Ref   QuIVer/RaBitQ
  ~95%     36,729       2,525          14.5x
  ~97%     21,021       1,728          12.2x
  ~99%     11,837         906          13.1x

注: RaBitQ 虽然有理论最优误差界 (O(1/sqrt(D)))，
    但其 IVF 架构（粗搜索+精排）无法匹配 QuIVer 的
    图导航架构（BQ-native graph + FP rerank）的吞吐。
    核心差异:
    - QuIVer: 图导航全程 RAM (BQ Hamming distance)，仅 top-K 精排
    - RaBitQ+IVF: 需要搜索大量 cluster + 精排大量候选
    - 在高 recall (>95%) 下，RaBitQ 需要增大 nprobe 和 k_factor，
      导致 QPS 急剧下降



PS C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB> python scripts/run_noavx512_bench.py
======================================================================
  QuIVer 无 AVX-512 消融实验
  数据集: Cohere-1M (768-d)
  SIMD 路径: AVX2 (AVX-512 已禁用)
  日志: C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB\research\bench_logs\cohere-1m-noavx512.log
  开始时间: 2026-05-12 22:09:26
======================================================================
   Compiling triviumdb v0.7.0 (C:\Users\Administrator\OneDrive\桌面\workspace\TriviumDB)
    Finished `bench` profile [optimized] target(s) in 20.06s
     Running benches\bench_sensitivity.rs (target\release\deps\bench_sensitivity-d2b34d7a5832ef72.exe)
加载 cohere-1m 数据集...
加载完成! 1000000 x 768 训练集, 1000 测试集, 耗时 2.34s

======================================================================
实验 1: 参数敏感性分析
数据集: cohere-1m (1000000 x 768)
======================================================================

计算 brute-force 基准 QPS (1000q)...
Brute-force 基准: QPS=18.4, latency=54.35ms/q

--- 1d: ef_search 精细 Recall-QPS 曲线 (m=32, ef_c=128, α=1.2) ---
ef          R@10(%)       MT-QPS       1T-QPS    lat(ms)      vs BF
------------------------------------------------------------------
8            44.92%       120195        11374      0.09    6532.5x
12           62.19%        91120         8179      0.12    4952.3x
16           72.26%        77771         7040      0.14    4226.8x
20           78.17%        70233         6133      0.16    3817.1x
24           82.30%        63338         5731      0.17    3442.4x
32           86.76%        56698         4894      0.20    3081.5x
40           90.36%        45536         4260      0.23    2474.8x
48           92.57%        40805         3807      0.26    2217.7x
56           93.67%        36766         3315      0.30    1998.2x
64           94.57%        34756         2978      0.34    1889.0x
80           95.66%        27235         2561      0.39    1480.2x
96           96.45%        24557         2244      0.45    1334.6x
112          96.86%        20011         1998      0.50    1087.6x
128          97.27%        19533         1824      0.55    1061.6x
160          97.76%        15698         1481      0.68     853.1x
192          98.13%        14021         1300      0.77     762.0x
224          98.37%        12117         1127      0.89     658.6x
256          98.68%        10880          979      1.02     591.3x
320          98.95%         8892          829      1.21     483.3x
384          99.08%         7537          714      1.40     409.7x
448          99.24%         6511          621      1.61     353.9x
512          99.31%         5819          541      1.85     316.3x
640          99.50%         4732          450      2.22     257.2x
768          99.57%         4018          381      2.62     218.4x
896          99.61%         3486          334      3.00     189.5x
1024         99.62%         3056          288      3.47     166.1x

--- 1e: 不同 m 的 Pareto 曲线 (ef_c=128, α=1.2) ---
m      ef          R@10(%)       MT-QPS      vs BF
--------------------------------------------------
8      32           48.82%       118413    6435.7x
8      64           65.69%        69756    3791.2x
8      128          77.88%        37807    2054.8x
8      256          84.42%        21127    1148.2x
8      512          88.31%        11238     610.8x
8      1024         91.09%         5662     307.7x

16     32           75.35%        80648    4383.2x
16     64           87.47%        51795    2815.0x
16     128          92.81%        28093    1526.8x
16     256          95.72%        14862     807.7x
16     512          97.52%         8141     442.5x
16     1024         98.29%         4224     229.6x

32     32           86.76%        55519    3017.4x
32     64           94.67%        33477    1819.5x
32     128          97.33%        18814    1022.6x
32     256          98.72%        10275     558.4x
32     512          99.33%         5460     296.7x
32     1024         99.62%         2912     158.2x

48     32           90.25%        42301    2299.0x
48     64           95.74%        24574    1335.6x
48     128          98.15%        14285     776.4x
48     256          99.20%         7764     421.9x
48     512          99.60%         4135     224.8x
48     1024         99.84%         2303     125.1x

64     32           91.45%        31903    1733.9x
64     64           96.39%        20253    1100.7x
64     128          98.50%        11829     642.9x
64     256          99.41%         6541     355.5x
64     512          99.74%         3550     192.9x
64     1024         99.92%         1977     107.4x


======================================================================
实验 2: 单线程 vs 多线程查询扩展性
数据集: cohere-1m (1000000 x 768)
======================================================================
索引构建: 108.3s (9230 vecs/s), Hot 675 MB

--- 2a: 不同 ef 下的 单线程 vs 多线程 QPS ---
ef             1T-QPS       MT-QPS      R@10(%)        加速比
----------------------------------------------------------
64               2814        34624       94.68%      12.3x
128              1780        19520       97.32%      11.0x
256              1005        11014       98.71%      11.0x
512               544         5893       99.31%      10.8x
1024              297         3076       99.62%      10.3x

--- 2b: 线程数扩展性 (ef=64) ---
线程数                 QPS        相对1线程
--------------------------------------
1                  3081         1.0x
2                  5927         1.9x
4                 11947         3.9x
8                 23165         7.5x
16                34843        11.3x

全部实验完成!

  ✅ 无 AVX-512 实验完成! 耗时 990s (16.5min)