Files
agentic-kvc/analysis/mb5_pd_ablation/fig1_reuse_fixed.json
Gahow Wang 9c105cf05a MB5 PD ablation: controlled-variable reuse/conc redo + campaign tooling
Reuse and concurrency axes redone with proper controlled variables, plus
the orchestration used to run them on dash0:

- run_reuse_fixed.sh: hold REAL prefill work (delta) constant, vary only
  cached prefix -> reuse = C/(C+U). Supersedes old fig1 (which held
  input=8192 and sliced prefix out, confounding "more reuse" with "less
  prefill").
- run_conc.sh: agentic-corner config (in=32768, delta=512, reuse=0.984,
  out=128) that exposes PD's structural KV-transfer tax. Supersedes old fig3.
- run_campaign{,2,3}.sh, backfill_d2048o128.sh: serial campaign drivers
  (strictly one driver at a time), out=128 sweeps, PD wall-cap for
  collapse-draining high-reuse arms, and flaked-arm backfill.
- mb5_run_gpu.sh: per-config bring-up / replay / teardown orchestrator.
- plot_pd_crossover.py: render the reuse_compare figures from fig_agg dumps.
- fig_agg.py: tolerate null stats from fully-collapsed arms (0 successes
  write the stat keys as null; `dict.get(k, {})` returns null, not {}).

Data: fig1_reuse_fixed.json, fig1_reuse_d{1024,2048}_o128.json
Figs: reuse_compare_AB.png, reuse_compare_ABC.png

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-01 01:03:27 +08:00

2 lines
9.0 KiB
JSON

[{"name": "reuse_p18432_d2048_2P+6D_rep1", "arm": "2P+6D", "n": 384, "req": 384, "e2e_p50": 2.533020637521986, "e2e_p90": 5.85391353177838, "e2e_p99": 13.18415429443412, "e2e_mean": 3.502040037624132, "ttft_p90": 4.246971374610439, "tpot_p99": 0.007761271612152603, "tps": 490.8994507324649, "wall": 200.2528213330079, "pu": 51.829787234042556, "du": 48.47872340425532, "apc": 0.7875}, {"name": "reuse_p18432_d2048_4P+4D_rep1", "arm": "4P+4D", "n": 384, "req": 384, "e2e_p50": 2.327055490517523, "e2e_p90": 4.852692579611903, "e2e_p99": 12.918203760247179, "e2e_mean": 3.1448965869362837, "ttft_p90": 3.0742682195210365, "tpot_p99": 0.00903629884385851, "tps": 526.2705626745112, "wall": 186.79365135001717, "pu": 26.928977272727273, "du": 70.5028409090909, "apc": 0.7875}, {"name": "reuse_p18432_d2048_6P+2D_rep1", "arm": "6P+2D", "n": 384, "req": 384, "e2e_p50": 3.1278362965094857, "e2e_p90": 4.839138368546264, "e2e_p99": 7.14371296691534, "e2e_mean": 3.424949068415723, "ttft_p90": 2.517161613609642, "tpot_p99": 0.012816217731421942, "tps": 514.6500556526703, "wall": 191.01134629303124, "pu": 18.974074074074075, "du": 95.7611111111111, "apc": 0.7875}, {"name": "reuse_p18432_d2048_8C-proxy_rep1", "arm": "colo", "n": 384, "req": 384, "e2e_p50": 2.692243392491946, "e2e_p90": 5.5457493721274655, "e2e_p99": 8.494194235785404, "e2e_mean": 3.2005772834381787, "ttft_p90": 2.1673677822982427, "tpot_p99": 0.030382820091552715, "tps": 528.3289416764078, "wall": 186.06589994498063, "pu": 45.94602272727273, "du": null, "apc": 0.7875}, {"name": "reuse_p2048_d2048_2P+6D_rep1", "arm": "2P+6D", "n": 384, "req": 384, "e2e_p50": 1.5452155300008599, "e2e_p90": 1.7081896169576791, "e2e_p99": 2.45207904072362, "e2e_mean": 1.6013075451828627, "ttft_p90": 0.34778050871100274, "tpot_p99": 0.005669485672979671, "tps": 996.0242196356846, "wall": 98.69639519002521, "pu": 23.893617021276597, "du": 88.177304964539, "apc": 0.4375}, {"name": "reuse_p2048_d2048_4P+4D_rep1", "arm": "4P+4D", "n": 384, "req": 384, "e2e_p50": 1.6793917149771005, "e2e_p90": 1.818927331600571, "e2e_p99": 2.323814849504852, "e2e_mean": 1.7090580133658477, "ttft_p90": 0.31593548320815895, "tpot_p99": 0.006266038190100945, "tps": 946.6728525267378, "wall": 103.84157498297282, "pu": 22.26530612244898, "du": 97.3265306122449, "apc": 0.4375}, {"name": "reuse_p2048_d2048_6P+2D_rep1", "arm": "6P+2D", "n": 384, "req": 384, "e2e_p50": 2.156280273484299, "e2e_p90": 2.297612584318267, "e2e_p99": 2.835462462223368, "e2e_mean": 2.157636388411902, "ttft_p90": 0.33686318337568083, "tpot_p99": 0.008273487242621223, "tps": 783.5302862115615, "wall": 125.46292304195231, "pu": 8.175, "du": 96.54166666666667, "apc": 0.4375}, {"name": "reuse_p2048_d2048_8C-proxy_rep1", "arm": "colo", "n": 384, "req": 384, "e2e_p50": 1.8087456520006526, "e2e_p90": 2.3159401921962854, "e2e_p99": 3.069265903531341, "e2e_mean": 1.8344004751229477, "ttft_p90": 0.30347971800365486, "tpot_p99": 0.010619336528881619, "tps": 837.246318714222, "wall": 117.41347534494707, "pu": 44.792410714285715, "du": null, "apc": 0.4375}, {"name": "reuse_p38912_d2048_2P+6D_rep1", "arm": "2P+6D", "n": 361, "req": 384, "e2e_p50": 6.57530459103873, "e2e_p90": 47.03490046202205, "e2e_p99": 448.50315178359153, "e2e_mean": 29.01171956352752, "ttft_p90": 45.17827109998325, "tpot_p99": 0.009629647705758752, "tps": 51.92238624874325, "wall": 1779.88737954502, "pu": 27.959832134292565, "du": 6.306754596322942, "apc": 0.8698754011405736}, {"name": "reuse_p38912_d2048_4P+4D_rep1", "arm": "4P+4D", "n": 327, "req": 384, "e2e_p50": 3.1876551050227135, "e2e_p90": 11.449897833599248, "e2e_p99": 29.874143838356257, "e2e_mean": 5.742165601527495, "ttft_p90": 9.523732434178246, "tpot_p99": 0.0121669072732512, "tps": 61.0543739943618, "wall": 1371.1056968290359, "pu": 9.472395023328149, "du": 10.68040435458787, "apc": 0.3990417067856459}, {"name": "reuse_p38912_d2048_6P+2D_rep1", "arm": "6P+2D", "n": 384, "req": 384, "e2e_p50": 4.693941325997002, "e2e_p90": 11.446271817904197, "e2e_p99": 25.98846597966217, "e2e_mean": 6.425895827477082, "ttft_p90": 7.717594898125389, "tpot_p99": 0.024569422919093177, "tps": 291.58829267237684, "wall": 337.1328769720276, "pu": 26.540880503144653, "du": 79.52830188679245, "apc": 0.83125}, {"name": "reuse_p38912_d2048_8C-proxy_rep1", "arm": "colo", "n": 384, "req": 384, "e2e_p50": 3.804246386018349, "e2e_p90": 9.94819127159426, "e2e_p99": 20.41812537532541, "e2e_mean": 5.546197232348732, "ttft_p90": 6.640114219562383, "tpot_p99": 0.06730832396853975, "tps": 322.8433493800264, "wall": 304.4944248929969, "pu": 45.20192307692308, "du": null, "apc": 0.83125}, {"name": "reuse_p4096_d2048_2P+6D_rep1", "arm": "2P+6D", "n": 384, "req": 384, "e2e_p50": 1.6006885370006785, "e2e_p90": 1.8821295853122137, "e2e_p99": 3.0432000186893857, "e2e_mean": 1.6879496743948341, "ttft_p90": 0.5026261055085341, "tpot_p99": 0.005706233633389954, "tps": 950.5167565946713, "wall": 103.42163809103658, "pu": 48.04081632653061, "du": 80.17006802721089, "apc": 0.5833333333333334}, {"name": "reuse_p4096_d2048_4P+4D_rep1", "arm": "4P+4D", "n": 384, "req": 384, "e2e_p50": 1.7440840654890053, "e2e_p90": 1.985663170699263, "e2e_p99": 2.674550873187256, "e2e_mean": 1.7983998170934683, "ttft_p90": 0.4889643796952441, "tpot_p99": 0.00646362871355286, "tps": 906.9274927042811, "wall": 108.3923475589836, "pu": 15.769607843137255, "du": 93.76470588235294, "apc": 0.5833333333333334}, {"name": "reuse_p4096_d2048_6P+2D_rep1", "arm": "6P+2D", "n": 384, "req": 384, "e2e_p50": 2.2896783250034787, "e2e_p90": 2.580035507993307, "e2e_p99": 3.2560947993275495, "e2e_mean": 2.329391661541498, "ttft_p90": 0.5062051271961536, "tpot_p99": 0.009080472360109436, "tps": 731.5427124539008, "wall": 134.37902985903202, "pu": 9.973958333333334, "du": 97.59375, "apc": 0.5833333333333334}, {"name": "reuse_p4096_d2048_8C-proxy_rep1", "arm": "colo", "n": 384, "req": 384, "e2e_p50": 1.904530945525039, "e2e_p90": 2.7650697994977236, "e2e_p99": 3.504905357128009, "e2e_mean": 1.9890683890278826, "ttft_p90": 0.41231827571755275, "tpot_p99": 0.012474335152598777, "tps": 782.9542862351619, "wall": 125.55522299098084, "pu": 46.62076271186441, "du": null, "apc": 0.5833333333333334}, {"name": "reuse_p512_d2048_2P+6D_rep1", "arm": "2P+6D", "n": 384, "req": 384, "e2e_p50": 1.5259871140005998, "e2e_p90": 1.6426021346589548, "e2e_p99": 2.4021258947759634, "e2e_mean": 1.565941817816262, "ttft_p90": 0.2866207461862359, "tpot_p99": 0.005887627696034079, "tps": 1007.858853784255, "wall": 97.53746730595594, "pu": 19.58695652173913, "du": 83.34057971014492, "apc": 0.175}, {"name": "reuse_p512_d2048_4P+4D_rep1", "arm": "4P+4D", "n": 384, "req": 384, "e2e_p50": 1.6439885564905126, "e2e_p90": 1.7445397889125163, "e2e_p99": 2.1885562700900483, "e2e_mean": 1.665460689239353, "ttft_p90": 0.21324421370518393, "tpot_p99": 0.006212426621711119, "tps": 966.5616230201254, "wall": 101.7048449459835, "pu": 8.515625, "du": 97.02083333333333, "apc": 0.175}, {"name": "reuse_p512_d2048_6P+2D_rep1", "arm": "6P+2D", "n": 384, "req": 384, "e2e_p50": 2.059083105996251, "e2e_p90": 2.20597213053843, "e2e_p99": 2.5390614521381236, "e2e_mean": 2.0654355626335623, "ttft_p90": 0.23428239471977577, "tpot_p99": 0.008075011566733285, "tps": 810.9640468666952, "wall": 121.21869074198185, "pu": 6.916666666666667, "du": 96.86206896551724, "apc": 0.175}, {"name": "reuse_p512_d2048_8C-proxy_rep1", "arm": "colo", "n": 384, "req": 384, "e2e_p50": 1.7308740815205965, "e2e_p90": 2.19734079890186, "e2e_p99": 2.574364611047204, "e2e_mean": 1.753096249226625, "ttft_p90": 0.2771356365934479, "tpot_p99": 0.009422459370876646, "tps": 875.0118823136878, "wall": 112.34590293798828, "pu": 46.7122641509434, "du": null, "apc": 0.175}, {"name": "reuse_p8192_d2048_2P+6D_rep1", "arm": "2P+6D", "n": 384, "req": 384, "e2e_p50": 1.7981248289870564, "e2e_p90": 2.62749446813832, "e2e_p99": 4.4945996468386165, "e2e_mean": 2.0330339767545715, "ttft_p90": 1.1596759190259036, "tpot_p99": 0.00635605895942004, "tps": 807.5382123177835, "wall": 121.73293907399056, "pu": 39.69298245614035, "du": 75.94152046783626, "apc": 0.7}, {"name": "reuse_p8192_d2048_4P+4D_rep1", "arm": "4P+4D", "n": 384, "req": 384, "e2e_p50": 1.8958783995185513, "e2e_p90": 2.529831808601739, "e2e_p99": 3.6714393736550135, "e2e_mean": 2.0412568975771137, "ttft_p90": 0.941303398378659, "tpot_p99": 0.0067997743117020415, "tps": 812.6337799219408, "wall": 120.9696205459768, "pu": 23.42543859649123, "du": 91.20614035087719, "apc": 0.7}, {"name": "reuse_p8192_d2048_6P+2D_rep1", "arm": "6P+2D", "n": 384, "req": 384, "e2e_p50": 2.4744975224894006, "e2e_p90": 3.0726254384964706, "e2e_p99": 4.066428293744687, "e2e_mean": 2.5728964008159587, "ttft_p90": 0.9646674327901564, "tpot_p99": 0.009585995291005914, "tps": 667.4391764241245, "wall": 147.28533096704632, "pu": 10.345238095238095, "du": 95.71428571428571, "apc": 0.7}, {"name": "reuse_p8192_d2048_8C-proxy_rep1", "arm": "colo", "n": 384, "req": 384, "e2e_p50": 2.1906960625201464, "e2e_p90": 3.31940173432813, "e2e_p99": 4.419624262939906, "e2e_mean": 2.272983053012316, "ttft_p90": 0.798827746004099, "tpot_p99": 0.014866376486218845, "tps": 697.1272118347367, "wall": 141.01300068502314, "pu": 48.098880597014926, "du": null, "apc": 0.7}]