effect: off (67b5048d8b51) -> on (97fa8425f080)  errors=[]
| prefill | conc | TTFT ms A | TTFT ms B | dTTFT | verdict | prefill tok/s A | prefill tok/s B | dprefill | verdict | groupings A / B |
|---:|---:|---:|---:|---:|---|---:|---:|---:|---|---|
| 16384 | 4 | 566.5±1.7 | 541.3±0.6 | -1.37% | NOISE | 4599 [4355-4900] | 4519 [4105-4902] | +1.28% | NOISE | 1x2,1x2,1x2,1x2 2x3,1x3,1x3,1x2 2x3,1x3,1x3,1x2 / 2x3,1x3,1x3,1x1 2x3,1x3,1x3,1x2 1x2,1x2,1x2,1x2 |
| 65536 | 4 | 548.3±0.0 | 555.2±1.2 | +1.26% | NOISE | 4365 [4225-4471] | 4360 [4294-4396] | -0.90% | NOISE | 2x3,1x3,1x3,1x1 2x3,1x3,1x3,1x2 1x2,1x2,1x2,1x2 / 1x2,1x2,1x2,1x2 1x2,1x2,1x2,1x2 1x2,1x2,1x2,1x2 |

control: on (97fa8425f080) -> null (d23b3c07579c)  errors=[]
| prefill | conc | TTFT ms A | TTFT ms B | dTTFT | verdict | prefill tok/s A | prefill tok/s B | dprefill | verdict | groupings A / B |
|---:|---:|---:|---:|---:|---|---:|---:|---:|---|---|
| 16384 | 4 | 515.3±0.9 | 519.4±2.6 | +2.06% | NOISE | 4950 [4874-5022] | 4730 [4450-4875] | -3.00% | NOISE | 1x2,1x2,1x2,1x2 2x3,1x3,1x3,1x2 2x3,1x3,1x3,1x2 / 1x2,1x2,1x2,1x2 2x3,1x3,1x3,1x2 1x2,1x2,1x2,1x2 |
| 65536 | 4 | 564.4±0.2 | 566.1±0.8 | +1.22% | LOSS | 4344 [4159-4458] | 4377 [4315-4444] | -1.02% | LOSS | 2x3,1x3,1x3,1x1 2x3,1x3,1x3,1x2 1x2,1x2,1x2,1x2 / 1x2,1x2,1x2,1x2 1x2,1x2,1x2,1x2 2x3,1x3,1x3,1x2 |

Effect beside the null control; an effect inside the control is a null,
whatever the per-run verdict says (that verdict's spread is one boot's):
  prefill  16384 conc 4:  ttft_ms effect -1.37% control +2.06%  tpot_ms effect +0.08% control -2.43%  tps_prefill effect +1.28% control -3.00%
  prefill  65536 conc 4:  ttft_ms effect +1.26% control +1.22%  tpot_ms effect +1.88% control -4.70%  tps_prefill effect -0.90% control -1.02%
streams: /home/arbi-admin/roh/runs/turn_c4_oh/effect.sse /home/arbi-admin/roh/runs/turn_c4_oh/control.sse  (stack left up; 'scripts/bench/console_build_ab.sh down' tears it down)
