FIFO · Jain index
–
Fair DRR · Jain index
–
FIFO · p99 e2e (s)
–
Fair DRR · p99 e2e (s)
–
FIFO · J/token
–
Fair DRR · J/token
–
Press Run.
TTFT scatter — static FIFO (blue) vs fair DRR (green)
x = request completion order, y = TTFT (s). Green fair cluster should sit lower & tighter under load; power cap shifts both up but saves J/token.
Per-user completed tokens
| user | FIFO tok | Fair tok |
|---|
Model: prefill 2 + 0.15·prompt ms · decode step 3 + 0.8·batch ms (×1.25 under cap) · max batch 8 · T=20 s window.