JUSTIN ZHANG - CARLETON UNIVERSITY, SWAN LAB
Thank you for the email. I wasn't able to use your feedback this time - a lot of my implementation was already done before yesterday's email landed.
We can discuss it - I've set aside a full segment for your email near the end.
Walking the orchestrator end to end - then how I'm sweeping the alpha & beta weights.
all.ymlEvery knob the sweep touches lives in one file - talking through the config while we're still in the code.
The worker pools run your workloads - and I can't abstract those away from the experiment.
Do you want me to design a cost model for “all” file types?
score = α·normPending + β·normWait7 CONFIGS ON THE RATIO DIAGONAL · α ∈ {0, 0.1, 0.25, 0.5, 0.75, 0.9, 1} · β = 1 − α · PURE WAIT → PURE PENDINGThe swept weights verifiably reached the scheduler - per-decision score fields differ combo to combo: α = 0 ⇒ score = normWait, α = 1 ⇒ score = normPending.
WORKLOAD: 135 TASKS (60 COMMA_REMOVAL · 45 BLEU_SCORE · 30 WORD_COUNT) · GATED UNIFORM ARRIVALS · 90 S WINDOW · 1 - 5 S SIMULATED PROCESSING · KEDA-SCALED WORKERS ON K3SThe fairness columns are flat across the entire sweep - these values hold at every α (running metrics are post-warm-up).
| METRIC | BLEU_SCORE | COMMA_REMOVAL | WORD_COUNT | ACROSS ALL α |
|---|---|---|---|---|
| JAIN (FINAL) | 0.695 | 0.644 | 0.754 | IDENTICAL AT EVERY α |
| JAIN (RUNNING, MIN) | 0.561 - 0.573 | 0.588 - 0.609 | 0.529 - 0.585 | NOISE BAND - NO TREND |
| MAX LAG (FINAL) | 22 | 32 | 13 | IDENTICAL AT EVERY α |
| MAX LAG (RUNNING) | 23 | 32 | 13 | PINNED AT EVERY α |
| WHO-WAITED SPREAD (ASSIGN P99) | MAX - MIN ACROSS TENANTS: 1.5 - 12.4 S · KEDA SPAWN / POD-START JITTER | NO α TREND | ||
Burst arrivals + multi-second processing - deep backlogs, the regime where the α:β ratio must matter.
Your feedback, point by point - everything we didn't get to apply this round.
Jain's Fairness Index fails when tasks are heterogeneous.Tenants? Workers? A collection of tasks?JUSTIN ZHANG - CARLETON UNIVERSITY, SWAN LAB