Direct validation of the runtime SLO guard

Establish the runtime behavior, recovery effectiveness, and compliance cost of the ITL-p99 SLO guard in the phase-decoupled power controller by directly testing cases in which the guard releases a lane setpoint after an SLO violation.

Background

The controller includes a runtime guard that monitors per-lane tail latency and releases a lane’s power cap or clock-window setpoint toward the hardware default when the ITL-p99 SLO is violated. All reported steady-state experiments armed the guard, but the guard never tripped in any recorded run. Consequently, the paper’s compliance results validate the latency-gated calibration procedure rather than observed runtime intervention, leaving direct fault-injection and recovery testing unresolved.

References

We are explicit about what the data does not isolate: the runtime guard never had cause to fire in any recorded run, so the compliance contrast is attributable to latency-gated calibration rather than to runtime intervention, and a direct test of the guard remains outstanding.

— Phase-Decoupled, Model-Calibrated Power Control for Disaggregated LLM Serving  (2609.11133 - Kim et al., 10 Sep 2026) in Introduction; Section 5.1; Section 6 / Future Work