Root cause of intermittent TTFT anomalies

Determine the root cause of the intermittent approximately 15-second TTFT blowups observed in the serving stack during the experiments.

Background

A small fraction of repetitions exhibited intermittent TTFT blowups exceeding approximately 15 seconds. These anomalies were treated as a serving-stack artifact and dropped pairwise because they were arm-independent, ensuring identical repetition sets across compared methods. However, the underlying cause was not resolved, leaving uncertainty about the source of this latency behavior and its implications for deployment robustness.

References

The intermittent TTFT anomalies in our baseline (\textasciitilde15\% of repetitions) are a serving-stack artifact whose root cause remains open.

— Phase-Decoupled, Model-Calibrated Power Control for Disaggregated LLM Serving  (2609.11133 - Kim et al., 10 Sep 2026) in Section 6, Limitations