Timestamp-exact replay for burst-sensitive routing evaluation

Develop timestamp-exact replay that preserves each key’s original timestamp sequence, in order to evaluate CacheRoute under realistic burstiness beyond aggregate inter-arrival-gap block bootstrapping.

Background

The burst-sensitivity analysis evaluates CacheRoute using marginal-CV-matched Gamma arrivals and a moving-block bootstrap of measured inter-arrival gaps. The bootstrap preserves short-range gap correlation but does not preserve the original global ordering or each key’s exact timestamp sequence.

The paper explicitly defers timestamp-exact replay, leaving unresolved whether preserving the complete per-key temporal structure would materially change the reported routing and capacity results.

References

The bootstrap draws contiguous blocks of 50 from 22,639 measured inter-arrival gaps. It retains short-range gap correlation, but neither the original global order nor each key's exact timestamp sequence. We leave timestamp-exact replay to future work.

CacheRoute: Planned Prefix-Affinity Routing for Large-Scale LLM Serving  (2608.19677 - Cheng, 20 Aug 2026) in Appendix, Section "Arrival-Process Sensitivity"