Measure shared-tier contention across multiple prefill engines

Measure the contention and performance effects that arise when several prefill engines concurrently access shared Valkey remote-tier instances, including the KV handoff to decode, rather than extrapolating from a single-engine prefill-only replay.

Background

The evaluation replays production traffic on one isolated PCP8 prefill engine and uses a private Valkey instance. It therefore does not directly establish how multiple prefill engines will contend for shared remote-tier capacity or how the remote KV handoff to decode will affect the reported behavior. The paper estimates that six engines would remain below the measured bandwidth, but explicitly identifies this multi-engine scenario as unmeasured.

References

Prefill only, one engine. The replays omit the KV handoff to decode and use one engine as the only Valkey client; contention from several engines on shared instances is estimated, not measured.

— When the KV Cache Is Slower Than the GPU: Scaling Remote KV Tiers for Context-Parallel Prefill  (2610.21500 - Ibrahim, 15 Oct 2026) in Section 6, Limitations, bullet “Prefill only, one engine”