Why it matters
Harshul Jain and Parv Khatri describe admission control before requests consume GPU cache, using request context alongside live engine pressure. Their benchmark method replays agent bursts, retries and sessions with and without a policy. It counts rejected arrivals in the success denominator and checks whether the chosen signal actually changes under load. A small-model demonstration found queue pressure while cache occupancy stayed low; its results are not general capacity estimates.
My takeaway: Replay traffic above measured capacity and report successful requests against all arrivals. Fail a benchmark when required telemetry is missing, compare queue and cache signals, and disclose model, hardware, workload and SLOs. Bound retries with backoff and jitter, and verify cancellation releases server resources.