Simulate many users asking the fleet for generations at the same time, and measure how it copes: waiting time, latency, throughput, bandwidth, cold model loads, per-machine load and fairness. The run opens on its own page with one panel per user.
What this is, honestly. Measured: how long the router takes to answer, bytes on the wire, errors. Modelled: queue wait, execution and utilisation, played on a virtual clock from the router's own exec / cold-load estimates. No GPU job is submitted and no real browser is driven: each user is a scripted client.