Pareto explorer

A configuration is on the front when no other configuration is at least as good on both metrics and strictly better on one. Everything off the front is beaten by something on it, for these two measures; which point on the front to choose depends on which measure matters more. Pick the metrics, a workload and the devices; tighten the SLO filter to keep only configurations whose p99 latencies at the reference load fit within a multiple of the workload's SLOs, and watch the front re-form.

The points are the sweep's 350 configurations, one set per workload. With every device shown and no SLO filter, the front drawn here is exactly the sweep's own Pareto flag for that pair of metrics (a unit test checks all of them). Hover a point for its configuration; click it to re-run it in the live simulator.