It's been difficult to understand the effects of changes to GPU Scheduling using our current user facing metrics like smoothness.  Perhaps it would be worth adding some lower level metrics so that we can understand the performance of scheduling, and how small or large changes affect it.

A core metric I am thinking of is queuing delay from when a high priority (i.e. Display Compositor) stream receives work and becomes schedule-able, until it becomes scheduled.  There may be other interesting metrics such as pre-emption frequency.  For cases like WebGL pre-emption can be detrimental to performance, so less is better.