This is something of a followup to the compositor is evil, but more of a guide on how to optimize latency, and less of a rant. I’m starting to collect performance measurements of piet-gpu running on Android, and ran into unexpectedly poor frame rates, which re-opened this topic. Simply increasing the number of images in the swapchain fixed the throughput problem, but I was also worried about the impact on latency. Exploring that led me down a deep rabbit-hole.