I decided to add a second summary graph to the profiler; this one (shown at the top, in purple) gives a view into the time taken on the audio thread - the critical one for good sound processing. It is shown as continuous here, but really it is discrete; it is somewhat easier to visualise like this I think. The graph goes more red the closer the audio thread comes to the max time slice it is given to generate audio. The image here shows a brief bottleneck which I've zoomed into in the bottom of the profiler. You can see that the yellow 'Audio' region is much bigger for this particular frame, compared to the next ones. Although I don't have instrumentation for it yet, I believe this is coinciding with the arrival of the first notes into the graph - since a known problem is that some memory allocation is done at this point on the first run (and this particular stress test is playing many notes all at once). Just something I know I need to get around to fixing. I'm satisfied the profiler is useful enough to leave alone for now, and move onto other things.