The previous implementation was O(N²lg(N)) due to sorting the full
runnables priority list once per runnable being scheduled. While not
confirmed, this is suspected to cause performance issues and
bottlenecking with the queue runner when the runnable list gets large
enough.
This commit changes the dispatcher to instead only sort runnables per
priority once per dispatch cycle. This has the drawback of being less
reactive to runnable priority changes: the previous code would react
immediately, while this might end up using "old" priorities until the
next dispatch cycle. However, dispatch cycles are not supposed to take
very long (seconds, not minutes/hours), so this is not expected to have
much or any practical impact.
Ideally runnables would be maintained in a sorted data structure instead
of the current approach of copying + sorting in the scheduler. This
would however be a much more invasive change to implement, and might
have to wait until we can confirm where the queue runner bottlenecks
actually lie.
Pierre Bourdon
b7c864c515