Current compilation delay strategy for lukewarm methods is based on a fixed threshold coefficient. It probably should also be make dependent on the compiler load. However, I now think that the threshold coefficient should also dynamically decrease with the load. Such that when the startup/warmup end (and there is no load on compilers) we should be free to get back to compile lukewarm methods.

So, the coefficient should rather be something like: scale = (1 + DelayCoefficient * load_average).
We need to port the load average compilation code from the leyden premain repo.