> or if they happen they might only converge to an stable state That one will always be dependent on the usage patterns. So the auto-tuner can't guarantee it.

> or if they happen they might only converge to an stable state That one will always be dependent on the usage patterns. So the auto-tuner can't guarantee it. Also, I imagine the risk of the feedback turning positive is related to the system load (not CPU, but the usage of the resources you are optimizing). If so, it will make your computer less able to manage load. But this can still be useful for optimizing for latency.