More precise spinning with back off idle strategy

[plokhotnyuk]

• allows usage with zero for maxSpins and maxYields properties without redundant ThreadHints.onSpinWait() and Thread.yield() calls
• fixes missing call of ThreadHints.onSpinWait() for 1st idle() call
• fixes redundant empty spin without Thread.yield() after switching to PARKING state

[plokhotnyuk]

@mikeb01 @wojciech-adaptive @vyazelenko Could you please review and merge this PR with squashed commits?

[vyazelenko]

@plokhotnyuk What is the motivation for this change? Have you considered writing a custom idle strategy instead?

[plokhotnyuk]

@plokhotnyuk What is the motivation for this change? Have you considered writing a custom idle strategy instead?

The motivation is simple: it makes the strategy easier to configure, saves CPU cycles by avoiding redundant loops and reducing contention with other threads that allows to reduce latency and increase total system throughput.

Yes, I'm using the proposed version in our classpath but it could be beneficial for all users. Especially it matters a lot for cases when there are dozens of services with several agents that have different latency requirements.

[vyazelenko]

/cc @mjpt777 @mikeb01

[mikeb01]

The intention of the back off idle strategy is that it will always going through a back off process regardless of the configuration, hence the reason that it will always do at least one spin and one yield before parking. We want to retain that behaviour.

Given that IdleStrategy is an interface, you can copy the existing strategy and build a new idle strategy that has the behaviour you desire. If you wanted to submit that new IdleStrategy implementation as a PR, we would consider merging that.