I am not skipping turns

Fragment of a discussion from Talk:CPU constant
Jump to navigation Jump to search

Yes, I've seen something like this in my profiler. I quickly look at the source, I think the problem is that robocode does not use System.nanoTime to compare times. Instead it using Java to trow exception if wait time is exceeded. It is possible that java treat it as low priority and skips the exception.

I noticed that with similar CPU demand from my bot, I am skipping significantly more against simple/fast bots. Sometimes even 30 turns which causes infamous ThreadDeath. I.e. if i run against Diamond, I see less skipped turns than against a rammer or HawkOnFire. Why is it so I do not understand. But it might be connected to your case.

Beaming (talk)04:01, 22 October 2017

I strongly suspect that if your thread doesn't pause often enough for long enough it will be paused by the JVM for a longer period for housekeeping tasks (GC, updating to JIT'd codepaths etc). I managed to create a fix for the ThreadDeath / long pause, which was simply adding a 1ms pause every tick. Maybe doing something like this would help on your side as well?

Skilgannon (talk)12:06, 22 October 2017

I tried 1ms Thread.Sleep at the end of each turn. Interestingly enough, sometimes it takes up to 10 ms It was never promised that it will sleep as much as asked, but factor of 10 is just nuts.

I see no significant change in the number of skipped turns.

I still get Thread.Death sometimes against simple/fast bots.

I really wish robocode itself would start garbage collector or at least call tread.sleep. Recall our discussion where we observe that Thread.Death disappears when in the GUI we set the desired FPS just a notch below max. I think the only difference that robocode inserts sleep in the cycle if FPS is below max.

Beaming (talk)23:04, 22 October 2017