ThreadDeath problem

Would you try modifing the source of robocode to make it ouput the result of checkSkippedTurn? I think the bug may be whether the skipped turns are counted incorrectly or some skipped turns are not logged to console.

Skipped turns are logged (and SkippedTurnEvents happens) only when the Robocode regains control over the execution of the thread. Since there is an unhandled exception being thrown right during the skipped turn checking, it makes sense that the skipped turns are not logged. I'm pretty sure you are skipping too many turns, Thread.stop() is being called as it is supposed to be (by looking at Robocode source), the exception is being thrown (because Thread.stop() is deprecated) and for some reason it's not being handled.

So, the error message is pretty much "you have skipped too many turns, I tried to stop your thread but for this I called a deprecated function which throws an error". The bot punishment for skipping too many turns is correct. The correct questions are: this should really happen? not the error, but the "too many skipped turns" thing. Does it happen when locking to 1000 FPS? If it was a "fair" turn skipping, it should happen while locked to 1000 FPS as well? If yes, why it's behaving differently when on full speed? Can others reproduce the same issue? (pick Roborio vs Neuromancer, for example, and go full speed). Why it's more likely to happen when a slow bot is against a fast bot? And so on.

Does you guys have any other interesting observations on that?

Ok. Roborio also fall into ThreadDeath in the battle:

rsalesc.roborio.Roborio 1.2.7  vs penguin.Joker .611wr

It does not happens at 1000 FPS, only if I push the slider to max.

I also see DrussGT falling into ThreadDeath in the battle:

jk.mega.DrussGT 3.2.1 vs penguin.Joker .611wr

It seems that in a pair of bots the most CPU demanding gets the exception.

Yeah, here is the heart of the issue in my opinion. Weird things going under the hood when going full speed. I still have not fully understood Robocode skipped turn source so I'm not sure what's happening. I'll do more tests tomorrow.

I looked at the internals. The difference between desired 1000 TPS and max is that in the later case the Thread.sleep is called (actually if you have more than 0.5 ms to spare). I am guessing that probabilistically Thread.Sleep is called couple times per round.

I do not know what java does when thread is asleep, but I would imagine that the garbage collector and other gizmos kick in during the sleep. While at max TPS the java chooses as it wishes when to run GC and if you are the unlucky one, the GC time will be counted against your bot. This would explain sporadic super long executions in my bot time profiling.

I would gladly try. But how do I call it within the AdvancedBot?

What I mean is that you hack in to the robocode to make it output some useful information about these bugs. There is no need for your bot to call that, as checkSkippedTurn is called by the game every turn.

I must be very distracted, the skipped turns should be logged, right?

They are logged before the piece of code which calls the thread stopping procedure. It's logged if the bot is still alive. That means that the bots which are getting the ThreadError are not alive at all?

Yes, that's very strange. And that's why I think hack into robocode source code and make it output some useful information in checkSkippedTurn would help.

This would help us find out 1. why skipped turns are getting greater than 30 eventually. 2. why there is no output before at all (is it still alive? )

I'm also thinking about the possibility that the main thread of robocode is also skipping turns itself ;p

I cannot remember how I get this knowledge but I recall that the logic is following. Robocode let you consume as much time as you want per tic (see the exception below), than if you exceed time quota, it will force you to skip as many turns as you excess time over permited time per tic. That is why you often see skipped turns in succession (though, not in my log above).

However, if a bot have a glitch and never returns control back to the robocode (let's say there is an infinite loop), then robocode have a way to force stop such bot. I would not be surprised that it first kill the bot and then executes checkSkippedTurn. In this case a bot will see nothing in console.

For your questions:

1. I/O robots (bots that read/write from the data directory) get 240 skipped turns. Apparently a robot can get marked as an I/O robot just by calling getDataDirectory().
2. This is where things stop making sense. The stacktrace shows RobotPeer.checkSkippedTurn() calling BasicRobotProxy.forceStopThread(), but the code shows that it can't happen without going through println("SYSTEM: ..."), there literally is no other code path. Are println() calls getting silently dropped?
3. The Robocode main thread can't "skip turns" because it controls the turn counter.

Relevant code:

public void checkSkippedTurn() {
    // Store last and current execution time for detecting skipped turns
    lastExecutionTime = currentExecutionTime;
    currentExecutionTime = battle.getTime();

    int numSkippedTurns = (currentExecutionTime - lastExecutionTime) - 1;
    if (numSkippedTurns >= 1) {
        events.get().clear(false);

        if (isAlive()) {
            for (int skippedTurn = lastExecutionTime + 1; skippedTurn < currentExecutionTime; skippedTurn++) {
                addEvent(new SkippedTurnEvent(skippedTurn));
                println("SYSTEM: " + getShortName() + " skipped turn " + skippedTurn);
            }
        }

        if ((!isIORobot && (numSkippedTurns > MAX_SKIPPED_TURNS))
                || (isIORobot && (numSkippedTurns > MAX_SKIPPED_TURNS_WITH_IO))) {
            println("SYSTEM: " + getShortName() + " has not performed any actions in a reasonable amount of time.");
            println("SYSTEM: No score will be generated.");
            setHalt(true);
            waitWakeupNoWait();
            punishBadBehavior(BadBehavior.SKIPPED_TOO_MANY_TURNS);
            robotProxy.forceStopThread();
        }
    }
}

Here is my current best guess.

Note that Beaming's stacktrace includes:

...
at net.sf.robocode.host.proxies.BasicRobotProxy.forceStopThread(BasicRobotProxy.java:44)
at net.sf.robocode.battle.peer.RobotPeer.checkSkippedTurn(RobotPeer.java:649)
...

Therefore, RobotPeer.checkSkippedTurn() called BasicRobotProxy.forceStopThread()

The source code of RobotPeer.checkSkippedTurn() shows one and only one possible call to BasicRobotProxy.forceStopThread():

public void checkSkippedTurn() {
    // ...
    if (/* snip */) {
        println("SYSTEM: " + getShortName() + " has not performed any actions in a reasonable amount of time.");
        println("SYSTEM: No score will be generated.");
        /* snip */
        robotProxy.forceStopThread();
    }
}

Therefore, control flow must have passed through this path.

Note that this code sequence includes a println("SYSTEM: <Bot> has not performed any actions...") call, which must have been executed, yet no such SYSTEM message is present in Beaming's log.

The source code of the println() method:

public void println(String s) {
    synchronized (proxyText) {
        battleText.append(s);
        battleText.append("\n");
    }
}

Note the following:

1. Robocode is multithreaded. Heavy multithreading causes race conditions.
2. The variable battleText is a StringBuilder, which is not thread-safe (unlike a StringBuffer), which is.
3. The code modifies battleText, but only synchronizes against proxyText.

Therefore, EvBotNG is skipping turns, but no log messages (from either Robocode or the bot) are printed, because they are silently dropped due to race conditions in the println() method. The only thing you ultimately see is the ThreadDeath exception caused by Thread.stop() forcibly terminating the thread.

Presumably, this occurs because of unlucky interaction between RobotPeer.println() and RobotPeer.readOutText().

readOutText()                                 | println()
============================================= | ===============================
final String robotText =                      |
battleText.toString() + proxyText.toString(); |
                                              | battleText.append(s);
                                              | battleText.append("\n");
battleText.setLength(0);                      |
proxyText.setLength(0);                       |
return robotText;                             |

Therefore, the solution is to add synchronized(battleText) { guards around readOutText() and println().

Or maybe I'm way off track.

I think you're on the right track, it definitely feels like race conditions to me. However in this case you're wrong - both code blocks (println and readOutText) are synchronized with the same object, so they can't run at the same time. It isn't important which one they are synchronized with, it could even be a different object. See the sharedLock example here.

I think a good place to start looking is in the 'fixes' to the skipped turn behavior, which were documented here, as well as a very busy discussion on the wiki.

Well, println() is definitely broken somehow. There is no way that checkSkippedTurn() could have called forceStopThread() (as is shown in the stacktrace), without going through println() first, yet the print message never made its way to the console (as is shown in Beaming's log).

battleText is also cleared in initializeRound() and cleanup(), both of which have no synchronization guards at all.

I think skipped turns are working as they should, they're just not being reported.

Good catch. I've added synchronization guards on the initializeRound() and cleanup() and now I get the SYSTEM: Neuromancer 5.4 has not performed any actions in a reasonable amount of time. printed inside the correct round. The ThreadDeath is still printed in the round after, but I can live with that I guess.

Now for the next question, why is this even happening?

Would you please share the patch? My multithreading skills are not that good.

It seems that Skilgannon has filed a bug report on this issue, including the necessary patch.

See it now. Thanks.

@ MultiplyByZer0, excellent analysis. I would never spotted it.

To add to mysteries. Why I do not see dramatic drop in scores if I run Voidious' RoboRunner? I see missing SYSTEM messages in logs but it does not affect scores the same drastic way as with the literumble client. I also see no ThreadDeath messages. I am guessing RoboRunner catches those somehow and process the score differently.

Just looked at those functions showing in error stack, in checkSkippedTurn(), it says: SYSTEM: No score will be generated. before the thread is killed. Therefore the good news is that these errors should not affect your scores.

And oracle says An instance of ThreadDeath is thrown in the victim thread when the Thread.stop() method is invoked. and The top-level error handler does not print out a message if ThreadDeath is never caught.

therefore the error message you see in the console could simply because robocode caught that error to make some clean-ups.

actually robocode kills your thread when you are skipping 30 turns in a row (or if you have file access, this will be increased to 240).

It somehow affects your score if you performed very well in this round :P Ok, the impact should be really small, and sometimes it may happen when you are doing bad, so it should balance, but you know... still a strange behavior.

If the bug happens randomly, your final score shouldn't be affected at all — just getting stable slower.

I think "no score will be generated" means that affected bot will get no points. The opponent still gets points, the winner bonus is assigned as the very least.

Searching on my system and old backups I was able to find Robocode versions for 1.7.4.2, 1.8.1.0. I was also able to download 1.8.0.0 and 1.9.0.0. So I ran some experiments using Neuromaner vs RaikoMicro. Note, dev version of Neuromancer I'm using here prints out each time it gets a SkippedTurnEvent.

• I can't reproduce it in at all 1.7.4.2. All output looks perfect - every round has printed SYSTEM lines either saying win or death, and Neuromancer correctly gets win/death events. Neuromancer doesn't skip any turns.

• I can't reproduce it at all in 1.8.0.0 - not even "SYSTEM: Neuromancer 5.4 has not performed any actions in a reasonable amount of time." However there are some rounds where nothing is printed, which is very suspicious. Still no skipped turn events.

• In 1.8.1.0 I can sometimes get a "SYSTEM: Neuromancer 5.4 has not performed any actions in a reasonable amount of time.". I saw this 3 times out of 10+ battles. There are also the occasional suspicious blank rounds where Neuromancer prints nothing, and SYSTEM prints nothing. Still no skipped turn events.

• In 1.9.0.0 I can reproduce ThreadDeath exceptions. On average 1 per battle, but sometimes 0 and max I saw was 4. Average of 3 skipped turn events per battle.

To note: all of them had similar CPU constants (between 5.3ms and 5.4ms).

Other things I tried:

• Changing the GC algorithm doesn't affect the number of ThreadDeaths in 1.9.0.0

• Doing a single getDataDirectory() at the beginning of each round doesn't affect the number of ThreadDeaths in 1.9.0.0

This makes me think it is a Robocode bug - possibly multiple bugs, since we have very different behaviors here over the different versions. It also greatly worries me that bots which were fine in 1.7.4.2 have problems in new versions.

Hi Skilgannon,

Isolating robocode issues from jdk version. Would it be correct to assume that you ran everything with openjdk-8?

Yes, all OpenJDK 8

openjdk version "1.8.0_131"
OpenJDK Runtime Environment (build 1.8.0_131-8u131-b11-2ubuntu1.16.04.3-b11)
OpenJDK 64-Bit Server VM (build 25.131-b11, mixed mode)

Here is the old installs I used to do these tests, I'd appreciate if someone could try them on Windows as well.

Tying into what Beaming said about the sleep in the internals not happening when set to max, perhaps this has been changed since earlier versions of Robocode. However there are more things that may be an issue as well, I don't think that is the total of what is buggy here. The missing prints in a round particularly concerns me (and yes, this still happens in recent versions).

If this don't happen when running the rumble, I suggest you use -Ddebug=true command line flag for robocode, which disables skipped turns & painting limit, to temporarily "solve" this problem.