[logback-dev] [JIRA] Commented: (LBCORE-97) Starvation on AppenderBase.doAppend

[Joern Huxhorn (JIRA)]

    [ http://jira.qos.ch/browse/LBCORE-97?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=11141#action_11141 ] 

Joern Huxhorn commented on LBCORE-97:
-------------------------------------

> Standard java synchronization was never guaranteed to be fair. Java 6 (or the 
> OS scheduler) is simply doing more optimizations that it didn't do previously
> but was always allowed to do.

+1

> However I don't see why this is a problem unless logging is your bottleneck

Logging is, indeed, the bottleneck of our application at the moment, but this is expected, "by design" and will change at a later time.

- We are rolling out a new version of our app so logging is still unusually high
- We are doing a proprietary CMS with XSLT as the template language. There are several XSLT developers receiving compilation specific INFO, WARN and ERROR events via my MulticastSocketAppender.
- The slowest receiver essentially dictates the overall speed of the application

> If this is the case I'd investigate why you're spending so much time in
> logging and if that's unfixable change to asynchronous logging.

The MultiplexAppender is already semi-asynchronous, meaning that it has a buffer of fixed size that is filled before the appender blocks.

See http://apps.sourceforge.net/trac/lilith/wiki/AppenderPitfalls

We've done this "abusive" logging for over a year with *very* acceptable performance.

Then we moved from a single-core Java 1.5 server to a multi-core Java 1.6 server and, suddenly, performance was extremely crappy. I narrowed the problem down to the issue discussed here and never really expected any discussion whether this patch should be applied or not because - honestly - I think I did a pretty good job citing all relevant documentation, showing up that we all worked with false presumptions - and this explicitly includes myself!

I *knew* that fairness wasn't guaranteed but I *assumed* that something like nearly-fair or somewhat-unfair could be expected. This assumption was wrong.

> because synchronous logging will in this case turn your application into a
> single threaded one

There's some truth in this statement.;)
However, large amounts of code won't log anything while other parts log in relatively big bursts so, theoretically, all CPUs should be able to work quite a bit, anyway, even if parts of the application are running semi-single-threaded.


> Starvation on AppenderBase.doAppend
> -----------------------------------
>
>                 Key: LBCORE-97
>                 URL: http://jira.qos.ch/browse/LBCORE-97
>             Project: logback-core
>          Issue Type: Bug
>          Components: Appender
>    Affects Versions: 0.9.15
>            Reporter: Joern Huxhorn
>            Assignee: Logback dev list
>            Priority: Critical
>         Attachments: AppenderBaseLock2.patch, SynchronizedVsFairLock.java
>
>
> The problem we are facing here is that several threads are trying to obtain the object monitor of the exact same resource, namely the Appender.
> The assumption that multiple threads waiting for ownership of a monitor would receive the monitor in the order that they tried to acquire it is simply incorrect. This is documented behavior.
> See the last paragraph of [1]:
> "Likewise, no assumptions should be made about the order in which threads are granted ownership of a monitor or the order in which threads wake in response to the notify or notifyAll method. An excellent reference for these topics is Chapter 9, "Threads," in Joshua Bloch's book Effective Java Programming Language Guide. "
> The documentation of Object.notifyAll() [2] states the following:
> "[..] The awakened threads will compete in the usual manner with any other threads that might be actively competing to synchronize on this object; for example, the awakened threads enjoy no reliable privilege or disadvantage in being the next thread to lock this object."
> The documentation in [5] also lists the following as a weak-spot of Built-in Synchronization in J2SE 1.4.x:
> "No way to alter the semantics of a lock, for example, with respect to reentrancy, read versus write protection, or fairness."
> In LBCORE-96 Ceki stated the following:
> "Logback is using the standard synchronization mechanism found in the JDK. You are saying that that mechanism is unsuitable which, truth be told, sounds quite suspicious."
> Yes, it's unsuitable in this situation because obtaining the object monitor is not guaranteed to be fair. It's not working in a "first come, first serve" manner. I assumed that, too, but it isn't the case. I had the exact same problem in Lilith some time ago because I made exactly this very same wrong assumption.
> Fairness of object monitor lock acquisition seems to be logical and a "good thing" but it's not specified that way, and for good reasons.
> Without fairness in place, the VM can optimize the execution of an application much better. A context switch is a costly operation for a CPU so performance is increased significantly if such a switch can be omitted.
> Concerning a test-case, this would be pretty hardcore to implement since it's 100% implementation dependent. One implementation *may* handle the locking of object monitors fairly while others don't.
> Therefore I'll try the following first:
> I assume I could convince you that object monitor acquisition (OMA) is not fair.
> If we take that for granted the following scenario should show my point:
> There are 4 Threads (e.g. 4 Threads handling concurrent webapp-requests) that have one single chokepoint, the appender.
> Since OMA isn't fair, it's possible that only some of those threads can work at all (see [3]). Exactly that is happening right now in our webapp running on a 4-core Solaris on JDK6.
> My next assumption is that we both agree that this isn't acceptable behavior.
> Logging should be handled "first come, first served" so that if Thread A is waiting to append an event and Thread B is waiting to append an event subsequently, the actual appending order should be A, then B.
> This is currently not the case. It *may* be the case but there is no guarantee of it.
> One could even argue that the logging system is working incorrectly (aside from the starvation problem) because events are appended in a different order than the actual execution of the logger call.
> The only way to prevent this is the introduction of fairness into the locking process. The way to do this is to use ReentrantLock [4].
> From the ReentrantLock javadoc:
> "The constructor for this class accepts an optional fairness parameter. When set true, under contention, locks favor granting access to the longest-waiting thread. Otherwise this lock does not guarantee any particular access order. Programs using fair locks accessed by many threads may display lower overall throughput (i.e., are slower; often much slower) than those using the default setting, but have smaller variances in times to obtain locks and guarantee lack of starvation."
> This is exactly what is absolutely necessary for a logging framework.
> Logging must *NOT* be able to introduce a starvation problem into a multi-threaded application! And it does exactly that right now.
> I doubt that the performance will decrease in a significant manner due to the use of a fair ReentrantLock but even if this would be the case it would be necessary anyway.
> Otherwise, appending simply isn't working correctly.
> [1] http://java.sun.com/j2se/1.5.0/docs/guide/vm/thread-priorities.html
> [2] http://java.sun.com/javase/6/docs/api/java/lang/Object.html#notifyAll()
> [3] http://en.wikipedia.org/wiki/Starvation_(computing)
> [4] http://java.sun.com/javase/6/docs/api/java/util/concurrent/locks/ReentrantLock.html
> [5] http://java.sun.com/developer/technicalArticles/J2SE/concurrency/

-- 
This message is automatically generated by JIRA.
-
If you think it was sent incorrectly contact one of the administrators: http://jira.qos.ch/secure/Administrators.jspa
-
For more information on JIRA, see: http://www.atlassian.com/software/jira