My favorite (having not read the post but just gone through CS undergrad in Java about four times by weight) was the mandatory locking in multithreaded programs in Java -- you can have parent threads and child threads, and you can set variables in your parent threads before the child threads execute, but unless you're doing it in a locking way (eg. with a shared semaphore) there's no guarantee that the parent thread actually executes before the child thread.
That was a fun one to explain in office hours.
"No no, it's not enough to just put this before that, you have to establish a Happens-Before relationship."
>you can set variables in your parent threads before the child threads execute, but unless you're doing it in a locking way (eg. with a shared semaphore) there's no guarantee that the parent thread actually executes before the child thread
There are definitely some surprising outcomes from the JMM, but I'm pretty sure this isn't one of them. If I recall correctly, starting a thread automatically synchronizes the current thread with the first thing the child thread does, without any explicit synchronization. So anything done prior to that point in the parent thread will be visible in the child.
I do not have the code here to show you, but if I am exaggerating it's a small fib.
Maybe the child thread was created some time before it was started, and the initialization happened some time between those events. It was an example that a friend was working for a class, and as they often do, came to me for help on why it wasn't behaving as expected.
I looked at it and I could see the problem, "you don't have any locks shared between parent and child." We added a lock, synchronized it once in each thread, and the problem went away. That's definitely a surprising outcome for anyone who is accustomed to procedural programming.
The implicit synchronization happens at the point the child thread is started, regardless of when it was allocated. Perhaps it was already started prior to the initialization, in which case some form of explicit synchronization would be required to ensure visibility.
FYI, the relevant section of the JLS is 17.4.4: "An action that starts a thread synchronizes-with the first action in the thread it starts." [1].
> The write of the default value (zero, false or null) to each variable synchronizes-with the first action in every thread. Although it may seem a little strange to write a default value to a variable before the object containing the variable is allocated, conceptually every object is created at the start of the program with its default initialized values.
I can't get that to jibe in my memory though against:
> The final action in a thread T1 synchronizes-with any action in another thread T2 that detects that T1 has terminated.
> T2 may accomplish this by calling T1.isAlive() or T1.join().
If we were looking into the threads, it would have been obvious that we should wait for the thread to finish before expecting to find a completed value in it other than the default. That's not what happened.
It's unfortunate that I didn't get a copy of the code in question, because it was a genuine idiosyncrasy in Java and by the time I got around to asking my friend for a copy, he had already rewritten his code in a different way that didn't hit the same bug.
Didn't use revision control, got his degree didn't care to repeat the same miserable exercises, now probably hacking ruby in his day job like me (except I promise I'm using revision control.)
I can guarantee you the class was not offered the option of using Java 7. I'll take your word for it if you say this is not a new feature, but I'm sure the story went just about how I've laid it out.
Yes internally. Somewhere deep in the bowels of the akka code, they are communicating state between threads (last I checked they were using a java concurrent queue by default).
At a bare minimum the changes to the location of the head/tail of a queue will need to be visible to multiple threads requiring volatile variables.
The java concurrent queues don't actually use synchronization blocks, they use native CAS operations but thinking that akka removes shared thread state entirely is dangerous.
Even lots of low latency standard Java is tending towards single threaded non blocking models without synchronisation.
The costs of synchronised blocks is high on the JVM. They defer to the operating system for thread scheduling, they enforce a memory barrier meaning all data is flushed out to RAM rather than CPU caches, and multithreaded code dirties CPU caches, reducing performance.
More server side code should be single threaded than not in my opinion.
You can write lock free algorithms that are not single threaded as well. If you design your system so that only 1 thread is doing the writing of data you avoid a lot of the negative performance implications you mention.
That said, I completely agree. Many times making a single threaded system faster will improve performance vs making it parallel.
If this is interesting to you then I'd highly recommend Java Concurrency in Practice [1]. It goes through the then-new java.util.concurrent package, but more importantly, does a really good job of making all of the theoretical concepts straightforward to grasp
Great book indeed. Along with effective Java, its one of my alltime favorites. But since its a Java book, it doesnt if I recall correctly detail how things are implemented at the core JVM level. Does anyone know a good book or source about that?
Fact #0. If you write 'synchronized' in Java code you are doing it wrong. Seriously. 'synchronized' indicates that the programmer is re-inventing and/or re-implementing a solved problem. In Java parallelism and concurrency for most real-wold problems are handled by frameworks or covered by well-known patterns.
That was a fun one to explain in office hours.
"No no, it's not enough to just put this before that, you have to establish a Happens-Before relationship."