RE: JavaMemoryModel: Immutablity and Final Fields

From: David Holmes (dholmes@dstc.edu.au)
Date: Sun Feb 25 2001 - 22:57:29 EST


> Assuming that everything is done more-or-less perfectly, then when you
> have finished initializing the field, you can do what you
> said originally and make the shared reference to the object volatile. You

> can then read that reference *once* in the reader thread and copy it to a
> non-volatile local. After that, use the non-volatile local instead of the

> shared volatile. You only need to perform the memory barrier once, after
all.

Copying into a thread-local variable certainly avoids the overhead of using
either a volatile reference or a synchronized accessor to the reference, but
requires a different approach to the way things are typically done at
present. If access to the object is within a simple scope then it is
feasible to use a local stack-based variable to hold the thread-local copy.
In general however these shared variables tend to be accessed from various
methods in various classes - take the Singleton pattern as an example. In
that case the only way to get a thread-local copy is to use a real
ThreadLocal accessed via the Singleton class.

public class Singleton {
    private static theInstance = null;

    private ThreadLocal localInstance = new ThreadLocal();

    public static Singleton getInstance() {
        Singleton local = (Singleton) localInstance.get();
        if (local == null) {
            synchronized (Singleton.class) {
                if (theInstance == null)
                    theInstance = new Singleton();
             }
             localInstance.set(theInstance);
             return theInstance;
        }
        return local;
    }
}

I wonder how the overhead of ThreadLocal compares to either a synchronized
accessor or a volatile reference. ;-)

Many of these situations boil down to what you describe: each thread needs
to get synchronized access to the reference to the shared object only the
first time they access it. Finding a way to pay for that synchronization
only once is one of the goals.

Cheers,
David

-------------------------------
JavaMemoryModel mailing list - http://www.cs.umd.edu/~pugh/java/memoryModel



This archive was generated by hypermail 2b29 : Thu Oct 13 2005 - 07:00:30 EDT