java - Synchronization on the local variables

Question

Today I was faced with the method constructServiceUrl() of the org.jasig.cas.client.util.CommonUtils class. I thought he was very strange:

final StringBuffer buffer = new StringBuffer();

synchronized (buffer)
{
    if (!serverName.startsWith("https://") && !serverName.startsWith("http://"))
    {
        buffer.append(request.isSecure() ? "https://" : "http://");
    }

    buffer.append(serverName);
    buffer.append(request.getRequestURI());

    if (CommonUtils.isNotBlank(request.getQueryString()))
    {
        final int location = request.getQueryString().indexOf(
                artifactParameterName + "=");

        if (location == 0)
        {
            final String returnValue = encode ? response.encodeURL(buffer.toString()) : buffer.toString();

            if (LOG.isDebugEnabled())
            {
                LOG.debug("serviceUrl generated: " + returnValue);
            }

            return returnValue;
        }

        buffer.append("?");

        if (location == -1)
        {
            buffer.append(request.getQueryString());
        }
        else if (location > 0)
        {
            final int actualLocation = request.getQueryString()
                    .indexOf("&" + artifactParameterName + "=");

            if (actualLocation == -1)
            {
                buffer.append(request.getQueryString());
            }
            else if (actualLocation > 0)
            {
                buffer.append(request.getQueryString().substring(0, actualLocation));
            }
        }
    }
}

Why did the author synchronizes a local variable?

question from:https://stackoverflow.com/questions/6123024/synchronization-on-the-local-variables

Answer 1

This is an example of manual "lock coarsening" and may have been done to get a performance boost.

Consider these two snippets:

StringBuffer b = new StringBuffer();
for(int i = 0 ; i < 100; i++){
    b.append(i);
}

versus:

StringBuffer b = new StringBuffer();
synchronized(b){
  for(int i = 0 ; i < 100; i++){
     b.append(i);
  }
}

In the first case, the StringBuffer must acquire and release a lock 100 times (because append is a synchronized method), whereas in the second case, the lock is acquired and released only once. This can give you a performance boost and is probably why the author did it. In some cases, the compiler can perform this lock coarsening for you (but not around looping constructs because you could end up holding a lock for long periods of time).

By the way, the compiler can detect that an object is not "escaping" from a method and so remove acquiring and releasing locks on the object altogether (lock elision) since no other thread can access the object anyway. A lot of work has been done on this in JDK7.

---

Update:

I carried out two quick tests:

1) WITHOUT WARM-UP:

In this test, I did not run the methods a few times to "warm-up" the JVM. This means that the Java Hotspot Server Compiler did not get a chance to optimize code e.g. by eliminating locks for escaping objects.

JDK                1.4.2_19    1.5.0_21    1.6.0_21    1.7.0_06
WITH-SYNC (ms)         3172        1108        3822        2786
WITHOUT-SYNC (ms)      3660         801         509         763
STRINGBUILDER (ms)      N/A         450         434         475

With JDK 1.4, the code with the external synchronized block is faster. However, with JDK 5 and above the code without external synchronization wins.

2) WITH WARM-UP:

In this test, the methods were run a few times before the timings were calculated. This was done so that the JVM could optimize code by performing escape analysis.

JDK                1.4.2_19    1.5.0_21    1.6.0_21    1.7.0_06
WITH-SYNC (ms)         3190         614         565         587
WITHOUT-SYNC (ms)      3593         779         563         610
STRINGBUILDER (ms)      N/A         450         434         475

Once again, with JDK 1.4, the code with the external synchronized block is faster. However, with JDK 5 and above, both methods perform equally well.

Here is my test class (feel free to improve):

public class StringBufferTest {

    public static void unsync() {
        StringBuffer buffer = new StringBuffer();
        for (int i = 0; i < 9999999; i++) {
            buffer.append(i);
            buffer.delete(0, buffer.length() - 1);
        }

    }

    public static void sync() {
        StringBuffer buffer = new StringBuffer();
        synchronized (buffer) {
            for (int i = 0; i < 9999999; i++) {
                buffer.append(i);
                buffer.delete(0, buffer.length() - 1);
            }
        }
    }

    public static void sb() {
        StringBuilder buffer = new StringBuilder();
        synchronized (buffer) {
            for (int i = 0; i < 9999999; i++) {
                buffer.append(i);
                buffer.delete(0, buffer.length() - 1);
            }
        }
    }    

    public static void main(String[] args) {

        System.out.println(System.getProperty("java.version"));

        // warm up
        for(int i = 0 ; i < 10 ; i++){
            unsync();
            sync();
            sb();
        }

        long start = System.currentTimeMillis();
        unsync();
        long end = System.currentTimeMillis();
        long duration = end - start;
        System.out.println("Unsync: " + duration);

        start = System.currentTimeMillis();
        sync();
        end = System.currentTimeMillis();
        duration = end - start;
        System.out.println("sync: " + duration);

        start = System.currentTimeMillis();
        sb();
        end = System.currentTimeMillis();
        duration = end - start;
        System.out.println("sb: " + duration);  
    }
}