Java Specialists' Java Training Europehome of the java specialists' newsletter

The Java Specialists' Newsletter
Issue 038a2001-12-28 Category: Performance Java version:

GitHub Subscribe Free RSS Feed

Counting Objects Clandestinely

by Dr. Heinz M. Kabutz

Welcome to the 38th edition of The Java(tm) Specialists' Newsletter, sent to 2226 Java experts in 70 countries, with Egypt as latest addition. I'm missing a few African countries, ok, it's worse than missing, I only have South Africa, Ghana, Nigeria, Zimbabwe and Egypt, so if you are from another African country, please send me an email.

Talking about Africa, I didn't include Mauritius in my list of African countries, although politically I probably should have. If you don't know what Mauritius is, it is a small island in the Indian Ocean inbetween South Africa and India. Mauritius was colonised by the French (who only drank and fought and never did much for the country except introduce bureaucracy ;-) and later properly colonised by the British. The result is that English is the official language, whereas French (and derivatives thereof) is the spoken language.

The people in Mauritius are amazing. Extremely friendly, and not only to sell you things. The beaches are fantastic, the water warm, lots of things to see and do, such as waterskiing, etc. Oh, and the hotels! Such service you've never seen anywhere.

Counting Objects Clandestinely

A few months ago I received a frantic phone call from a friend who had gotten himself seriously stuck. His program was running out of memory and he didn't know where the objects were being created. He ran his program in a commercial memory analyser, which only had the effect that the memory analyser crashed. (If you want to have a laugh, try running any big Java application (e.g. Together/J or JBuilder 4) in a memory analyser). If it runs, you'll see that the success of your product does not depend on good programming but on good marketing.

The trick I'm going to show you in this newsletter essentially saved my friend's project, and it is sooo easy and cheap, you won't believe it. Please note however, that this is only an initial implementation and that you should extend the functionality if you want to use it a lot. I have myself used this idea to effectively find memory problems in my own applications.

Imagine for a second that there was some way of knowing when an Object was created. Well, if you really want that functionality, why don't you add it? Yes, why don't you change java/lang/Object? Sure, it's intrusive, but very effective. The test code that you add to your code to find the memory leaks would be removed anyway (or commented out?) before the final build, so why not? It took me a while to get all this right, which is why you didn't see a newsletter for a while, but here is the code, which I tried on Windows JDK 1.3.1. On other VMs it might reformat your harddrive, sell your credit card details or promote you to sales manager.

package java.lang;
import java.util.*;
public class Object {
  // ... stick this at the bottom of Object.java:
  private static HashMap countMap = new HashMap();
  private static boolean counting = true;
  private static int totalObjects = 0;

  private class Counter {
    int value = 1;
  }

  public Object() {
    synchronized(Object.class) {
      if (counting) {
        counting = false;
        totalObjects++;
        Counter c = (Counter)countMap.get(getClass());
        if (c == null) {
          countMap.put(getClass(), new Counter());
        } else {
          c.value++;
        }
        counting = true;
      }
    }
  }
  public synchronized static void ___resetObjectCreationStats() {
    counting = false;
    totalObjects = 0;
    countMap.clear();
    counting = true;
  }
  public static void ___printObjectCreationStats() {
    ___printObjectCreationStats(System.out);
  }
  public synchronized static void ___printObjectCreationStats(java.io.PrintStream out) {
    counting = false;
    out.println("Total number of objects: " + totalObjects);
    TreeSet sorted = new TreeSet(new Comparator() {
      public int compare(Object o1, Object o2) {
        int value1 = ((Counter)((Map.Entry)o1).getValue()).value;
        int value2 = ((Counter)((Map.Entry)o2).getValue()).value;
        int result = value2 - value1;
        if (result == 0) {
          String classname1 = ((Class)((Map.Entry)o1).getKey()).getName();
          String classname2 = ((Class)((Map.Entry)o2).getKey()).getName();
          return classname1.compareTo(classname2);
        }
        return result;
      }
    });
    sorted.addAll(countMap.entrySet());
    Iterator it = sorted.iterator();
    while(it.hasNext()) {
      Map.Entry entry = (Map.Entry)it.next();
      out.println("\t" + ((Counter)entry.getValue()).value
        + "\t" + ((Class)entry.getKey()).getName());
    }
    out.println();
    counting = true;
  }
}

You can now know exactly how many objects were created since you last reset the counters. [Actually, not really exactly. This approach will not tell you when arrays are created, but you will find out about any non-array objects that were made.] The way you would use this in your code is as follows:

import java.util.*;
public class Test {
  public static void testStringConcatenation() {
    ___resetObjectCreationStats();
    int localHost = 0x7F000001;
    String s =
      ((localHost>>24)&0xff) + "."
      + ((localHost>>16)&0xff) + "."
      + ((localHost>>8)&0xff) + "."
      + (localHost&0xff);
    System.out.println("Objects created to make an IP address String");
    System.out.println("--------------------------------------------");
    ___printObjectCreationStats();
  }
  public static void testLinkedListCreation() {
    ___resetObjectCreationStats();
    LinkedList daysOfWeek = new LinkedList();
    System.out.println("Objects created to make a LinkedList");
    System.out.println("------------------------------------");
    ___printObjectCreationStats();
    daysOfWeek.add("Sunday");
    daysOfWeek.add("Monday");
    daysOfWeek.add("Tuesday");
    daysOfWeek.add("Wednesday");
    daysOfWeek.add("Thursday");
    daysOfWeek.add("Friday");
    daysOfWeek.add("Saturday");
    System.out.println("and after adding 7 elements to the LinkedList");
    System.out.println("---------------------------------------------");
    ___printObjectCreationStats();
  }
  public static void testHashMapCreation() {
    ___resetObjectCreationStats();
    HashMap monthsVsLength = new HashMap();
    System.out.println("Objects created to make a HashMap");
    System.out.println("---------------------------------");
    ___printObjectCreationStats();
    monthsVsLength.put("January", new Integer(31));
    monthsVsLength.put("February", new Integer(28));
    monthsVsLength.put("March", new Integer(31));
    monthsVsLength.put("April", new Integer(30));
    monthsVsLength.put("May", new Integer(31));
    monthsVsLength.put("June", new Integer(30));
    monthsVsLength.put("July", new Integer(31));
    monthsVsLength.put("August", new Integer(31));
    monthsVsLength.put("September", new Integer(30));
    monthsVsLength.put("October", new Integer(31));
    monthsVsLength.put("November", new Integer(30));
    monthsVsLength.put("December", new Integer(31));
    System.out.println("and after adding 12 elements to the HashMap");
    System.out.println("-------------------------------------------");
    ___printObjectCreationStats();
  }
  public static void main(String args[]) {
    System.out.println("Objects created to get the VM started");
    System.out.println("-------------------------------------");
    ___printObjectCreationStats();
    testStringConcatenation();
    testLinkedListCreation();
    testHashMapCreation();
  }
}

When you compile the test class you should also at the same time point it to your Object.class otherwise you'll get a compiler moan. You have to run Test by including the new Object.class file in your bootclasspath, for example java -Xbootclasspath/p:. Test, which will prepend the current directory to the boot classpath. The output is rather long and would be different for each VM version.

Objects created to get the VM started
-------------------------------------
Total number of objects: 784
        514     java.util.Hashtable$Entry
        104     java.lang.String
        25      java.lang.StringBuffer
        22      java.util.Locale
        13      java.io.File
        11      sun.security.action.GetPropertyAction
*snip - run it yourself to see the rest*

Objects created to make an IP address String
--------------------------------------------
Total number of objects: 6
        5       java.lang.String
        1       java.lang.StringBuffer

Objects created to make a LinkedList
------------------------------------
Total number of objects: 2
        1       java.util.LinkedList
        1       java.util.LinkedList$Entry

and after adding 7 elements to the LinkedList
---------------------------------------------
Total number of objects: 9
        8       java.util.LinkedList$Entry
        1       java.util.LinkedList

Objects created to make a HashMap
---------------------------------
Total number of objects: 1
        1       java.util.HashMap

and after adding 12 elements to the HashMap
-------------------------------------------
Total number of objects: 25
        12      java.lang.Integer
        12      java.util.HashMap$Entry
        1       java.util.HashMap

What are the disadvantages of this approach, and what are the alternatives? The disadvantages that I can think of immediately are that you don't see the arrays like this and that you cannot see where objects were created from. I don't know how to solve the problems with the arrays, but you could easily change the code to also remember the stack trace of each object that gets constructed. You could then analyse the stack traces and find out exactly where the objects are constructed from. The alternatives I can think of is to use java -Xrunhprof (to perform JVMPI heap, cpu, or monitor profiling) or to use an expensive commercial tool. The disadvantage of some of the commercial tools is that if your program quickly eats memory the memory tools tend to fall on their back and play dead.

Have you ever wondered why Swing is so slow? Have a look at this test code:

import javax.swing.*;
import java.awt.BorderLayout;
import java.awt.event.*;

public class TestGUI extends JFrame {
  public TestGUI() {
    super("TestGUI");
    getContentPane().add(new JTextArea(), BorderLayout.CENTER);
    JButton press = new JButton("Press Me!");
    press.addActionListener(new ActionListener() {
      public void actionPerformed(ActionEvent e) {
        ___printObjectCreationStats();
        ___resetObjectCreationStats();
      }
    });
    getContentPane().add(press, BorderLayout.SOUTH);
    setSize(500, 600);
    setDefaultCloseOperation(JFrame.EXIT_ON_CLOSE);
    show();
  }
  public static void main(String[] args) {
    new TestGUI();
  }
}

I ran this code and moved my mouse in circles for 10 seconds in the JTextArea. Here are the first couple of entries that were shown when I pressed the "Press Me!" button:

     Total number of objects: 15622
        2867    java.lang.String
        1545    java.lang.ref.Finalizer
        1384    java.awt.event.MouseEvent
        1326    java.awt.Point
        1200    java.lang.StringBuffer
        1047    java.util.Hashtable$Entry
        769     java.util.WeakHashMap$WeakKey
        750     java.awt.EventQueueItem
        706     sun.awt.EventQueueItem
        648     java.awt.Rectangle
        316     sun.java2d.loops.GraphicsPrimitiveProxy
        266     sun.awt.MostRecentKeyValue
        253     java.awt.geom.AffineTransform
        134     java.util.HashMap$Entry

Why so many Strings? I can only assume that that has to do with the pluggable look & feel. The reason I say that is because if I press the button again (after moving the mouse in circles a few times), I get the following output:

Total number of objects: 5257
        891     java.lang.ref.Finalizer
        831     java.awt.event.MouseEvent
        796     java.awt.Point
        456     java.util.WeakHashMap$WeakKey
        436     java.awt.EventQueueItem

It gets quite interesting when we look at different VMs, just remember the warning at the beginning of this newsletter ;-)

That's all for this week and this year. I wish you all the best for 2002, may you find much opportunity to tell others about this newsletter ;-) [and make me prosper in the process ;]

Kind regards, and now I must get back to spending a few sunny days with my wife & two kiddies.

Heinz

Performance Articles Related Java Course

Java Master
Java Concurrency
Design Patterns
In-House Courses



© 2010-2014 Heinz Kabutz - All Rights Reserved Sitemap
Oracle and Java are registered trademarks of Oracle and/or its affiliates. Other names may be trademarks of their respective owners. JavaSpecialists.eu is not connected to Oracle, Inc. and is not sponsored by Oracle, Inc.
@CORE_THE_BAND #RBBJGR