The Java Explorer

Tips and insights on Java

  • Subscribe

  • If you find this blog useful, please enter your email address to subscribe and receive notifications of new posts by email.

    Join 39 other followers


Posted by Eyal Schneider on June 17, 2009

Any Java programmer has encountered at least once the disturbing OutOfMemoryError (OOME) message. The interesting thing is that this error occurs for different reasons, and the diagnosis and treatment in each case is completely different. In this post I will try to cover some of these cases, show how to reproduce them and how to deal with them.
Some important things to remember about OOME:

  • As the name implies, this is an Error, and not an Exception. Errors indicate a serious condition in the JVM. Catching Errors is usually useless – the code doesn’t have the means to recover from them. Since they are usually not caught, they result in some thread getting terminated.
  • The stack trace is not always present in OOME, since its creation also requires memory, which may not be available.
  • The circumstances where an OOME is thrown may vary between different JVM implementations, due to a different memory management model. The textual description of the error may also vary. This post refers to to my own experience with Sun’s JVM, although I believe that most of it is relevant for other implementations as well.
  • As explained above, OOME is not recoverable at runtime. Therefore, the right strategy is to avoid its recurring once it happens, by means of configuration, design, or hardware modifications.


Java memory organization

Let’s first start with a very short overview of how a Java process organizes its address space. Basically, the memory is divided into heap memory and non-heap memory.  Modern JVM implementations use a generational heap model, in order to improve garbage collection performance. In this model, the heap area is usually split into 3 generations: young, old (“tenured”) and permanent. The young generation is where new objects created with the new operator are allocated, and it is designed to hold “young” objects. Older objects are moved to the tenured generation during GC. The permanent generation is not common to all JVM implementations, and it differs from the other generations – it stores (among other things) metadata of classes that have been loaded. The permanent generation is sometimes regarded as part of the heap, and sometimes not (The heap sizing flags -Xmx and -Xms, for example, do not include the permanent generation size).
Non-heap memory is the additional memory space outside of the areas reserved for the heap, and it can be used for different purposes. I will refer to this memory as native memory.

java.lang.OutOfMemoryError: Java heap space

This common error indicates that an object creation failed because there was no available heap space for it. It will never be thrown unless the JVM tries a full garbage collection first, including removal of softly/weakly referenced objects. When the stack trace is present for this particular OOME, it is usually irrelevant – the specific allocation that could not be satisfied gives us no clue regarding previous allocations that probably caused the problem. Same with the thread where the error occured – It may occur in any thread, since they all share the same heap.
Following is a sample code that eventually generates this OOME:
List<String> list = new ArrayList<String>();
    list.add(new String("Consume more memory!"));
1) Check for memory leaks
  • Use the memory tab in JConsole in order to examine the graph of memory usage of your application over time. Focus on the “Tenured generation”, and click on “perform GC” every once in a while, to see whether the used memory returns to some expected baseline level after every garbage collection.
  • Run your application with the following JVM flags: -verbose:gc and -XX:+PrintGCDetails. This will add console outputs with details on every garbage collection that takes place. Using these outputs, you can analyze the memory consumption pattern, and see whether the memory usage increases constantly where it shouldn´t.                                    

 The following tools can help you analyze the reasons for the memory leak:

  • Run jmap -histo <java process id> at different stages of your application execution. This command prints the number of instances and total used memory per Java class. You can identify classes that have an unreasonable number of instances, or classes for which the number of instances increases unexpectedly. This can limit the search space in your code, when looking for the bug.
  • Generate a heap dump using JConsole or jmap, and analyze it using a heap analyzer (e.g. JHat, IBM HeapAnalyzer). The heap dump contains the full state of the heap at the point of time where the dump was created. JHat allows examining this heap snapshot, navigating through the different instances, and even executing complex queries (OQL language). An advanced feature allows comparing two snapshots taken at different points of time, in order to identify the allocations that occurred between the two. If you want the dump to be created automatically when the OOME occurs, add the -XX:+HeapDumpOnOutOfMemoryError JVM flag.

2) Increase the heap size

Use the JVM flags -Xms and -Xmx to set the initial and maximal heap size, respectively. Make sure not to exceed the physical memory size, otherwise you encourage paging, which affects performance. Sometimes installing additional physical memory is required.

3) Reconsider the design

An inadequate design can easily consume too much heap space and cause this OOME. Sometimes, it is useful to change the design in a way that compromises some desired system property (such as performance), but reduces memory use dramatically. For example, disk/database storage can replace some of the data structures in RAM, and memory sensitive mechanisms such as weak/soft references can be introduced.

java.lang.OutOfMemoryError: PermGen space

When the permanent generation becomes full, this error occurs, regardless of the available space in the object allocation area (young + old generations). Note that this specific message is issued by Sun’s JVM, and other JVM implementations may have different messages, specially those implementations with a different policy for storing class information (such as IBM’s JVM).
I will reproduce this OOME in two different ways. The first one uses a custom class loader to load too many classes. Assume that we have a custom class loader named MyClassLoader, capable of loading the class Dummy:
List<Class<?>> classes = new ArrayList<Class<?>>(); while(true){ MyClassLoader cl = new MyClassLoader(); try{ classes.add(cl.loadClass("Dummy")); }catch (ClassNotFoundException e) { e.printStackTrace(); } }
Note that we are actually loading the same class over and over, with a different class loader (remember that a class is identified by the combination of its name and its class loader object). Also, note that we store references to the loaded classes. If we don’t do so, the garbage collector realizes that the classes are not used, and performs garbage collection on the permanent generation.

A second way of filling up the permanent generation, is by exploiting the String intern pool. The class String manages a pool of strings that have been interned, using String.intern(…). This pool is also a part of the permanent generation.

List<String> list = new ArrayList<String>();
int i=0;
    list.add(("Consume more memory!"+(i++)).intern());

The examples above highlight the common cases where this OOME occurs in practice:

  • Applications which load classes dynamically, such as web servers and application containers
  • A class loader with a memory leak
  • Applications that make use of string interning, in order to improve performance or reducy heap usage


Start by analyzing. A good place to start with is the stack trace. If it contains the java.lang.String.intern(..) call, then the reason is probably string interning. Alternatively, if it contains the java.lang.ClassLoader.defineClass(..) call, then you are probably dealing with a class loading problem. In the latter case, you may want to track the loading/unloading of classes, using the JVM flags -XX:+TraceClassLoading(or -verbose:class) and -XX:+TraceClassUnloading. Also, make sure you don’t have the flag -Xnoclassgc, because it prevents GC on classes, and can easily cause this OOME in applications that keep loading classes dynamically. As a part of the analysis, you can also run the utility jmap -permstat <process id> (not available for Windows). It provides details on all class loaders that have been used since the application was launched. 

Once you have pinpointed the problematic piece of code, check whether the design can be changed to consume less memory. If this is not an option, use the -XX:MaxPermSize flag to increase the maximal size of the permanent generation (e.g. -XX:MaxPermSize=256m).
Remember that increasing the values of -Xmx and -Xms flags instead will not change a thing. As explained before, these flags control the size of the young+old generations only, and do not affect the permanent generation.


java.lang.OutOfMemoryError: unable to create new native thread

Whenever we start a a  thread in Java, the JVM allocates space in native memory, designated for the thread’s call stack. This space is freed only after the thread is terminated. In case that there is not enough native memory for the stack space, we get this error. This can easily be reproduced with the following thread spawner code:
    new Thread(new Runnable(){
        public void run() {
            try {
            } catch(InterruptedException e) { }        
This error is common in applications that try to manage too many live threads, as in the example above. As always, first ask yourself whether the design is reasonable. There are some generic solutions for reducing the number of threads. For example, an application that creates a thread per connection has a scalability problem in its design. The IO multiplexing feature introduced in Java 1.4 can solve this, and reduce dramatically the number of threads (See class java.nio.channels.Selector).
If you wish to try solving this problem by changing the memory settings only, then any of the following can help:
1) Reduce the heap size using -Xmx and -Xms flags. Sometimes, inexperienced programmers increase the heap size when they encounter this OOME, but it only makes the condition worse.
2) Reduce the permanent generation size using -XX:MaxPermSize
3) Reduce the size of memory allocated per stack using -Xss (e.g. -Xss512k). The default values depends on the environment, and range between256k and 1024k. Remember that when you reduce this value, you are increasing the risk of StackOverflowError, specially if you have deep recursion somewhere in your code or libraries that you use.

Other native memory OOMEs

There are many ways to use native memory other than thread stack allocation discussed before. Allocations can be made by the JVM itself, third party libraries, utilities in the JDK libraries,  and JNI code that you write. When the native memory gets exhausted, we may get an OOME of some kind. They are not consistent with their message, but one of them looks as follows:
java.lang.OutOfMemoryError: requested <bytes> bytes for <reason>. Out of swap space?
We can also cause a different message, by using the direct buffer feature of the java.nio package, in order to consume all available native memory:
List<ByteBuffer> list = new ArrayList<ByteBuffer>();
In this case we will get the message:
java.lang.OutOfMemoryError: Direct buffer memory
1)  Check whether there is a memory leak in some native code you use with JNI
2) Reduce the heap size using -Xmx and -Xms flags, to make more space for native memory.
3) Increase the swap space on disk. The way to do this depends on the operating system you use.


There are several types of OutOfMemoryError messages. It is important to know how to identify and analyze them,  and then how to find an appropriate solution. Solutions are in terms of prevention and not recovery at runtime.  A recovery attempt after an OOME may be too late – the data integrity may be broken, and there is no guarantee that there will be enough resources to complete it successfully. Instead, we should re-launch the application with a different setup.
An automatic recovery mechanism that occurs before the OOME can be effective, though. Since Java 1.5, we have the MemoryMXBean class, as part of the JMX services. An example of such a memory tracking mechanism was written by Dr. Heinz Kabutz in his 92nd issue of his newsletter.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: