--- layout: default title: Identifying Memory Allocators ---

Identifying Memory Allocators

In its simplest mode, InMemProfiler tracks memory allocations and attributes these to the name of the class being allocated. This is nothing new – the same information can be derived by using a tool such as Netbeans Profiler on a live process or Eclipse Memory Analysis Tool to examine a heap dump.

However, it is often the case that the dominant class which is being allocated is the char[] or byte[] class. This raises the much more interesting question of where the memory is being allocated from. To answer this question you can’t use the Eclipse Memory Analysis Tool as the heapdump does not store information about where an object was allocated. You have to use a profiling tool such as the Netbeans Profiler.

Example

Consider the following dummy webapp:

package com.yourorg.app;

import org.logging.Logger;
import org.server.Server.Request;

public class App
{
  public static void handleRequest(Request req) {
    Logger.log("Time: " + System.currentTimeMillis() + ", got request!");
    // App logic here
    Storage.store(System.currentTimeMillis());
  }

  private static class Storage
  {
    public static void store(long currentTimeMillis) {
      Logger.log("Time: " + currentTimeMillis + ", store request!");
      // Storage logic here
    }
  }
}

This code is invoked by a webapp container (e.g. Tomcat) and writes a log every time it handles a request and every time it stores something.

Netbeans Profiler

Lets start by using Netbeans Profiler to take a look at what memory is being allocated.

The main tab of output is the per class summaries. This shows that char[] arrays are by far the most common allocation.

We can drill into this further by asking Netbeans Profiler to show the Allocation Stack Traces for char[].

The stack traces are shown in a common tree with the % column showing how important each stack trace element is as a fraction of the total number of bytes allocated for instances of this class.

The problem with this view is that you probably don’t care about the fact that java.utils.Arrays allocates char[] instances. You are much more likely to care about entries further up in the stack indicating which classes within the com.yourorg package are doing this allocation. Furthermore, the Netbeans Profiler shows the stack traces in a tree so by the time you find stack frames with classes in your package you can’t see a single figure for the amount of memory that your com.yourorg classes are allocating.

InMemProfiler

InMemProfiler in #trace mode can handle this case much better.

First try – no filtering

The first time we run with InMemProfiler we can try these options:

#trace#include-[C#outputlimit-5

This traces all char[] allocations and will output the top 5 traces and the top 5 allocating classes:

Allocated objects:
  93448:1107 - [C

    Allocation Sites:
      10400:100
        java.util.Arrays.copyOf(Arrays.java)
        java.lang.AbstractStringBuilder.expandCapacity(AbstractStringBuilder.java)
        java.lang.AbstractStringBuilder.append(AbstractStringBuilder.java)
        java.lang.StringBuilder.append(StringBuilder.java)
        com.yourorg.app.App$Storage.store(App.java)
        com.yourorg.app.App.handleRequest(App.java)
        org.server.Server$ServerThread.run(Server.java)
        java.lang.Thread.run(Thread.java)
      10400:100
        java.util.Arrays.copyOf(Arrays.java)
        java.lang.AbstractStringBuilder.expandCapacity(AbstractStringBuilder.java)
        java.lang.AbstractStringBuilder.append(AbstractStringBuilder.java)
        java.lang.StringBuilder.append(StringBuilder.java)
        org.server.Server$ServerThread.run(Server.java)
        java.lang.Thread.run(Thread.java)
      10400:100
        java.util.Arrays.copyOf(Arrays.java)
        java.lang.AbstractStringBuilder.expandCapacity(AbstractStringBuilder.java)
        java.lang.AbstractStringBuilder.append(AbstractStringBuilder.java)
        java.lang.StringBuilder.append(StringBuilder.java)
        com.yourorg.app.App.handleRequest(App.java)
        org.server.Server$ServerThread.run(Server.java)
        java.lang.Thread.run(Thread.java)
      8800:100
        java.util.Arrays.copyOfRange(Arrays.java)
        java.lang.String.(String.java)
        java.lang.StringBuilder.toString(StringBuilder.java)
        com.yourorg.app.App$Storage.store(App.java)
        com.yourorg.app.App.handleRequest(App.java)
        org.server.Server$ServerThread.run(Server.java)
        java.lang.Thread.run(Thread.java)
      8800:100
        java.util.Arrays.copyOfRange(Arrays.java)
        java.lang.String.(String.java)
        java.lang.StringBuilder.toString(StringBuilder.java)
        org.server.Server$ServerThread.run(Server.java)
        java.lang.Thread.run(Thread.java)

    Allocation Classes:
      java.lang.StringBuilder
        83608:1001 - All methods
          36824:343 - append
          28368:329 - toString
          18416:329 - 
      java.lang.Thread
        82808:994 - All methods
          82552:988 - run
          256:6 - 
          176:4 - init
      org.server.Server$ServerThread
        82552:988 - All methods
          82552:988 - run
      java.util.Arrays
        67688:694 - All methods
          37792:349 - copyOf
          29896:345 - copyOfRange
      java.lang.AbstractStringBuilder
        56328:682 - All methods
          36824:343 - expandCapacity
          36824:343 - append
          19504:339 - 

This output isn’t great – note that java.lang.StringBuilder is the top class as it is common to all of the string construction traces. This isn’t a very interesting finding as StringBuilder is used to construct Strings throughout Java code.

Second try – ignore library classes

This time we use the following options

#trace#include-[C#outputlimit-5#traceignore-java

This only affects the "Allocation Classes" part of the output and will ignore all "Allocation Classes" which begin with the string "java".

    Allocation Classes:
      org.server.Server$ServerThread
        82552:988 - All methods
          82552:988 - run
      com.yourorg.app.App
        51056:620 - All methods
          51056:620 - handleRequest
      com.yourorg.app.App$Storage
        24800:300 - All methods
          24800:300 - store
      sun.misc.Launcher$AppClassLoader
        19016:199 - All methods
          19016:199 - loadClass
          4456:58 - getPermissions
      sun.misc.URLClassPath
        11120:93 - All methods
          11120:93 - getResource
          1088:9 - getLoader

This is better but now the ServerThread is being blamed for all of the allocations so we can refine the #traceignore settings further:

#trace#include-[C#outputlimit-5#traceignore-java,sun,org.server,org.logging

This gives the following output:

    Allocation Classes:
      com.yourorg.app.App
        51056:620 - All methods
          51056:620 - handleRequest
      com.yourorg.app.App$Storage
        24800:300 - All methods
          24800:300 - store

This is much better! InMemProfiler can now identify exactly two classes and show the memory allocation.

However, looking at the code it is a little surprising that App is being blamed for almost twice as much allocation as App$Storage. This is actually a red herring and show up because InMemProfiler attributes each memory allocation to every non ignored class in the stack trace that allocated it. This means that all the allocations within App$Storage are also attributed to App.

Third try – target application classes

The output can be improved a little further by using #tracetarget to let InMemProfiler know where in the stack to stop attributing memory allocations to classes.

#trace#include-[C#outputlimit-5#traceignore-java,sun,org.server,org.logging#tracetarget-com.yourorg

Once again, this only affects the "Allocation Classes" part of the output. InMemProfiler will ignore all "Allocation Classes" further up in a stack once it finds a single frame beginning with "com.yourorg".

    Allocation Classes:
      com.yourorg.app.App
        26256:320 - All methods
          26256:320 - handleRequest
      com.yourorg.app.App$Storage
        24800:300 - All methods
          24800:300 - store

This final output is now accurate and as useful as we can make it. This clearly shows how many bytes are allocated for char[] by each of these two classes. It should be noted that the higher char[] allocation count for the App class is due to char[]s allocated during classloading of the classes which App depends on.