Investigating Your RAM Usage

Because Android is designed for mobile devices, you should always be careful about how much random-access memory (RAM) your application uses. Although Dalvik and ART perform routine garbage collection (GC), this doesn’t mean you can ignore when and where your application allocates and releases memory. In order to provide a stable user experience that allows the system to quickly switch between apps, it is important that your application does not needlessly consume memory when the user is not interacting with it.

Even if you follow all the best practices for Managing Your App Memory during development (which you should), you still might leak objects or introduce other memory bugs. The only way to be certain your application is using as little memory as possible is to analyze your app’s memory usage with tools. This guide shows you how to do that.

Interpreting Log Messages

The simplest place to begin investigating your application’s memory usage is the runtime log messages. Sometimes when a GC occurs, a message is printed to logcat. The logcat output is also available in the Device Monitor or directly in IDEs such as Eclipse and Android Studio.

Dalvik Log Messages

In Dalvik (but not ART), every GC prints the following information to logcat:

D/dalvikvm: <GC_Reason> <Amount_freed>, <Heap_stats>, <External_memory_stats>, <Pause_time>

Example:

D/dalvikvm( 9050): GC_CONCURRENT freed 2049K, 65% free 3571K/9991K, external 4703K/5261K, paused 2ms+2ms

GC Reason

What triggered the GC and what kind of collection it is. Reasons that may appear include:

GC_CONCURRENT

A concurrent GC that frees up memory as your heap begins to fill up.

GC_FOR_MALLOC

A GC caused because your application attempted to allocate memory when your heap was already full, so the system had to stop your application and reclaim memory.

GC_HPROF_DUMP_HEAP

A GC that occurs when you request to create an HPROF file to analyze your heap.

GC_EXPLICIT

An explicit GC, such as when you call gc() (which you should avoid calling and instead trust the GC to run when needed).

GC_EXTERNAL_ALLOC

This happens only on API level 10 and lower (newer versions allocate everything in the Dalvik heap). A GC for externally allocated memory (such as the pixel data stored in native memory or NIO byte buffers).

Amount freed

The amount of memory reclaimed from this GC.

Heap stats

Percentage free of the heap and (number of live objects)/(total heap size).

External memory stats

Externally allocated memory on API level 10 and lower (amount of allocated memory) / (limit at which collection will occur).

Pause time

Larger heaps will have larger pause times. Concurrent pause times show two pauses: one at the beginning of the collection and another near the end.

As these log messages accumulate, look out for increases in the heap stats (the 3571K/9991K value in the above example). If this value continues to increase, you may have a memory leak.

ART Log Messages

Unlike Dalvik, ART doesn't log messqages for GCs that were not explicity requested. GCs are only printed when they are they are deemed slow. More precisely, if the GC pause exceeds than 5ms or the GC duration exceeds 100ms. If the application is not in a pause perceptible process state, then none of its GCs are deemed slow. Explicit GCs are always logged.

ART includes the following information in its garbage collection log messages:

I/art: <GC_Reason> <GC_Name> <Objects_freed>(<Size_freed>) AllocSpace Objects, <Large_objects_freed>(<Large_object_size_freed>) <Heap_stats> LOS objects, <Pause_time(s)>

Example:

I/art : Explicit concurrent mark sweep GC freed 104710(7MB) AllocSpace objects, 21(416KB) LOS objects, 33% free, 25MB/38MB, paused 1.230ms total 67.216ms

GC Reason

What triggered the GC and what kind of collection it is. Reasons that may appear include:

Concurrent

A concurrent GC which does not suspend application threads. This GC runs in a background thread and does not prevent allocations.

Alloc

The GC was initiated because your application attempted to allocate memory when your heap was already full. In this case, the garbage collection occurred in the allocating thread.

Explicit

The garbage collection was explicitly requested by an application, for instance, by calling gc() or gc(). As with Dalvik, in ART it is recommended that you trust the GC and avoid requesting explicit GCs if possible. Explicit GCs are discouraged since they block the allocating thread and unnecessarily was CPU cycles. Explicit GCs could also cause jank if they cause other threads to get preempted.

NativeAlloc

The collection was caused by native memory pressure from native allocations such as Bitmaps or RenderScript allocation objects.

CollectorTransition

The collection was caused by a heap transition; this is caused by switching the GC at run time. Collector transitions consist of copying all the objects from a free-list backed space to a bump pointer space (or visa versa). Currently collector transitions only occur when an application changes process states from a pause perceptible state to a non pause perceptible state (or visa versa) on low RAM devices.

HomogeneousSpaceCompact

Homogeneous space compaction is free-list space to free-list space compaction which usually occurs when an application is moved to a pause imperceptible process state. The main reasons for doing this are reducing RAM usage and defragmenting the heap.

DisableMovingGc

This is not a real GC reason, but a note that collection was blocked due to use of GetPrimitiveArrayCritical. while concurrent heap compaction is occuring. In general, the use of GetPrimitiveArrayCritical is strongly discouraged due to its restrictions on moving collectors.

HeapTrim

This is not a GC reason, but a note that collection was blocked until a heap trim finished.

GC Name

ART has various different GCs which can get run.

Concurrent mark sweep (CMS)

A whole heap collector which frees collects all spaces other than the image space.

Concurrent partial mark sweep

A mostly whole heap collector which collects all spaces other than the image and zygote spaces.

Concurrent sticky mark sweep

A generational collector which can only free objects allocated since the last GC. This garbage collection is run more often than a full or partial mark sweep since it is faster and has lower pauses.

Marksweep + semispace

A non concurrent, copying GC used for heap transitions as well as homogeneous space compaction (to defragement the heap).

Objects freed

The number of objects which were reclaimed from this GC from the non large object space.

Size freed

The number of bytes which were reclaimed from this GC from the non large object space.

Large objects freed

The number of object in the large object space which were reclaimed from this garbage collection.

Large object size freed

The number of bytes in the large object space which were reclaimed from this garbage collection.

Heap stats

Percentage free and (number of live objects)/(total heap size).

Pause times

In general pause times are proportional to the number of object references which were modified while the GC was running. Currently, the ART CMS GCs only has one pause, near the end of the GC. The moving GCs have a long pause which lasts for the majority of the GC duration.

If you are seeing a large amount of GCs in logcat, look for increases in the heap stats (the 25MB/38MB value in the above example). If this value continues to increase and doesn't ever seem to get smaller, you could have a memory leak. Alternatively, if you are seeing GC which are for the reason "Alloc", then you are already operating near your heap capacity and can expect OOM exceptios in the near future.

Viewing Heap Updates

To get a little information about what kind of memory your application is using and when, you can view real-time updates to your app's heap in the Device Monitor:

1. Open the Device Monitor.

   From your <sdk>/tools/ directory, launch the monitor tool.

2. In the Debug Monitor window, select your app's process from the list on the left.
3. Click Update Heap above the process list.
4. In the right-side panel, select the Heap tab.

The Heap view shows some basic stats about your heap memory usage, updated after every GC. To see the first update, click the Cause GC button.

Figure 1. The Device Monitor tool, showing the [1] Update Heap and [2] Cause GC buttons. The Heap tab on the right shows the heap results.

Continue interacting with your application to watch your heap allocation update with each garbage collection. This can help you identify which actions in your application are likely causing too much allocation and where you should try to reduce allocations and release resources.

Tracking Allocations

As you start narrowing down memory issues, you should also use the Allocation Tracker to get a better understanding of where your memory-hogging objects are allocated. The Allocation Tracker can be useful not only for looking at specific uses of memory, but also to analyze critical code paths in an application such as scrolling.

For example, tracking allocations when flinging a list in your application allows you to see all the allocations that need to be done for that behavior, what thread they are on, and where they came from. This is extremely valuable for tightening up these paths to reduce the work they need and improve the overall smoothness of the UI.

To use Allocation Tracker:

1. Open the Device Monitor.

   From your <sdk>/tools/ directory, launch the monitor tool.

2. In the DDMS window, select your app's process in the left-side panel.
3. In the right-side panel, select the Allocation Tracker tab.
4. Click Start Tracking.
5. Interact with your application to execute the code paths you want to analyze.
6. Click Get Allocations every time you want to update the list of allocations.

The list shows all recent allocations, currently limited by a 512-entry ring buffer. Click on a line to see the stack trace that led to the allocation. The trace shows you not only what type of object was allocated, but also in which thread, in which class, in which file and at which line.

Figure 2. The Device Monitor tool, showing recent application allocations and stack traces in the Allocation Tracker.

Note: You will always see some allocations from DdmVmInternal and else where that come from the allocation tracker itself.

Although it's not necessary (nor possible) to remove all allocations for your performance critical code paths, the allocation tracker can help you identify important issues in your code. For instance, some apps might create a new Paint object on every draw. Moving that object into a global member is a simple fix that helps improve performance.

Viewing Overall Memory Allocations

For further analysis, you may want to observe how your app's memory is divided between different types of RAM allocation with the following adb command:

adb shell dumpsys meminfo <package_name|pid> [-d]

The -d flag prints more info related to Dalvik and ART memory usage.

The output lists all of your app's current allocations, measured in kilobytes.

When inspecting this information, you should be familiar with the following types of allocation:

Private (Clean and Dirty) RAM

This is memory that is being used by only your process. This is the bulk of the RAM that the system can reclaim when your app’s process is destroyed. Generally, the most important portion of this is “private dirty” RAM, which is the most expensive because it is used by only your process and its contents exist only in RAM so can’t be paged to storage (because Android does not use swap). All Dalvik and native heap allocations you make will be private dirty RAM; Dalvik and native allocations you share with the Zygote process are shared dirty RAM.

Proportional Set Size (PSS)

This is a measurement of your app’s RAM use that takes into account sharing pages across processes. Any RAM pages that are unique to your process directly contribute to its PSS value, while pages that are shared with other processes contribute to the PSS value only in proportion to the amount of sharing. For example, a page that is shared between two processes will contribute half of its size to the PSS of each process.

A nice characteristic of the PSS measurement is that you can add up the PSS across all processes to determine the actual memory being used by all processes. This means PSS is a good measure for the actual RAM weight of a process and for comparison against the RAM use of other processes and the total available RAM.

For example, below is the the output for Map’s process on a Nexus 5 device. There is a lot of information here, but key points for discussion are listed below.

adb shell dumpsys meminfo com.google.android.apps.maps -d

Note: The information you see may vary slightly from what is shown here, as some details of the output differ across platform versions.

** MEMINFO in pid 18227 [com.google.android.apps.maps] **
                   Pss  Private  Private  Swapped     Heap     Heap     Heap
                 Total    Dirty    Clean    Dirty     Size    Alloc     Free
                ------   ------   ------   ------   ------   ------   ------
  Native Heap    10468    10408        0        0    20480    14462     6017
  Dalvik Heap    34340    33816        0        0    62436    53883     8553
 Dalvik Other      972      972        0        0
        Stack     1144     1144        0        0
      Gfx dev    35300    35300        0        0
    Other dev        5        0        4        0
     .so mmap     1943      504      188        0
    .apk mmap      598        0      136        0
    .ttf mmap      134        0       68        0
    .dex mmap     3908        0     3904        0
    .oat mmap     1344        0       56        0
    .art mmap     2037     1784       28        0
   Other mmap       30        4        0        0
   EGL mtrack    73072    73072        0        0
    GL mtrack    51044    51044        0        0
      Unknown      185      184        0        0
        TOTAL   216524   208232     4384        0    82916    68345    14570

 Dalvik Details
        .Heap     6568     6568        0        0
         .LOS    24771    24404        0        0
          .GC      500      500        0        0
    .JITCache      428      428        0        0
      .Zygote     1093      936        0        0
   .NonMoving     1908     1908        0        0
 .IndirectRef       44       44        0        0

 Objects
               Views:       90         ViewRootImpl:        1
         AppContexts:        4           Activities:        1
              Assets:        2        AssetManagers:        2
       Local Binders:       21        Proxy Binders:       28
       Parcel memory:       18         Parcel count:       74
    Death Recipients:        2      OpenSSL Sockets:        2

Here is an older dumpsys on Dalvik of the gmail app:

** MEMINFO in pid 9953 [com.google.android.gm] **
                 Pss     Pss  Shared Private  Shared Private    Heap    Heap    Heap
               Total   Clean   Dirty   Dirty   Clean   Clean    Size   Alloc    Free
              ------  ------  ------  ------  ------  ------  ------  ------  ------
  Native Heap      0       0       0       0       0       0    7800    7637(6)  126
  Dalvik Heap   5110(3)    0    4136    4988(3)    0       0    9168    8958(6)  210
 Dalvik Other   2850       0    2684    2772       0       0
        Stack     36       0       8      36       0       0
       Cursor    136       0       0     136       0       0
       Ashmem     12       0      28       0       0       0
    Other dev    380       0      24     376       0       4
     .so mmap   5443(5) 1996    2584    2664(5) 5788    1996(5)
    .apk mmap    235      32       0       0    1252      32
    .ttf mmap     36      12       0       0      88      12
    .dex mmap   3019(5) 2148       0       0    8936    2148(5)
   Other mmap    107       0       8       8     324      68
      Unknown   6994(4)    0     252    6992(4)    0       0
        TOTAL  24358(1) 4188    9724   17972(2)16388    4260(2)16968   16595     336

 Objects
               Views:    426         ViewRootImpl:        3(8)
         AppContexts:      6(7)        Activities:        2(7)
              Assets:      2        AssetManagers:        2
       Local Binders:     64        Proxy Binders:       34
    Death Recipients:      0
     OpenSSL Sockets:      1

 SQL
         MEMORY_USED:   1739
  PAGECACHE_OVERFLOW:   1164          MALLOC_SIZE:       62

Generally, you should be concerned with only the Pss Total and Private Dirty columns. In some cases, the Private Clean and Heap Alloc columns also offer interesting data. Here is some more information about the different memory allocations (the rows) you should observe:

Dalvik Heap

The RAM used by Dalvik allocations in your app. The Pss Total includes all Zygote allocations (weighted by their sharing across processes, as described in the PSS definition above). The Private Dirty number is the actual RAM committed to only your app’s heap, composed of your own allocations and any Zygote allocation pages that have been modified since forking your app’s process from Zygote.

Note: On newer platform versions that have the Dalvik Other section, the Pss Total and Private Dirty numbers for Dalvik Heap do not include Dalvik overhead such as the just-in-time compilation (JIT) and GC bookkeeping, whereas older versions list it all combined under Dalvik.

The Heap Alloc is the amount of memory that the Dalvik and native heap allocators keep track of for your app. This value is larger than Pss Total and Private Dirty because your process was forked from Zygote and it includes allocations that your process shares with all the others.

.so mmap and .dex mmap

The RAM being used for mmapped .so (native) and .dex (Dalvik or ART) code. The Pss Total number includes platform code shared across apps; the Private Clean is your app’s own code. Generally, the actual mapped size will be much larger—the RAM here is only what currently needs to be in RAM for code that has been executed by the app. However, the .so mmap has a large private dirty, which is due to fix-ups to the native code when it was loaded into its final address.

.oat mmap

This is the amount of RAM used by the code image which is based off of the preloaded classes which are commonly used by multiple apps. This image is shared across all apps and is unaffected by particular apps.

.art mmap

This is the amount of RAM used by the heap image which is based off of the preloaded classes which are commonly used by multiple apps. This image is shared across all apps and is unaffected by particular apps. Even though the ART image contains Object instances, it does not count towards your heap size.

.Heap (only with -d flag)

This is the amount of heap memory for your app. This excludes objects in the image and large object spaces, but includes the zygote space and non-moving space.

.LOS (only with -d flag)

This is the amount of RAM used by the ART large object space. This includes zygote large objects. Large objects are all primitive array allocations larger than 12KB.

.GC (only with -d flag)

This is the amount of internal GC accounting overhead for your app. There is not really any way to reduce this overhead.

.JITCache (only with -d flag)

This is the amount of memory used by the JIT data and code caches. Typically, this is zero since all of the apps will be compiled at installed time.

.Zygote (only with -d flag)

This is the amount of memory used by the zygote space. The zygote space is created during device startup and is never allocated into.

.NonMoving (only with -d flag)

This is the amount of RAM used by the ART non-moving space. The non-moving space contains special non-movable objects such as fields and methods. You can reduce this section by using fewer fields and methods in your app.

.IndirectRef (only with -d flag)

This is the amount of RAM used by the ART indirect reference tables. Usually this amount is small, but if it is too high, it may be possible to reduce it by reducing the number of local and global JNI references used.

Unknown

Any RAM pages that the system could not classify into one of the other more specific items. Currently, this contains mostly native allocations, which cannot be identified by the tool when collecting this data due to Address Space Layout Randomization (ASLR). As with the Dalvik heap, the Pss Total for Unknown takes into account sharing with Zygote, and Private Dirty is unknown RAM dedicated to only your app.

TOTAL

The total Proportional Set Size (PSS) RAM used by your process. This is the sum of all PSS fields above it. It indicates the overall memory weight of your process, which can be directly compared with other processes and the total available RAM.

The Private Dirty and Private Clean are the total allocations within your process, which are not shared with other processes. Together (especially Private Dirty), this is the amount of RAM that will be released back to the system when your process is destroyed. Dirty RAM is pages that have been modified and so must stay committed to RAM (because there is no swap); clean RAM is pages that have been mapped from a persistent file (such as code being executed) and so can be paged out if not used for a while.

ViewRootImpl

The number of root views that are active in your process. Each root view is associated with a window, so this can help you identify memory leaks involving dialogs or other windows.

AppContexts and Activities

The number of application Context and Activity objects that currently live in your process. This can be useful to quickly identify leaked Activity objects that can’t be garbage collected due to static references on them, which is common. These objects often have a lot of other allocations associated with them and so are a good way to track large memory leaks.

Note: A View or Drawable object also holds a reference to the Activity that it's from, so holding a View or Drawable object can also lead to your application leaking an Activity.

Capturing a Heap Dump

A heap dump is a snapshot of all the objects in your app's heap, stored in a binary format called HPROF. Your app's heap dump provides information about the overall state of your app's heap so you can track down problems you might have identified while viewing heap updates.

To retrieve your heap dump:

1. Open the Device Monitor.

   From your <sdk>/tools/ directory, launch the monitor tool.

2. In the DDMS window, select your app's process in the left-side panel.
3. Click Dump HPROF file, shown in figure 3.
4. In the window that appears, name your HPROF file, select the save location, then click Save.

Figure 3. The Device Monitor tool, showing the [1] Dump HPROF file button.

If you need to be more precise about when the dump is created, you can also create a heap dump at the critical point in your application code by calling dumpHprofData().

The heap dump is provided in a format that's similar to, but not identical to one from the Java HPROF tool. The major difference in an Android heap dump is due to the fact that there are a large number of allocations in the Zygote process. But because the Zygote allocations are shared across all application processes, they don’t matter very much to your own heap analysis.

To analyze your heap dump, you can use a standard tool like jhat or the Eclipse Memory Analyzer Tool (MAT). However, first you'll need to convert the HPROF file from Android's format to the J2SE HPROF format. You can do this using the hprof-conv tool provided in the <sdk>/platform-tools/ directory. Simply run the hprof-conv command with two arguments: the original HPROF file and the location to write the converted HPROF file. For example:

hprof-conv heap-original.hprof heap-converted.hprof

Note: If you're using the version of DDMS that's integrated into Eclipse, you do not need to perform the HPROF converstion—it performs the conversion by default.

You can now load the converted file in MAT or another heap analysis tool that understands the J2SE HPROF format.

When analyzing your heap, you should look for memory leaks caused by:

• Long-lived references to an Activity, Context, View, Drawable, and other objects that may hold a reference to the container Activity or Context.
• Non-static inner classes (such as a Runnable, which can hold the Activity instance).
• Caches that hold objects longer than necessary.

Using the Eclipse Memory Analyzer Tool

The Eclipse Memory Analyzer Tool (MAT) is just one tool that you can use to analyze your heap dump. It's also quite powerful so most of its capabilities are beyond the scope of this document, but here are a few tips to get you started.

Once you open your converted HPROF file in MAT, you'll see a pie chart in the Overview, showing what your largest objects are. Below this chart, are links to couple of useful features:

• The Histogram view shows a list of all classes and how many instances there are of each.

  You might want to use this view to find extra instances of classes for which you know there should be only a certain number. For example, a common source of leaks is additional instance of your Activity class, for which you should usually have only one instance at a time. To find a specific class instance, type the class name into the <Regex> field at the top of the list.

  When you find a class with too many instances, right-click it and select List objects > with incoming references. In the list that appears, you can determine where an instance is retained by right-clicking it and selecting Path To GC Roots > exclude weak references.

• The Dominator tree shows a list of objects organized by the amount of retained heap.

  What you should look for is anything that's retaining a portion of heap that's roughly equivalent to the memory size you observed leaking from the GC logs, heap updates, or allocation tracker.

  When you see something suspicious, right-click on the item and select Path To GC Roots > exclude weak references. This opens a new tab that traces the references to that object which is causing the alleged leak.

  Note: Most apps will show an instance of Resources near the top with a good chunk of heap, but this is usually expected when your application uses lots of resources from your res/ directory.

Figure 4. The Eclipse Memory Analyzer Tool (MAT), showing the Histogram view and a search for "MainActivity".

For more information about MAT, watch the Google I/O 2011 presentation, Memory management for Android apps, which includes a walkthrough using MAT beginning at about 21:10. Also refer to the Eclipse Memory Analyzer documentation.

Comparing heap dumps

You may find it useful to compare your app's heap state at two different points in time in order to inspect the changes in memory allocation. To compare two heap dumps using MAT:

1. Create two HPROF files as described above, in Capturing a Heap Dump.
2. Open the first HPROF file in MAT (File > Open Heap Dump).
3. In the Navigation History view (if not visible, select Window > Navigation History), right-click on Histogram and select Add to Compare Basket.
4. Open the second HPROF file and repeat steps 2 and 3.
5. Switch to the Compare Basket view and click Compare the Results (the red "!" icon in the top-right corner of the view).

Triggering Memory Leaks

While using the tools described above, you should aggressively stress your application code and try forcing memory leaks. One way to provoke memory leaks in your application is to let it run for a while before inspecting the heap. Leaks will trickle up to the top of the allocations in the heap. However, the smaller the leak, the longer you need to run the application in order to see it.

You can also trigger a memory leak in one of the following ways:

1. Rotate the device from portrait to landscape and back again multiple times while in different activity states. Rotating the device can often cause an application to leak an Activity, Context, or View object because the system recreates the Activity and if your application holds a reference to one of those objects somewhere else, the system can't garbage collect it.
2. Switch between your application and another application while in different activity states (navigate to the Home screen, then return to your app).

Tip: You can also perform the above steps by using the "monkey" test framework. For more information on running the monkey test framework, read the monkeyrunner documentation.