Random-access memory (RAM) is a valuable resource in any software development environment, but it's even more valuable on a mobile operating system where physical memory is often constrained. Although Android's Dalvik virtual machine performs routine garbage collection, this doesn't allow you to ignore when and where your app allocates and releases memory.
In order for the garbage collector to reclaim memory from your app, you need to avoid
introducing memory leaks (usually caused by holding onto object references in global members) and
Reference objects at the appropriate time (as defined by
lifecycle callbacks discussed further below). For most apps, the Dalvik garbage collector takes
care of the rest: the system reclaims your memory allocations when the corresponding objects leave
the scope of your app's active threads.
This document explains how Android manages app processes and memory allocation, and how you can proactively reduce memory usage while developing for Android. For more information about general practices to clean up your resources when programming in Java, refer to other books or online documentation about managing resource references. If you’re looking for information about how to analyze your app’s memory once you’ve already built it, read Investigating Your RAM Usage.
Android does not offer swap space for memory, but it does use paging and memory-mapping (mmapping) to manage memory. This means that any memory you modify—whether by allocating new objects or touching mmapped pages—remains resident in RAM and cannot be paged out. So the only way to completely release memory from your app is to release object references you may be holding, making the memory available to the garbage collector. That is with one exception: any files mmapped in without modification, such as code, can be paged out of RAM if the system wants to use that memory elsewhere.
In order to fit everything it needs in RAM, Android tries to share RAM pages across processes. It can do so in the following ways:
.odexfile for direct mmapping), app resources (by designing the resource table to be a structure that can be mmapped and by aligning the zip entries of the APK), and traditional project elements like native code in
Due to the extensive use of shared memory, determining how much memory your app is using requires care. Techniques to properly determine your app's memory use are discussed in Investigating Your RAM Usage.
Here are some facts about how Android allocates then reclaims memory from your app:
To maintain a functional multi-tasking environment, Android sets a hard limit on the heap size
for each app. The exact heap size limit varies between devices based on how much RAM the device
has available overall. If your app has reached the heap capacity and tries to allocate more
memory, it will receive an
In some cases, you might want to query the system to determine exactly how much heap space you
have available on the current device—for example, to determine how much data is safe to keep in a
cache. You can query the system for this figure by calling
getMemoryClass(). This returns an integer indicating the number of
megabytes available for your app's heap. This is discussed further below, under
Check how much memory you should use.
Instead of using swap space when the user switches between apps, Android keeps processes that are not hosting a foreground ("user visible") app component in a least-recently used (LRU) cache. For example, when the user first launches an app, a process is created for it, but when the user leaves the app, that process does not quit. The system keeps the process cached, so if the user later returns to the app, the process is reused for faster app switching.
If your app has a cached process and it retains memory that it currently does not need, then your app—even while the user is not using it—is constraining the system's overall performance. So, as the system runs low on memory, it may kill processes in the LRU cache beginning with the process least recently used, but also giving some consideration toward which processes are most memory intensive. To keep your process cached as long as possible, follow the advice in the following sections about when to release your references.
More information about how processes are cached while not running in the foreground and how Android decides which ones can be killed is available in the Processes and Threads guide.
You should consider RAM constraints throughout all phases of development, including during app design (before you begin development). There are many ways you can design and write code that lead to more efficient results, through aggregation of the same techniques applied over and over.
You should apply the following techniques while designing and implementing your app to make it more memory efficient.
If your app needs a service to perform work in the background, do not keep it running unless it's actively performing a job. Also be careful to never leak your service by failing to stop it when its work is done.
When you start a service, the system prefers to always keep the process for that service running. This makes the process very expensive because the RAM used by the service can’t be used by anything else or paged out. This reduces the number of cached processes that the system can keep in the LRU cache, making app switching less efficient. It can even lead to thrashing in the system when memory is tight and the system can’t maintain enough processes to host all the services currently running.
The best way to limit the lifespan of your service is to use an
IntentService, which finishes
itself as soon as it's done handling the intent that started it. For more information, read
Running in a Background Service
Leaving a service running when it’s not needed is one of the worst memory-management mistakes an Android app can make. So don’t be greedy by keeping a service for your app running. Not only will it increase the risk of your app performing poorly due to RAM constraints, but users will discover such misbehaving apps and uninstall them.
When the user navigates to a different app and your UI is no longer visible, you should release any resources that are used by only your UI. Releasing UI resources at this time can significantly increase the system's capacity for cached processes, which has a direct impact on the quality of the user experience.
To be notified when the user exits your UI, implement the
onTrimMemory() callback in your
Activity classes. You should use this
method to listen for the
which indicates your UI is now hidden from view and you should free resources that only your UI
Notice that your app receives the
onTrimMemory() callback with
only when all the UI components of your app process become hidden from the user.
This is distinct
onStop() callback, which is called when an
Activity instance becomes hidden, which occurs even when the user moves to
another activity in your app. So although you should implement
onStop() to release activity resources such as a network connection or to unregister broadcast
receivers, you usually should not release your UI resources until you receive
onTrimMemory(TRIM_MEMORY_UI_HIDDEN). This ensures
that if the user navigates back from another activity in your app, your UI resources are
still available to resume the activity quickly.
During any stage of your app's lifecycle, the
onTrimMemory() callback also tells you when
the overall device memory is getting low. You should respond by further releasing resources based
on the following memory levels delivered by
Your app is running and not considered killable, but the device is running low on memory and the system is actively killing processes in the LRU cache.
Your app is running and not considered killable, but the device is running much lower on memory so you should release unused resources to improve system performance (which directly impacts your app's performance).
Your app is still running, but the system has already killed most of the processes in the LRU cache, so you should release all non-critical resources now. If the system cannot reclaim sufficient amounts of RAM, it will clear all of the LRU cache and begin killing processes that the system prefers to keep alive, such as those hosting a running service.
Also, when your app process is currently cached, you may receive one of the following
The system is running low on memory and your process is near the beginning of the LRU list. Although your app process is not at a high risk of being killed, the system may already be killing processes in the LRU cache. You should release resources that are easy to recover so your process will remain in the list and resume quickly when the user returns to your app.
The system is running low on memory and your process is near the middle of the LRU list. If the system becomes further constrained for memory, there's a chance your process will be killed.
The system is running low on memory and your process is one of the first to be killed if the system does not recover memory now. You should release everything that's not critical to resuming your app state.
Note: When the system begins killing processes in the LRU cache, although it primarily works bottom-up, it does give some consideration to which processes are consuming more memory and will thus provide the system more memory gain if killed. So the less memory you consume while in the LRU list overall, the better your chances are to remain in the list and be able to quickly resume.
As mentioned earlier, each Android-powered device has a different amount of RAM available to the
system and thus provides a different heap limit for each app. You can call
getMemoryClass() to get an estimate of your app's available heap in
megabytes. If your app tries to allocate more memory than is available here, it will receive an
In very special situations, you can request a larger heap size by setting the
attribute to "true" in the manifest
tag. If you do so, you can call
getLargeMemoryClass() to get an estimate of the large heap size.
However, the ability to request a large heap is intended only for a small set of apps that can justify the need to consume more RAM (such as a large photo editing app). Never request a large heap simply because you've run out of memory and you need a quick fix—you should use it only when you know exactly where all your memory is being allocated and why it must be retained. Yet, even when you're confident your app can justify the large heap, you should avoid requesting it to whatever extent possible. Using the extra memory will increasingly be to the detriment of the overall user experience because garbage collection will take longer and system performance may be slower when task switching or performing other common operations.
Additionally, the large heap size is not the same on all devices and, when running on
devices that have limited RAM, the large heap size may be exactly the same as the regular heap
size. So even if you do request the large heap size, you should call
getMemoryClass() to check the regular heap size and strive to always
stay below that limit.
When you load a bitmap, keep it in RAM only at the resolution you need for the current device's screen, scaling it down if the original bitmap is a higher resolution. Keep in mind that an increase in bitmap resolution results in a corresponding (increase2) in memory needed, because both the X and Y dimensions increase.
Note: On Android 2.3.x (API level 10) and below, bitmap objects always appear as the same size in your app heap regardless of the image resolution (the actual pixel data is stored separately in native memory). This makes it more difficult to debug the bitmap memory allocation because most heap analysis tools do not see the native allocation. However, beginning in Android 3.0 (API level 11), the bitmap pixel data is allocated in your app's Dalvik heap, improving garbage collection and debuggability. So if your app uses bitmaps and you're having trouble discovering why your app is using some memory on an older device, switch to a device running Android 3.0 or higher to debug it.
For more tips about working with bitmaps, read Managing Bitmap Memory.
Take advantage of optimized containers in the Android framework, such as
LongSparseArray. The generic
implementation can be quite memory
inefficient because it needs a separate entry object for every mapping. Additionally, the
SparseArray classes are more efficient because they avoid the system's need
the key and sometimes value (which creates yet another object or two per entry). And don't be
afraid of dropping down to raw arrays when that makes sense.
Be knowledgeable about the cost and overhead of the language and libraries you are using, and keep this information in mind when you design your app, from start to finish. Often, things on the surface that look innocuous may in fact have a large amount of overhead. Examples include:
HashMaprequires the allocation of an additional entry object that takes 32 bytes (see the previous section about optimized data containers).
A few bytes here and there quickly add up—app designs that are class- or object-heavy will suffer from this overhead. That can leave you in the difficult position of looking at a heap analysis and realizing your problem is a lot of small objects using up your RAM.
Often, developers use abstractions simply as a "good programming practice," because abstractions can improve code flexibility and maintenance. However, abstractions come at a significant cost: generally they require a fair amount more code that needs to be executed, requiring more time and more RAM for that code to be mapped into memory. So if your abstractions aren't supplying a significant benefit, you should avoid them.
Protocol buffers are a language-neutral, platform-neutral, extensible mechanism designed by Google for serializing structured data—think XML, but smaller, faster, and simpler. If you decide to use protobufs for your data, you should always use nano protobufs in your client-side code. Regular protobufs generate extremely verbose code, which will cause many kinds of problems in your app: increased RAM use, significant APK size increase, slower execution, and quickly hitting the DEX symbol limit.
For more information, see the "Nano version" section in the protobuf readme.
Using a dependency injection framework such as Guice or RoboGuice may be attractive because they can simplify the code you write and provide an adaptive environment that's useful for testing and other configuration changes. However, these frameworks tend to perform a lot of process initialization by scanning your code for annotations, which can require significant amounts of your code to be mapped into RAM even though you don't need it. These mapped pages are allocated into clean memory so Android can drop them, but that won't happen until the pages have been left in memory for a long period of time.
External library code is often not written for mobile environments and can be inefficient when used for work on a mobile client. At the very least, when you decide to use an external library, you should assume you are taking on a significant porting and maintenance burden to optimize the library for mobile. Plan for that work up-front and analyze the library in terms of code size and RAM footprint before deciding to use it at all.
Even libraries supposedly designed for use on Android are potentially dangerous because each
library may do things differently. For example, one library may use nano protobufs while another
uses micro protobufs. Now you have two different protobuf implementations in your app. This can and
will also happen with different implementations of logging, analytics, image loading frameworks,
caching, and all kinds of other things you don't expect. ProGuard won't save you here because these
will all be lower-level dependencies that are required by the features for which you want the
library. This becomes especially problematic when you use an
subclass from a library (which
will tend to have wide swaths of dependencies), when libraries use reflection (which is common and
means you need to spend a lot of time manually tweaking ProGuard to get it to work), and so on.
Also be careful not to fall into the trap of using a shared library for one or two features out of dozens of other things it does; you don't want to pull in a large amount of code and overhead that you don't even use. At the end of the day, if there isn't an existing implementation that is a strong match for what you need to do, it may be best if you create your own implementation.
A variety of information about optimizing your app's overall performance is available in other documents listed in Best Practices for Performance. Many of these documents include optimizations tips for CPU performance, but many of these tips also help optimize your app's memory use, such as by reducing the number of layout objects required by your UI.
The ProGuard tool shrinks, optimizes, and obfuscates your code by removing unused code and renaming classes, fields, and methods with semantically obscure names. Using ProGuard can make your code more compact, requiring fewer RAM pages to be mapped.
If you do any post-processing of an APK generated by a build system (including signing it with your final production certificate), then you must run zipalign on it to have it re-aligned. Failing to do so can cause your app to require significantly more RAM, because things like resources can no longer be mmapped from the APK.
Note: Google Play Store does not accept APK files that are not zipaligned.
Once you achieve a relatively stable build, begin analyzing how much RAM your app is using throughout all stages of its lifecycle. For information about how to analyze your app, read Investigating Your RAM Usage.
If it's appropriate for your app, an advanced technique that may help you manage your app's memory is dividing components of your app into multiple processes. This technique must always be used carefully and most apps should not run multiple processes, as it can easily increase—rather than decrease—your RAM footprint if done incorrectly. It is primarily useful to apps that may run significant work in the background as well as the foreground and can manage those operations separately.
An example of when multiple processes may be appropriate is when building a music player that plays music from a service for long period of time. If the entire app runs in one process, then many of the allocations performed for its activity UI must be kept around as long as it is playing music, even if the user is currently in another app and the service is controlling the playback. An app like this may be split into two process: one for its UI, and the other for the work that continues running in the background service.
You can specify a separate process for each app component by declaring the
for each component in the manifest file. For example, you can specify that your service should run
in a process separate from your app's main process by declaring a new process named "background"
(but you can name the process anything you like):
<service android:name=".PlaybackService" android:process=":background" />
Your process name should begin with a colon (':') to ensure that the process remains private to your app.
Before you decide to create a new process, you need to understand the memory implications. To illustrate the consequences of each process, consider that an empty process doing basically nothing has an extra memory footprint of about 1.4MB, as shown by the memory information dump below.
adb shell dumpsys meminfo com.example.android.apis:empty ** MEMINFO in pid 10172 [com.example.android.apis:empty] ** 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 1864 1800 63 Dalvik Heap 764 0 5228 316 0 0 5584 5499 85 Dalvik Other 619 0 3784 448 0 0 Stack 28 0 8 28 0 0 Other dev 4 0 12 0 0 4 .so mmap 287 0 2840 212 972 0 .apk mmap 54 0 0 0 136 0 .dex mmap 250 148 0 0 3704 148 Other mmap 8 0 8 8 20 0 Unknown 403 0 600 380 0 0 TOTAL 2417 148 12480 1392 4832 152 7448 7299 148
Note: More information about how to read this output is provided in Investigating Your RAM Usage. The key data here is the Private Dirty and Private Clean memory, which shows that this process is using almost 1.4MB of non-pageable RAM (distributed across the Dalvik heap, native allocations, book-keeping, and library-loading), and another 150K of RAM for code that has been mapped in to execute.
This memory footprint for an empty process is fairly significant and it can quickly grow as you start doing work in that process. For example, here is the memory use of a process that is created only to show an activity with some text in it:
** MEMINFO in pid 10226 [com.example.android.helloactivity] ** 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 3000 2951 48 Dalvik Heap 1074 0 4928 776 0 0 5744 5658 86 Dalvik Other 802 0 3612 664 0 0 Stack 28 0 8 28 0 0 Ashmem 6 0 16 0 0 0 Other dev 108 0 24 104 0 4 .so mmap 2166 0 2824 1828 3756 0 .apk mmap 48 0 0 0 632 0 .ttf mmap 3 0 0 0 24 0 .dex mmap 292 4 0 0 5672 4 Other mmap 10 0 8 8 68 0 Unknown 632 0 412 624 0 0 TOTAL 5169 4 11832 4032 10152 8 8744 8609 134
The process has now almost tripled in size, to 4MB, simply by showing some text in the UI. This leads to an important conclusion: If you are going to split your app into multiple processes, only one process should be responsible for UI. Other processes should avoid any UI, as this will quickly increase the RAM required by the process (especially once you start loading bitmap assets and other resources). It may then be hard or impossible to reduce the memory usage once the UI is drawn.
Additionally, when running more than one process, it's more important than ever that you keep your code as lean as possible, because any unnecessary RAM overhead for common implementations are now replicated in each process. For example, if you are using enums (though you should not use enums), all of the RAM needed to create and initialize those constants is duplicated in each process, and any abstractions you have with adapters and temporaries or other overhead will likewise be replicated.
Another concern with multiple processes is the dependencies that exist between them. For example, if your app has a content provider that you have running in the default process which also hosts your UI, then code in a background process that uses that content provider will also require that your UI process remain in RAM. If your goal is to have a background process that can run independently of a heavy-weight UI process, it can't have dependencies on content providers or services that execute in the UI process.