Designing for Performance

An Android application will run on a mobile device with limited computing power and storage, and constrained battery life. Because of this, it should be efficient. Battery life is one reason you might want to optimize your app even if it already seems to run "fast enough". Battery life is important to users, and Android's battery usage breakdown means users will know if your app is responsible draining their battery.

Note that although this document primarily covers micro-optimizations, these will almost never make or break your software. Choosing the right algorithms and data structures should always be your priority, but is outside the scope of this document.

Introduction

There are two basic rules for writing efficient code:

  • Don't do work that you don't need to do.
  • Don't allocate memory if you can avoid it.

Optimize Judiciously

This document is about Android-specific micro-optimization, so it assumes that you've already used profiling to work out exactly what code needs to be optimized, and that you already have a way to measure the effect (good or bad) of any changes you make. You only have so much engineering time to invest, so it's important to know you're spending it wisely.

(See Closing Notes for more on profiling and writing effective benchmarks.)

This document also assumes that you made the best decisions about data structures and algorithms, and that you've also considered the future performance consequences of your API decisions. Using the right data structures and algorithms will make more difference than any of the advice here, and considering the performance consequences of your API decisions will make it easier to switch to better implementations later (this is more important for library code than for application code).

(If you need that kind of advice, see Josh Bloch's Effective Java, item 47.)

One of the trickiest problems you'll face when micro-optimizing an Android app is that your app is pretty much guaranteed to be running on multiple hardware platforms. Different versions of the VM running on different processors running at different speeds. It's not even generally the case that you can simply say "device X is a factor F faster/slower than device Y", and scale your results from one device to others. In particular, measurement on the emulator tells you very little about performance on any device. There are also huge differences between devices with and without a JIT: the "best" code for a device with a JIT is not always the best code for a device without.

If you want to know how your app performs on a given device, you need to test on that device.

Avoid Creating Unnecessary Objects

Object creation is never free. A generational GC with per-thread allocation pools for temporary objects can make allocation cheaper, but allocating memory is always more expensive than not allocating memory.

If you allocate objects in a user interface loop, you will force a periodic garbage collection, creating little "hiccups" in the user experience. The concurrent collector introduced in Gingerbread helps, but unnecessary work should always be avoided.

Thus, you should avoid creating object instances you don't need to. Some examples of things that can help:

  • If you have a method returning a string, and you know that its result will always be appended to a StringBuffer anyway, change your signature and implementation so that the function does the append directly, instead of creating a short-lived temporary object.
  • When extracting strings from a set of input data, try to return a substring of the original data, instead of creating a copy. You will create a new String object, but it will share the char[] with the data. (The trade-off being that if you're only using a small part of the original input, you'll be keeping it all around in memory anyway if you go this route.)

A somewhat more radical idea is to slice up multidimensional arrays into parallel single one-dimension arrays:

  • An array of ints is a much better than an array of Integers, but this also generalizes to the fact that two parallel arrays of ints are also a lot more efficient than an array of (int,int) objects. The same goes for any combination of primitive types.
  • If you need to implement a container that stores tuples of (Foo,Bar) objects, try to remember that two parallel Foo[] and Bar[] arrays are generally much better than a single array of custom (Foo,Bar) objects. (The exception to this, of course, is when you're designing an API for other code to access; in those cases, it's usually better to trade good API design for a small hit in speed. But in your own internal code, you should try and be as efficient as possible.)

Generally speaking, avoid creating short-term temporary objects if you can. Fewer objects created mean less-frequent garbage collection, which has a direct impact on user experience.

Performance Myths

Previous versions of this document made various misleading claims. We address some of them here.

On devices without a JIT, it is true that invoking methods via a variable with an exact type rather than an interface is slightly more efficient. (So, for example, it was cheaper to invoke methods on a HashMap map than a Map map, even though in both cases the map was a HashMap.) It was not the case that this was 2x slower; the actual difference was more like 6% slower. Furthermore, the JIT makes the two effectively indistinguishable.

On devices without a JIT, caching field accesses is about 20% faster than repeatedly accesssing the field. With a JIT, field access costs about the same as local access, so this isn't a worthwhile optimization unless you feel it makes your code easier to read. (This is true of final, static, and static final fields too.)

Prefer Static Over Virtual

If you don't need to access an object's fields, make your method static. Invocations will be about 15%-20% faster. It's also good practice, because you can tell from the method signature that calling the method can't alter the object's state.

Avoid Internal Getters/Setters

In native languages like C++ it's common practice to use getters (e.g. i = getCount()) instead of accessing the field directly (i = mCount). This is an excellent habit for C++, because the compiler can usually inline the access, and if you need to restrict or debug field access you can add the code at any time.

On Android, this is a bad idea. Virtual method calls are expensive, much more so than instance field lookups. It's reasonable to follow common object-oriented programming practices and have getters and setters in the public interface, but within a class you should always access fields directly.

Without a JIT, direct field access is about 3x faster than invoking a trivial getter. With the JIT (where direct field access is as cheap as accessing a local), direct field access is about 7x faster than invoking a trivial getter. This is true in Froyo, but will improve in the future when the JIT inlines getter methods.

Use Static Final For Constants

Consider the following declaration at the top of a class:

static int intVal = 42;
static String strVal = "Hello, world!";

The compiler generates a class initializer method, called <clinit>, that is executed when the class is first used. The method stores the value 42 into intVal, and extracts a reference from the classfile string constant table for strVal. When these values are referenced later on, they are accessed with field lookups.

We can improve matters with the "final" keyword:

static final int intVal = 42;
static final String strVal = "Hello, world!";

The class no longer requires a <clinit> method, because the constants go into static field initializers in the dex file. Code that refers to intVal will use the integer value 42 directly, and accesses to strVal will use a relatively inexpensive "string constant" instruction instead of a field lookup. (Note that this optimization only applies to primitive types and String constants, not arbitrary reference types. Still, it's good practice to declare constants static final whenever possible.)

Use Enhanced For Loop Syntax

The enhanced for loop (also sometimes known as "for-each" loop) can be used for collections that implement the Iterable interface and for arrays. With collections, an iterator is allocated to make interface calls to hasNext() and next(). With an ArrayList, a hand-written counted loop is about 3x faster (with or without JIT), but for other collections the enhanced for loop syntax will be exactly equivalent to explicit iterator usage.

There are several alternatives for iterating through an array:

    static class Foo {
        int mSplat;
    }
    Foo[] mArray = ...

    public void zero() {
        int sum = 0;
        for (int i = 0; i < mArray.length; ++i) {
            sum += mArray[i].mSplat;
        }
    }

    public void one() {
        int sum = 0;
        Foo[] localArray = mArray;
        int len = localArray.length;

        for (int i = 0; i < len; ++i) {
            sum += localArray[i].mSplat;
        }
    }

    public void two() {
        int sum = 0;
        for (Foo a : mArray) {
            sum += a.mSplat;
        }
    }

zero() is slowest, because the JIT can't yet optimize away the cost of getting the array length once for every iteration through the loop.

one() is faster. It pulls everything out into local variables, avoiding the lookups. Only the array length offers a performance benefit.

two() is fastest for devices without a JIT, and indistinguishable from one() for devices with a JIT. It uses the enhanced for loop syntax introduced in version 1.5 of the Java programming language.

To summarize: use the enhanced for loop by default, but consider a hand-written counted loop for performance-critical ArrayList iteration.

(See also Effective Java item 46.)

Consider Package Instead of Private Access with Private Inner Classes

Consider the following class definition:

public class Foo {
    private class Inner {
        void stuff() {
            Foo.this.doStuff(Foo.this.mValue);
        }
    }

    private int mValue;

    public void run() {
        Inner in = new Inner();
        mValue = 27;
        in.stuff();
    }

    private void doStuff(int value) {
        System.out.println("Value is " + value);
    }
}

The key things to note here are that we define a private inner class (Foo$Inner) that directly accesses a private method and a private instance field in the outer class. This is legal, and the code prints "Value is 27" as expected.

The problem is that the VM considers direct access to Foo's private members from Foo$Inner to be illegal because Foo and Foo$Inner are different classes, even though the Java language allows an inner class to access an outer class' private members. To bridge the gap, the compiler generates a couple of synthetic methods:

/*package*/ static int Foo.access$100(Foo foo) {
    return foo.mValue;
}
/*package*/ static void Foo.access$200(Foo foo, int value) {
    foo.doStuff(value);
}

The inner class code calls these static methods whenever it needs to access the mValue field or invoke the doStuff method in the outer class. What this means is that the code above really boils down to a case where you're accessing member fields through accessor methods. Earlier we talked about how accessors are slower than direct field accesses, so this is an example of a certain language idiom resulting in an "invisible" performance hit.

If you're using code like this in a performance hotspot, you can avoid the overhead by declaring fields and methods accessed by inner classes to have package access, rather than private access. Unfortunately this means the fields can be accessed directly by other classes in the same package, so you shouldn't use this in public API.

Use Floating-Point Judiciously

As a rule of thumb, floating-point is about 2x slower than integer on Android devices. This is true on a FPU-less, JIT-less G1 and a Nexus One with an FPU and the JIT. (Of course, absolute speed difference between those two devices is about 10x for arithmetic operations.)

In speed terms, there's no difference between float and double on the more modern hardware. Space-wise, double is 2x larger. As with desktop machines, assuming space isn't an issue, you should prefer double to float.

Also, even for integers, some chips have hardware multiply but lack hardware divide. In such cases, integer division and modulus operations are performed in software — something to think about if you're designing a hash table or doing lots of math.

Know And Use The Libraries

In addition to all the usual reasons to prefer library code over rolling your own, bear in mind that the system is at liberty to replace calls to library methods with hand-coded assembler, which may be better than the best code the JIT can produce for the equivalent Java. The typical example here is String.indexOf and friends, which Dalvik replaces with an inlined intrinsic. Similarly, the System.arraycopy method is about 9x faster than a hand-coded loop on a Nexus One with the JIT.

(See also Effective Java item 47.)

Use Native Methods Judiciously

Native code isn't necessarily more efficient than Java. For one thing, there's a cost associated with the Java-native transition, and the JIT can't optimize across these boundaries. If you're allocating native resources (memory on the native heap, file descriptors, or whatever), it can be significantly more difficult to arrange timely collection of these resources. You also need to compile your code for each architecture you wish to run on (rather than rely on it having a JIT). You may even have to compile multiple versions for what you consider the same architecture: native code compiled for the ARM processor in the G1 can't take full advantage of the ARM in the Nexus One, and code compiled for the ARM in the Nexus One won't run on the ARM in the G1.

Native code is primarily useful when you have an existing native codebase that you want to port to Android, not for "speeding up" parts of a Java app.

If you do need to use native code, you should read our JNI Tips.

(See also Effective Java item 54.)

Closing Notes

One last thing: always measure. Before you start optimizing, make sure you have a problem. Make sure you can accurately measure your existing performance, or you won't be able to measure the benefit of the alternatives you try.

Every claim made in this document is backed up by a benchmark. The source to these benchmarks can be found in the code.google.com "dalvik" project.

The benchmarks are built with the Caliper microbenchmarking framework for Java. Microbenchmarks are hard to get right, so Caliper goes out of its way to do the hard work for you, and even detect some cases where you're not measuring what you think you're measuring (because, say, the VM has managed to optimize all your code away). We highly recommend you use Caliper to run your own microbenchmarks.

You may also find Traceview useful for profiling, but it's important to realize that it currently disables the JIT, which may cause it to misattribute time to code that the JIT may be able to win back. It's especially important after making changes suggested by Traceview data to ensure that the resulting code actually runs faster when run without Traceview.

↑ Go to top