Pury — New Way to Profile Your Android Application

Nikita Kozlov
Sep 22, 2016 · 8 min read

Applications are all about helping users to do what they need to, while providing best User Experience possible. Significant part of it is performance. Sometimes in the name of efficiency developers are writing low quality code that is difficult to support and maintain, without necessarily achieving positive results. In this regards I would like to quote Michael A. Jackson:

“The First Rule of Program Optimisation: Don’t do it. The Second Rule of Program Optimisation (for experts only!): Don’t do it yet.”

Before starting any optimisation we need to identify the problem. As a first step we should collect general data about application performance. For example, time from startActivity() call until data is displayed on the new screen. Or time it takes to load next page of content for a RecyclerView. Comparing this information with acceptable thresholds can tell us if there is a problem. In case something takes longer then expected, we can check more deeply and identify method or API call that caused the issue.

Luckily for us, there are some tools for profiling Android applications:

  1. Hugo — library that provides annotation-triggered method call logging. You need to annotate your method with @DebugLog, and it will log arguments, return values, and the time it took to execute. I like it, but it is limited to a single method call, so you can’t measure time it takes from opening Activity till displaying content.
  2. Android Studio toolset, for example System Trace, is very precise and provides a lot of information. But you need to spend a significant amount of time to collect and analyze the data.
  3. Backend solutions, for example JMeter. They have a lot of functionality, but it takes lots of efforts and time, first, to learn and, second, to use them. And after all, how often do you need to profile an application under heavy concurrent load? It doesn’t look like a common case in Android Development.

Missing tool

  1. Execution time for a particular method or API call. This can be covered by Hugo, for example.
  2. Time between two different events, that could happened in independent pieces of the code, but connected in terms of logic. Android Studio toolset can cover this topic, but, as already mentioned, you need to spend significant amount of time to profile.

While I was checking available profiling tools and trying to imagine all the possible cases, I realised that at least one tool is missing. So I came up with following requirements for it:

  1. Start and stop profiling should be triggered with two independent events, that will allow us to be as flexible as it needs.
  2. If we want to monitor application performance, then having just start and stop events is not enough. Sometimes we want to know more about what happens in between. Information about all intermediate stages should be united into one big report. That will allow us to easier understand and share the data.
  3. Sometimes we have scenarios that repeats frequently, for example, loading next page of content for a RecyclerView. Then one round of profiling is not enough, having statistical data, like average, minimal and maximal times, can give us more insights.

That is why I created Pury.

Introduction to Pury

Output for launching an example application:

App Start --> 0ms
  Splash Screen --> 5ms
    Splash Load Data --> 37ms
    Splash Load Data <-- 1042ms, execution = 1005ms
  Splash Screen <-- 1042ms, execution = 1037ms
  Main Activity Launch --> 1043ms 
    onCreate() --> 1077ms 
    onCreate() <-- 1100ms, execution = 23ms
    onStart() --> 1101ms 
    onStart() <-- 1131ms, execution = 30ms
  Main Activity Launch <-- 1182ms, execution = 139ms
App Start <-- 1182ms

As you can see, Pury measured time for launching the application, including intermediate stages, like loading data on splash screen and activity’s lifecycle methods. For each stage start and stop timestamps are displayed and so as execution time. Apart from the normal profiling, it could be useful for performance monitoring, to be sure that some changes don’t introduce an unexpected latency.

Output for a screen with pagination:

Get Next Page --> 0ms
  Load --> avg = 1.80ms, min = 1ms, max = 3ms, for 5 runs
  Load <-- avg = 258.40ms, min = 244ms, max = 278ms, for 5 runs
  Process --> avg = 261.00ms, min = 245ms, max = 280ms, for 5 runs
  Process <-- avg = 114.20ms, min = 99ms, max = 129ms, for 5 runs
Get Next Page <-- avg = 378.80ms, min = 353ms, max = 411ms, for 5 runs

In this example, as you can see, Pury collected information about loading next page 5 times and then took the average. For each stage start timestamp and execution time are displayed.

Inner structure and limitations

Performance measurements are done by Profilers. Each Profiler contains a list of Runs. Multiple Profilers can work in parallel, but only a single Run per each Profiler can be active. Once all Runs in a single Profiler are finished, result is reported. Amount of runs defines by runsCounter parameter.

Two Profilers running in parallel. First has a single Run with an active stage. Second has one stopped Run and one active, each of Runs contains a root Stage with two nested Stages. Active stages are green, stopped stages are red.

Run has a root Stage inside. Each Stage has a name, an order number and an arbitrary amount of nested Stages. Stage can have only one active nested Stage. If you stop a parent Stage, then all nested Stages are also stopped.

Profiling with Pury

1. StartProfiling — triggers an event to start Stage or Run. Profiling will start before method execution.

@StartProfiling(profilerName = "List pagination", runsCounter = 3, stageName = "Loading", stageOrder = 0)
private void loadNextPage() { }

StartProfiling can accept up to 5 arguments:

  • profilerName — name of the profiler is displayed in the result. Along with runsCounter identifies the Profiler.
  • runsCounter — amount of runs for Profiler to wait for. Result is available only after all runs are stopped.
  • stageName — identifies a stage to start. Name is displayed in the result.
  • stageOrder — stage order reflects the hierarchy of stages. In order to start a new stage, it must be bigger then order of current most nested active stage. Stage order is a subject to one more limitation: first start event must have order number equal zero.
  • enabled — if set to false, an annotation is skipped.

I want to emphasise one fact. Profiler is identified by combination of profilerName and runsCounter. So if you are using same profilerName, but different runsCounter, then you will get two separate results, instead of a combined one.

2. StopProfiling — triggers an event to stop Stage or Run. Profiling is stopped after method execution. Once Stage or Run is stopped, all nested stages are also stopped.

@StopProfiling(profilerName = "List pagination", runsCounter = 3, stageName = "Loading")
private void displayNextPage() { }

It has the same arguments as StartProfiling, except stageOrder.

3. MethodProfiling — combination of StartProfiling and StopProfiling.

@MethodProfiling(profilerName = "List pagination", runsCounter = 3, stageName = "Process", stageOrder = 1)
private List<String> processNextPage() { }

It has exact same arguments as StartProfiling with one small remark. If stageName is empty then it will be generated from method’s name and class. This is made in order to be able to use MethodProfiling without any arguments and get a meaningful result.

Since Java 7 doesn’t support repeatable annotations, I made group annotations for each of annotation above:

@StartProfilings(StartProfiling[] value)@StopProfilings(StopProfiling[] value)@MethodProfilings(MethodProfiling[] value)

As already mentioned, you can call start or stop profiling with a direct call:


Arguments are exactly the same as in corresponding annotations, except enabled, of course.

Logging Results

public interface Logger {
    void result(String tag, String message);

    void warning(String tag, String message);

    void error(String tag, String message);

By default result goes to Log.d, warning to Log.w and error to Log.e.

How to start using Pury?

buildscript {
    repositories {
    dependencies {
        classpath 'com.nikitakozlov:weaverlite:1.0.0'
apply plugin: 'com.nikitakozlov.weaverlite'

You can enable/disable it on a debug and/or release build. Default configuration looks like this.

weaverLite {
    enabledForDebug = true
    enabledForRelease = false

Second, include following dependencies:

dependencies {
   compile 'com.nikitakozlov.pury:annotations:1.0.1'
   debugCompile 'com.nikitakozlov.pury:pury:1.0.2'

If you want to profile on release, then use compile instead of compileDebug for a second dependency.

UPDATE: no-op version is released. Check GitHub repo.

Small recommendation

public final class StartApp {
    public static final String PROFILER_NAME = "App Start";

    public static final String TOP_STAGE ="App Start";
    public static final int TOP_STAGE_ORDER = 0;

    public static final String SPLASH_SCREEN = "Splash Screen";
    public static final int SPLASH_SCREEN_ORDER = TOP_STAGE_ORDER + 1;

    public static final String MAIN_ACTIVITY_LAUNCH = "Main Activity Launch";
    public static final int MAIN_ACTIVITY_LAUNCH_ORDER = SPLASH_SCREEN_ORDER + 1;

    public static final String MAIN_ACTIVITY_CREATE = "onCreate()";

As you can see, every ORDER constant depends on the parent’s stage, it is very handy. You can also add a constant for runsCounter to be sure that you are always using the same one. If you add here an enabled flag then you can easily disable one particular scenario from a single place.


I would like to hear your opinion about this solution. If you have any suggestions please feel free to rise an issue on GitHub. You can also contact me via Gitter.

Nikita Kozlov

Written by

Android Software Engineer at Zalando SE, Berlin