Android
devices have a lot of cores, so writing smooth apps is a simple task for
anyone, right? Wrong. As everything on Android can be done in a lot of
different ways, picking the best option can be tough. If you want to choose the
most efficient method, you have to know what’s happening under the hood.
Luckily, you don’t have to rely on your feelings or sense of smell, since
there’s a lot of tools out there that can help you find bottlenecks by
measuring and describing what’s going on. Properly optimized and smooth apps
greatly improve the user experience, and also drain less battery.
Let’s see some numbers first to consider
how important optimization really is. According to a Nimbledroid post, 86% of users (including me) have
uninstalled apps after using them only once due to poor performance. If you’re
loading some content, you have less than 11 seconds to show it to the user.
Only every third user will give you more time. You might also get a lot of bad
reviews on Google Play because of it.
The first thing every user notices over and
over is the app’s startup time. According to another
Nimbledroid post, out of the 100 top apps, 40
start in under 2 seconds, and 70 start in under 3 seconds. So if possible, you
should generally display some content as soon as possible and delay the
background checks and updates a bit.
Always remember, premature optimization is
the root of all evil. You should also not waste too much time with micro
optimization. You will see the most benefit of optimizing code that runs often.
For example, this includes the
onDraw() function,
which runs every frame, ideally 60 times per second. Drawing is the slowest
operation out there, so try redrawing only what you have to. More about this
will come later.
Performance
Tips
Enough
theory, here is a list of some of the things you should consider if performance
matters to you.
1.
String vs StringBuilder
Let’s
say that you have a String, and for some reason you want to append more Strings
to it 10 thousand times. The code could look something like this.
String string = "hello";for (int i = 0; i < 10000; i++) {
string += " world";}
You
can see on the Android Studio Monitors how inefficient some String
concatenation can be. There’s tons of Garbage Collections (GC) going on.
This
operation takes around 8 seconds on my fairly good device, which has Android
5.1.1. The more efficient way of achieving the same goal is using a
StringBuilder, like this.
StringBuilder sb = new StringBuilder("hello");for (int i = 0; i < 10000; i++) {
sb.append(" world");}String string = sb.toString();
On
the same device this happens almost instantly, in less than 5ms. The CPU and
Memory visualizations are almost totally flat, so you can imagine how big this
improvement is. Notice though, that for achieving this difference, we had to
append 10 thousand Strings, which you probably don’t do often. So in case you
are adding just a couple Strings once, you will not see any improvement. By the
way, if you do:
String string = "hello" + " world";
It
gets internally converted to a StringBuilder, so it will work just fine.
You
might be wondering, why is concatenating Strings the first way so slow? It is
caused by the fact that Strings are immutable, so once they are created, they
cannot be changed. Even if you think you are changing the value of a String,
you are actually creating a new String with the new value. In an example like:
String myString = "hello";myString += " world";
What
you will get in memory is not 1 String “hello world”, but actually 2 Strings.
The String myString will contain “hello world”, as you would expect. However,
the original String with the value “hello” is still alive, without any
reference to it, waiting to be garbage collected. This is also the reason why
you should store passwords in a char array instead of a String. If you store a
password as a String, it will stay in the memory in human-readable format until
the next GC for an unpredictable length of time. Back to the immutability
described above, the String will stay in the memory even if you assign it
another value after using it. If you, however, empty the char array after using
the password, it will disappear from everywhere.
2.
Picking the Correct Data Type
Before you start writing code, you should
decide what data types you will use for your collection. For example, should
you use a
Vector or an ArrayList?
Well, it depends on your usecase. If you need a thread-safe collection, which
will allow only one thread at once to work with it, you should pick a Vector,
as it is synchronized. In other cases you should probably stick to an ArrayList,
unless you really have a specific reason to use vectors.
How about the case when you want a
collection with unique objects? Well, you should probably pick a
Set.
They cannot contain duplicates by design, so you will not have to take care of
it yourself. There are multiple types of sets, so pick one that fits your use
case. For a simple group of unique items, you can use a HashSet.
If you want to preserve the order of items in which they were inserted, pick a LinkedHashSet.
A TreeSet sorts
items automatically, so you will not have to call any sorting methods on it. It
should also sort the items efficiently, without you having to think of sorting
algorithms.
Data
dominates. If you've chosen the right data structures and organized things
well, the algorithms will almost always be self-evident. Data structures, not
algorithms, are central to programming.
— Rob Pike's 5 Rules of Programming
— Rob Pike's 5 Rules of Programming
Sorting integers or strings is pretty
straightforward. However, what if you want to sort a class by some property?
Let’s say you are writing a list of meals you eat, and store their names and
timestamps. How would you sort the meals by timestamp from the lowest to
highest? Luckily, it’s pretty simple. It’s enough to implement the
Comparable interface
in the Meal class
and override the compareTo() function.
To sort the meals by lowest timestamp to highest, we could write something like
this.@Override
public int compareTo(Object object) {
Meal meal = (Meal) object; if (this.timestamp < meal.getTimestamp()) { return -1; } else if (this.timestamp > meal.getTimestamp()) { return 1; } return 0;}
3.
Location Updates
There are a lot of apps out there which
collect the user’s location. You should use the Google Location Services API
for that purpose, which contains a lot of useful functions. There is a separate article about
using it, so I will not repeat it.
I’d
just like to stress some important points from a performance perspective.
First of all, use only the most precise
location as you need. For example, if you are doing some weather forecasting,
you don’t need the most accurate location. Getting just a very rough area based
on the network is faster, and more battery efficient. You can achieve it by
setting the priority to
LocationRequest.PRIORITY_LOW_POWER.
You can also use a function of
LocationRequest called setSmallestDisplacement().
Setting this in meters will cause your app to not be notified about location
change if it was smaller than the given value. For example, if you have a map
with nearby restaurants around you, and you set the smallest displacement to 20
meters, the app will not be making requests for checking restaurants if the
user is just walking around in a room. The requests would be useless, as there
wouldn’t be any new nearby restaurant anyway.
The second rule is requesting location
updates only as often as you need them. This is quite self explanatory. If you
are really building that weather forecast app, you do not need to request the
location every few seconds, as you probably don’t have such precise forecasts
(contact me if you do). You can use the
setInterval() function
for setting the required interval in which the device will be updating your app
about the location. If multiple apps keep requesting the user’s location, every
app will be notified at every new location update, even if you have a higher setInterval() set. To
prevent your app from being notified too often, make sure to always set a
fastest update interval with setFastestInterval().
And finally, the third rule is requesting
location updates only if you need them. If you are displaying some nearby
objects on the map every x seconds and the app goes in background, you do not
need to know the new location. There is no reason to update the map if the user
cannot see it anyway. Make sure to stop listening for location updates when
appropriate, preferably in
onPause().
You can then resume the updates in onResume().
4.
Network Requests
There is a high chance that your app is
using the internet for downloading or uploading data. If it is, you have
several reasons to pay attention to handling
network requests. One of them is mobile data,
which is very limited to a lot of people and you shouldn’t waste it.
So,
what can you do about it? Batch. To avoid waking up the radio every couple
seconds, prefetch things that the user might need in the upcoming minutes. The
proper way of batching is highly dynamic depending on your app, but if it is
possible, you should download the data the user might need in the next 3-4
minutes. One could also edit the batch parameters based on the user’s internet
type, or charging state. For example, if the user is on WiFi while charging,
you can prefetch a lot more data than if the user is on mobile internet with
low battery. Taking all these variables into account can be a tough thing,
which only few people would do. Luckily though, there is GCM Network Manager to
the rescue!
GCM Network Manager is a really helpful
class with a lot of customizable attributes. You can easily schedule both
repeating and one-off tasks. At repeating tasks you can set the lowest, as well
as the highest repeat interval. This will allow batching not only your
requests, but also requests from other apps. The radio has to be woken up only
once per some period, and while it’s up, all apps in the queue download and
upload what they are supposed to. This Manager is also aware of the device’s
network type and charging state, so you can adjust accordingly. You can find
more details and samples in this
article, I urge you to check it out. An example
task looks like this:
Task task = new OneoffTask.Builder() .setService(CustomService.class) .setExecutionWindow(0, 30) .setTag(LogService.TAG_TASK_ONEOFF_LOG) .setUpdateCurrent(false) .setRequiredNetwork(Task.NETWORK_STATE_CONNECTED) .setRequiresCharging(false) .build();
By the way, since Android 3.0, if you do a
network request on the main thread, you will get a
NetworkOnMainThreadException.
That will definitely warn you not to do that again.
5.
Reflection
Reflection is the ability of classes and
objects to examine their own constructors, fields, methods, and so on. It is
used usually for backward compatibility, to check if a given method is
available for a particular OS version. If you have to use reflection for that
purpose, make sure to cache the response, as using reflection is pretty slow.
Some widely used libraries use Reflection too, like Roboguice for dependency
injection. That’s the reason why you should prefer Dagger 2. For more details
about reflection, you can check a separate
post.
6.
Autoboxing
Autoboxing and unboxing are processes of
converting a primitive type to an Object type, or vice versa. In practice it
means converting an int to an Integer. For achieving that, the compiler uses
the
Integer.valueOf()function
internally. Converting is not just slow, Objects also take a lot more memory
than their primitive equivalents. Let’s look at some code.Integer total = 0;for (int i = 0; i < 1000000; i++) {
total += i;}
While
this takes 500ms on average, rewriting it to avoid autoboxing will speed it up
drastically.
int total = 0;
for (int i = 0; i < 1000000; i++) {
total += i;}
This
solution runs at around 2ms, which is 25 times faster. If you don’t believe me,
test it out. The numbers will be obviously different per device, but it should
still be a lot faster. And it’s also a really simple step to optimize.
Okay, you probably don’t create a variable
of type Integer like this often. But what about the cases when it is more
difficult to avoid? Like in a map, where you have to use Objects, like
Map<Integer,
Integer>? Look at the solution many people use.Map<Integer, Integer> myMap = new HashMap<>();for (int i = 0; i < 100000; i++) {
myMap.put(i, random.nextInt());}
Inserting
100k random ints in the map takes around 250ms to run. Now look at the solution
with SparseIntArray.
SparseIntArray myArray = new SparseIntArray();for (int i = 0; i < 100000; i++) {
myArray.put(i, random.nextInt());}
This
takes a lot less, roughly 50ms. It’s also one of the easier methods for
improving performance, as nothing complicated has to be done, and the code also
stays readable. While running a clear app with the first solution took 13MB of
my memory, using primitive ints took something under 7MB, so just the half of
it.
SparseIntArray is just one of the cool
collections that can help you avoid autoboxing. A map like
Map<Integer, Long> could
be replaced by SparseLongArray,
as the value of the map is of type Long.
If you look at the source
code of SparseLongArray,
you will see something pretty interesting. Under the hood, it is basically just
a pair of arrays. You can also use a SparseBooleanArray similarly.
If you read the source code, you might have
noticed a note saying that
SparseIntArray can be
slower than HashMap.
I’ve been experimenting a lot, but for me SparseIntArray was
always better both memory and performance wise. I guess it’s still up to you
which you choose, experiment with your use cases and see which fits you the
most. Definitely have the SparseArrays in your
head when using maps.
7.
OnDraw
As I’ve said above, when you are optimizing
performance, you will probably see the most benefit in optimizing code which
runs often. One of the functions running a lot is
onDraw().
It may not surprise you that it’s responsible for drawing views on the screen.
As the devices usually run at 60 fps, the function is run 60 times per second.
Every frame has 16 ms to be fully handled, including its preparation and
drawing, so you should really avoid slow functions. Only the main thread can
draw on the screen, so you should avoid doing expensive operations on it. If
you freeze the main thread for several seconds, you might get the infamous
Application Not Responding (ANR) dialog. For resizing images, database work,
etc., use a background thread.
If you think your users
won't notice that drop in frame rate, you are wrong!
I’ve
seen some people trying to shorten their code, thinking that it will be more
efficient that way. That definitely isn’t the way to go, as shorter code
totally doesn’t mean faster code. Under no circumstances should you measure the
quality of code by the number of lines.
One of the things you should avoid in
onDraw() is
allocating objects like Paint. Prepare everything in the constructor, so it’s
ready when drawing. Even if you have onDraw() optimized,
you should call it only as often as you have to. What is better than calling an
optimized function? Well, not calling any function at all. In case you want to
draw text, there is a pretty neat helper function called drawText(),
where you can specify things like the text, coordinates, and the text color.
8.
ViewHolders
You probably know this one, but I cannot
skip it. The Viewholder design pattern is a way of making scrolling lists
smoother. It is a kind of view caching, which can seriously reduce the calls to
findViewById() and
inflating views by storing them. It can look something like this.static class ViewHolder {
TextView title; TextView text; public ViewHolder(View view) { title = (TextView) view.findViewById(R.id.title); text = (TextView) view.findViewById(R.id.text); }}
Then, inside the
getView() function
of your adapter, you can check if you have a useable view. If not, you create
one.ViewHolder viewHolder;if (convertView == null) {
convertView = inflater.inflate(R.layout.list_item, viewGroup, false); viewHolder = new ViewHolder(convertView); convertView.setTag(viewHolder);} else { viewHolder = (ViewHolder) convertView.getTag();} viewHolder.title.setText("Hello World");
You
can find a lot of useable info about this pattern around the internet. It can
also be used in cases when your list view has multiple different types of
elements in it, like some section headers.
9.
Resizing Images
Chances are, your app will contain some
images. In case you are downloading some JPGs from the web, they can have
really huge resolutions. However, the devices they will be displayed on will be
a lot smaller. Even if you take a photo with the camera of your device, it
needs to be downsized before displaying as the photo resolution is a lot bigger
than the resolution of the display. Resizing images before displaying them is a
crucial thing. If you’d try displaying them in full resolutions, you’d run out
of memory pretty quickly. There is a
whole lot written about displaying bitmaps efficiently in the Android docs, I
will try summing it up.
So
you have a bitmap, but you don’t know anything about it. There’s a useful flag
of Bitmaps called inJustDecodeBounds at your service, which allows you to find
out the bitmap’s resolution. Let’s assume that your bitmap is 1024x768, and the
ImageView used for displaying it is just 400x300. You should keep dividing the
bitmap’s resolution by 2 until it’s still bigger than the given ImageView. If
you do, it will downsample the bitmap by a factor of 2, giving you a bitmap of
512x384. The downsampled bitmap uses 4x less memory, which will help you a lot
with avoiding the famous OutOfMemory error.
Now that you know how to do it, you should
not do it. … At least, not if your app relies on images heavily, and it’s not
just 1-2 images. Definitely avoid stuff like resizing and recycling images
manually, use some third party libraries for that. The most popular ones are Picasso by
Square, Universal
Image Loader, Fresco by
Facebook, or my favourite, Glide.
There is a huge active community of developers around it, so you can find a lot
of helpful people at the issues section on GitHub as well.
10.
Strict Mode
Strict Mode is a quite useful developer
tool that many people don’t know about. It’s usually used for detecting network
requests or disk accesses from the main thread. You can set what issues Strict
Mode should look for and what penalty it should trigger. A google sample looks
like this:
public void onCreate() {
if (DEVELOPER_MODE) { StrictMode.setThreadPolicy(new StrictMode.ThreadPolicy.Builder() .detectDiskReads() .detectDiskWrites() .detectNetwork() .penaltyLog() .build()); StrictMode.setVmPolicy(new StrictMode.VmPolicy.Builder() .detectLeakedSqlLiteObjects() .detectLeakedClosableObjects() .penaltyLog() .penaltyDeath() .build()); } super.onCreate();}
If you want to detect every issue Strict
Mode can find, you can also use
detectAll().
As with many performance tips, you should not blindly try fixing everything
Strict Mode reports. Just investigate it, and if you are sure it’s not an
issue, leave it alone. Also make sure to use Strict Mode only for debugging,
and always have it disabled on production builds.
Debugging
Performance: The Pro Way
Let’s
now see some tools that can help you find bottlenecks, or at least show that
something is wrong.
1.
Android Monitor
This
is a tool built into Android Studio. By default, you can find the Android
Monitor at the bottom left corner, and you can switch between 2 tabs there.
Logcat and Monitors. The Monitors section contains 4 different graphs. Network,
CPU, GPU, and Memory. They are pretty self explanatory, so I will just quickly
go through them. Here is a screenshot of the graphs taken while parsing some
JSON as it is downloaded.
The Network part shows the incoming and
outgoing traffic in KB/s. The CPU part displays the CPU usage in percent. The
GPU monitor displays how much time it takes to render the frames of a UI
window. This is the most detailed monitor out of these 4, so if you want more
details about it, read
this.
Lastly we have the Memory
monitor, which you will probably use the most. By
default it shows the current amount of Free and Allocated memory. You can force
a Garbage Collection with it too, to test if the amount of used memory drops
down. It has a useful feature called Dump Java Heap, which will create a HPROF
file which can be opened with the HPROF
Viewer and Analyzer. That will enable you to see
how many objects you have allocated, how much memory is taken by what, and maybe
which objects are causing memory leaks. Learning how to use this analyzer is
not the simplest task out there, but it is worth it. The next thing you can do
with the Memory Monitor is do some timed Allocation Tracking, which you can
start and stop as you wish. It could be useful at many cases, for example when
scrolling or rotating the device.
2.
GPU Overdraw
While
seeing true color is the best, you will always see some overdraws, especially
around texts, navigation drawers, dialogs and more. So don’t try getting rid of
it fully. If your app is blueish or greenish, that’s probably fine. However, if
you see too much red on some simple screens, you should investigate what’s
going on. It might be too many fragments stacked onto each other, if you keep
adding them instead of replacing. As I’ve mentioned above, drawing is the
slowest part of apps, so there is no sense drawing something if there will be
more than 3 layers drawn onto it. Feel free to check out your favourite apps
with it. You will see that even apps with over a billion downloads have red
areas, so just take it easy when you are trying to optimize.
3.
GPU Rendering
The bars consist of 3-4 colors, and
according to the Android docs, their size indeed matters. The smaller, the
better. At the bottom you have blue, which represents the time used to create
and update the View’s display lists. If this part is too tall, it means that
there is a lot of custom view drawing, or a lot of work done in the
onDraw() functions.
If you have Android 4.0+, you will see a purple bar above the blue one. This
represents the time spent transferring resources to the render thread. Then
comes the red part, which represents the time spent by Android’s 2D renderer
issuing commands to OpenGL to draw and redraw display lists. At the top is the
orange bar, which represents the time the CPU is waiting for the GPU to finish
its work. If it’s too tall, the app is doing too much work on the GPU.
If
you are good enough, there is one more color above the orange. It is a green
line representing the 16 ms threshold. As your goal should be running your app
at 60 fps, you have 16 ms to draw every frame. If you don’t make it, some
frames might be skipped, the app could become jerky, and the user would
definitely notice. Pay special attention to animations and scrolling, that’s
where the smoothness matters the most. Even though you can detect some skipped
frames with this tool, it won’t really help you figuring out where exactly the
problem is.
4.
Hierarchy Viewer
This is one of my favourite tools out
there, as it’s really powerful. You can start it from Android Studio through
Tools -> Android -> Android Device Monitor, or it is also in your
sdk/tools folder as “monitor”. You can also find a standalone hierarachyviewer
executable there, but as it’s deprecated you should open the monitor. However
you open the Android Device Monitor, switch to the Hierarchy Viewer
perspective. If you don’t see any running apps assigned to your device, there are a couple
things you can do to fix it. Also try checking outthis issue
thread, there are people with all kinds of issues and all kinds of solutions.
Something should work for you too.
With
Hierarchy Viewer, you can get a really neat overview of your view hierarchies
(obviously). If you see every layout in a separate XML, you might easily spot
useless views. However, if you keep combining the layouts, it can easily get
confusing. A tool like this makes it simple to spot, for example, some
RelativeLayout, which has just 1 child, another RelativeLayout. That makes one
of them removable.
Avoid calling
requestLayout(),
as it causes traversing of the entire view hierarchy, to find out how big each
view should be. If there is some conflict with the measurements, the hierarchy
might be traversed multiple times, which if happens during some animation, it
will definitely make some frames be skipped. If you want to find out more about
how Android draws its views, you can read
this. Let’s look at one view as seen in
Hierarchy Viewer.
The
top right corner contains a button for maximizing the preview of the particular
view in a standalone window. Under it you can also see the actual preview of
the view in the app. The next item is a number, which represents how many
children the given view has, including the view itself. If you select a node
(preferably the root one) and press “Obtain layout times” (3 colored circles),
you will have 3 more values filled, together with colored circles appearing
labelled measure, layout, and draw. It might not be shocking that the measure
phase represents the time it took to measure the given view. The layout phase
is about the rendering time, while the drawing is the actual drawing operation.
These values and colors are relative to each other. Green one means that the
view renders in the top 50% of all views in the tree. Yellow means rendering in
the slower 50% of all views in the tree, red means that the given view is one
of the slowest. As these values are relative, there will always be red ones.
You simply cannot avoid them.
Under
the values you have the class name, such as “TextView”, an internal view ID of
the object, and the android:id of the view, which you set in the XML files. I
urge you to build a habit of adding IDs to all views, even if you don’t
reference them in the code. It will make identifying the views in Hierarchy
Viewer really simple, and in case you have automated tests in your project, it
will also make targeting the elements a lot faster. That will save some time
for you and your colleagues writing them. Adding IDs to elements added in XML
files is pretty straightforward. But what about the dynamically added elements?
Well, it turns out to be really simple too. Just create an ids.xml file inside
your values folder and type in the required fields. It can look like this:
<resources> <item name="item_title" type="id"/> <item name="item_body" type="id"/></resources>
Then in the code, you can use
setId(R.id.item_title).
It couldn’t be simpler.
There are a couple more things to pay
attention to when optimizing UI. You should generally avoid deep hierarchies
while preferring shallow, maybe wide ones. Do not use layouts you don’t need.
For example, you can probably replace a group of nested
LinearLayouts with
either a RelativeLayout,
or a TableLayout.
Feel free to experiment with different layouts, don’t just always use LinearLayout and RelativeLayout.
Also try creating some custom views when needed, it can improve the performance
significantly if done well. For instance, did you know that Instagram doesn’t
use TextViews for displaying
comments?
You can find some more info about Hierarchy
Viewer on the Android
Developers site with
descriptions of different panes, using the Pixel Perfect tool, etc. One more
thing I would point out is capturing the views in a .psd file, which can be
done by the “Capture the window layers” button. Every view will be in a
separate layer, so it’s really simple to hide or change it in Photoshop or
GIMP. Oh, that’s another reason to add an ID to every view you can. It will
make the layers have names that actually make sense.
You
will find a lot more debugging tools in Developer options, so I advise you to
activate them and see what are they doing. What could possibly go wrong?
The Android developers site contains a set
of best
practices for performance. They cover a lot of
different areas, including memory management, which I haven’t really talked
about. I silently ignored it, because handling memory and tracking memory leaks
is a whole separate story. Using a third party library for efficiently
displaying images will help a lot, but if you still have memory issues, check
out Leak canary made by
Square, or read this.
Wrapping Up
So, this was the good news. The bad new is,
optimizing Android apps
is a lot more complicated. There are a lot of ways of doing everything, so you
should be familiar with the pros and cons of them. There usually isn’t any
silver bullet solution which has only benefits. Only by understanding what’s
happening behind the scenes will you be able to pick the solution which is best
for you. Just because your favorite developer says that something is good, it
doesn’t necessarily mean that it’s the best solution for you. There are a lot
more areas to discuss and more profiling tools which are more advanced, so we
might get to them next time.
Make sure you learn from the top developers
and top companies. You can find a couple hundred engineering blogs at this link.
It’s obviously not just Android related stuff, so if you are interested only in
Android, you have to filter the particular blog. I would highly recommend the
blogs of Facebook and Instagram. Even though the Instagram UI on Android is
questionable, their engineering blog has some really cool articles. For me it’s
awesome that it’s so easy to see how things are done in companies which are
handling hundreds of millions of users daily, so not reading their blogs seems
crazy. The world is changing really fast, so if you aren’t constantly trying to
improve, learn from others and use new tools, you will be left behind. As Mark
Twain said, a person who doesn’t read has no advantage over one who can’t read.
This article was written by Tibor Kaputa, a
Toptal Java developer.
No comments:
Post a Comment