Skip to main content

Why are annotations under Android such a performance issue (slow)?



I'm the lead author of ORMLite which uses Java annotations on classes to build database schemas. A big startup performance problem for our package turns out to be the calling of annotation methods under Android 1.6. I see the same behavior up through 3.0.





We are seeing that the following simple annotation code is incredibly GC intensive and a real performance problem. 1000 calls to an annotation method takes almost a second on a fast box. The same code under Java can do 28 million (sic) calls in the same time. We have an annotation that has 25 methods in it and we'd like to do more than 50 of these a second.





Does anyone know why this is happening and if there is any work around? There are certainly things that ORMLite can do in terms of caching this information but is there anything that we can do to "fix" annotations under Android? Thanks.







public void testAndroidAnnotations() throws Exception {

Field field = Foo.class.getDeclaredField("field");

MyAnnotation myAnnotation = field.getAnnotation(MyAnnotation.class);

long before = System.currentTimeMillis();

for (int i = 0; i < 1000; i++)

myAnnotation.foo();

Log.i("test", "in " + (System.currentTimeMillis() - before) + "ms");

}

@Target(FIELD) @Retention(RUNTIME)

private static @interface MyAnnotation {

String foo();

}

private static class Foo {

@MyAnnotation(foo = "bar")

String field;

}







This results in the following log output:







I/TestRunner( 895): started: testAndroidAnnotations

D/dalvikvm( 895): GC freed 6567 objects / 476320 bytes in 85ms

D/dalvikvm( 895): GC freed 8951 objects / 599944 bytes in 71ms

D/dalvikvm( 895): GC freed 7721 objects / 524576 bytes in 68ms

D/dalvikvm( 895): GC freed 7709 objects / 523448 bytes in 73ms

I/test ( 895): in 854ms







EDIT:





After @candrews pointed me in the right direction, I did some poking around the code. The performance problem looks to be caused by some terrible, gross code in Method.equals() . It is calling the toString() of both methods and then comparing them. Each toString() use StringBuilder with a bunch of append methods without a good initializing size. Doing the .equals by comparing fields would be significantly faster.





EDIT:





An interesting reflection performance improvement was given to me. We are now using reflection to peek inside the AnnotationFactory class to read the list of fields directly. This makes the reflection class 20 times faster for us since it bypasses the invoke which is using the method.equals() call. It is not a generic solution but here's the Java code from ORMLite SVN repository .



Source: Tips4all

Comments

  1. Google has acknowledged the issue and fixed it "post-Honeycomb"


    https://code.google.com/p/android/issues/detail?id=7811


    So at least they know about it and have supposedly fixed it for some future version.

    ReplyDelete
  2. Both GC and reflection are multitudes slower on dalvik compared to JVMs. From what I´ve manged to pick up it's basically because of design decisions made for the mobile platform.

    Can you do anything about the @Retention(RUNTIME)?

    You are supposed to reuse objects by overwriting value. See for example how getView is implemented on adapters.

    getView(int position, View convertView, ViewGroup parent) {
    ImageView imageView;


    where you reuse an old view and write new values to it instead of throwing it away to make a new one.

    ReplyDelete
  3. I think if you manage to change the RUNTIME retention policy, it should not be that slow.

    EDIT: I know, for your project that may not be an option. Perhaps it is more a problem of what you are doing with that annotation rather than bad performance in general.

    ReplyDelete

Post a Comment

Popular posts from this blog

[韓日関係] 首相含む大幅な内閣改造の可能性…早ければ来月10日ごろ=韓国

div not scrolling properly with slimScroll plugin

I am using the slimScroll plugin for jQuery by Piotr Rochala Which is a great plugin for nice scrollbars on most browsers but I am stuck because I am using it for a chat box and whenever the user appends new text to the boxit does scroll using the .scrollTop() method however the plugin's scrollbar doesnt scroll with it and when the user wants to look though the chat history it will start scrolling from near the top. I have made a quick demo of my situation http://jsfiddle.net/DY9CT/2/ Does anyone know how to solve this problem?

Why does this javascript based printing cause Safari to refresh the page?

The page I am working on has a javascript function executed to print parts of the page. For some reason, printing in Safari, causes the window to somehow update. I say somehow, because it does not really refresh as in reload the page, but rather it starts the "rendering" of the page from start, i.e. scroll to top, flash animations start from 0, and so forth. The effect is reproduced by this fiddle: http://jsfiddle.net/fYmnB/ Clicking the print button and finishing or cancelling a print in Safari causes the screen to "go white" for a sec, which in my real website manifests itself as something "like" a reload. While running print button with, let's say, Firefox, just opens and closes the print dialogue without affecting the fiddle page in any way. Is there something with my way of calling the browsers print method that causes this, or how can it be explained - and preferably, avoided? P.S.: On my real site the same occurs with Chrome. In the ex