Re: [9] Review Request: 8071306 GUI perfomance are very slow compared java 1.6.0_45

[Anton V. Tarasov]

On 15.05.2015 1:57, Sergey Bylokhov wrote:

Hi, Anton.

+     * Determines the bounds of a visible part of the component relative to its
+     * parents.
Did you mean "to its parent"?

Yep, new version:
http://cr.openjdk.java.net/~serb/8071306/webrev.05/

Looks good to me.

Regards,
Anton.

2.

 100      * The components in this container.
 101      * @see #add
 102      * @see #getComponents
 103      */
 104     private java.util.List<Component> component = new ArrayList<>();

May be it's worth to rename the field? The "component" name is odd...

I suppose it wasn't changed, because this name is used in the serialization for a long time. Plus there are a bunch of the similar vars like: tmpComponent etc. I can do it later for jdk9 only.

Ok, thanks. It's up to you.

Regards,
Anton.

Regards,
Anton.

On 07.05.2015 3:39, Sergey Bylokhov wrote:

Hello.
Please review the fix for a jdk9. I plan to backport it to jdk8u60.

Description.

An UIworks really slowly, when an application has a lot of components in one container, and these components should be disabled one by one.

The reason is the next sequence of methods calls:

Component.setEnabled->updateCursorImmediately()-> some cursor related staff->GlobalCursorManager._updateCursor->Container.findComponentAt()-iteration over all components in the container.....-> twice....

You can imagine how it works in case of 10000 components in the container.

Note that in the bug report described difference jdk6 vs jdk8 -> 1sec vs 6 sec. This was caused by the two fixes, one of which adds checkTreeLock() and in another one a simple array of components was replaced by the ArrayList. Since code was added to the really hot method we got so big slowdown.

To fix the problem I suggest two different approaches:
 - Container.java: Fix a general case, by eliminating a second iteration in the 
hot loop.

- Component.java: Totally eliminates cursor machinery, if component cannot affect current cursor.

Some speedup measurements on my local system:

- Simple removing of checkTreeLock() will partly solve a regression reported by the user(12 sec -> 5 sec).

 - Changes in the loop will speedup the code in the worse case(5->2 sec)
 - The changes in the Component.java will change the performance from 2 sec to 
100< ms

Test case which was added was improved from 10 seconds to 80 ms.

JMH test: 
http://cr.openjdk.java.net/~serb/8071306/SetEnabledPerformanceTest.java

Fixedversion:

Benchmark                                       Mode Cnt       Score       
Error   Units
SetEnabledPerformanceTest.testContains thrpt    5  301300,813 ± 17338,045  
ops/ms
SetEnabledPerformanceTest.testFindComponentAt thrpt    5      20,521 ±     
0,269  ops/ms
SetEnabledPerformanceTest.testGetComponentAt thrpt    5      22,297 ±     1,264 
 ops/ms
SetEnabledPerformanceTest.testSetEnabled thrpt    5     711,120 ±    19,837  
ops/ms

Base version:

Benchmark                                       Mode Cnt       Score      Error 
  Units
SetEnabledPerformanceTest.testContains thrpt    5  299145,642 ± 2120,183  ops/ms
SetEnabledPerformanceTest.testFindComponentAt thrpt    5       1,101 ±    0,012 
 ops/ms
SetEnabledPerformanceTest.testGetComponentAt thrpt    5       6,792 ±    0,097  
ops/ms
SetEnabledPerformanceTest.testSetEnabled thrpt    5       0,464 ±    0,020  
ops/ms


Bug: https://bugs.openjdk.java.net/browse/JDK-8071306
Webrev can be found at: http://cr.openjdk.java.net/~serb/8071306/webrev.03

--

Best regards, Sergey.

--
Best regards, Sergey.

--
Best regards, Sergey.