Bugzilla – Full Text Bug Listing
| Summary: | Optimize startup performance by replacing EventListenerList implementation | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Product: | [Tools] GEF | Reporter: | Iwan Birrer <iwanbirrer> | ||||||||
| Component: | GEF-Legacy Draw2d | Assignee: | Anthony Hunter <ahunter.eclipse> | ||||||||
| Status: | NEW --- | QA Contact: | |||||||||
| Severity: | normal | ||||||||||
| Priority: | P3 | CC: | ahunter.eclipse, hudsonr, igor.burilo, tfletche | ||||||||
| Version: | 3.4 | ||||||||||
| Target Milestone: | --- | ||||||||||
| Hardware: | PC | ||||||||||
| OS: | Windows XP | ||||||||||
| Whiteboard: | |||||||||||
| Attachments: |
|
||||||||||
Created attachment 118061 [details]
Executable gef snippet to profile initialization times
We need to look into this for 3.5 if a performance gain is to be had. The implementation of EventListenerList is similar in performance to the implementation of EventTable in SWT, or ListenerList in org.eclipse.core.runtime. Insertion performance is O(n^2). With the patch, it is still O(n^2), but the performance for each type of listener interface is isolated from the others. A better approach would be to grow Object array's size geometrically, as is done in ArrayList#ensureCapacity(). That way the number of calls to arrayCopy is Log(n) instead of N. I have the same problem too. For 87000 edit parts it takes about 80 seconds. Also the problem isn't only in EventListenerList but in Figure.addPropertyChangeListener, I'll add an attachment of my profiling. Target milestone for this bug is set to 3.5, but 3.5 is already completed, is there any plans for this bug for 3.6? I'm using GEF 3.4.1. Created attachment 159408 [details]
add profiling screenshot
Unset target milestone as the specified one is already passed. |
Created attachment 118060 [details] Replacement for the draw2d EventListenerList Build ID: M20080911-1700 Steps To Reproduce: 1. Create a new Pluin project, add the dependencies listed below to the Manifest and copy the attached EventListenerPerformance.java to the src folder. - org.eclipse.jface 3.4.0 - org.eclipse.gef 3.4.1 - org.eclipse.core.runtime 3.4.0 - com.ibm.icu 3.8.1 2. Run the EventListenerPerformance.java. Initialization times are printed to the console. 3. Import the org.eclipse.draw2d plug-in as a source project and replace the org.eclipse.draw2d.EventListenerList with the one in the attachenent. 4. Re-run the EventListenerPerformance.java. Initialization times are printed to the console. More information: With a growing number of edit parts (> 5000) the initialization time of a gef editor can increase to several seconds. After some profiling we found that a great deal of the time is taken by the org.eclipse.draw2d.EventListenerList.addListener() class due to a System.arraycopy() for each listener added to the list. The attached snippets demonstrates the slow down caused by the EventListenerList class for a growing number of edit parts by measuring the time for EditorViewer.setContents(). 2000 edit parts: 141 ms 10000 edit parts: 1828 ms 20000 edit parts: 7109 ms 40000 edit parts: 31235 ms By removing the array copy in the EventListenerList (see attached EventListenerList.java) and replacing the internal array with a map, we got the following numbers: 2000 edit parts: 110 ms 10000 edit parts: 375 ms 20000 edit parts: 594 ms 40000 edit parts: 1297 ms The performance gain is quite obvious. It boils down to a different approach on the array copy strategy: The attached EventListenerList implementation does the copy of the array when iterating over the list while the draw2d/GEF version performs the copy operation when adding a listener to the list. However the attached version may slow down updating the edit parts/figures. Our tests with and editable editor hasn't shown a perceivable slow down though. We hope somebody can look at this issue and let us know if there's a chance that the EventListenerList implementation is changed for the next release. Test environment: - Intel Core 2, T7200 @ 2 GHz, 2 GB Ram - JavaSE-1.6