Community
Participate
Working Groups
We appear to have a significant decrease in performance for the "Open LCD on Targets tab" test. Inquiring minds want to know why?
Logged bug 130935 for the filtering slowdown in the launch configuration dialog. I have altered the OpenLaunchConfigurationDialogTests to turn off all of the filtering via the debug preferences.
We have seen improvements with changes to the LCD for filtering etc.
This test opens many many dialogs. We are seeing mininal regression on Linux but major regression on Windows. I hope to do more profiling but it appears to be an issue down below us.
Let me know how you get on.
Tweak released to AntElementNode to have a less expensive hashcode/equals
Is there anything I should be looking into about dialog performance on Windows and Linux?
I am going to spend time on this today so I can be more helpful in pinning down what is causing the differences.
Created attachment 38433 [details] CPU Statistics for Eclipse 3.2 M3 run of Ant LCD test
Created attachment 38434 [details] CPU Statistics for Eclipse 3.2 N20060411-0010 run of Ant LCD test
The major diff that jumps out at me is the change in org.eclipse.swt.internal.win32.OS.CallWindowProcW(int, int, int, int, int) This is on WinXP on Sun JDK 1.4.2_08
CallWindowProcW() is just calling the window proc for whatever the particular control is and asking windows to process a WM_ style message for a control. The interesting this is that it should be getting called about the same number of times if the same controls exist in the test for M3 and N20060411-0010. What does the dialog look like? Is the code UI code the same? Have label controls been replaced by text controls? Can you attach a picture of the dialog?
Ok...I will get method call information and more info on who is doing the calling. The dialog (the debug launch configuration dialog) has changed to include a toolbar that used to just be some buttons for creation, duplication, removal etc of launch configurations. I will check on how this change maps to when the regressions showed up...I do not believe it is a direct match. Is it at all interesting that almost no regression on Linux but large on WinXP? Are there some controls that are just known to be worse on WinXP than Linux?
>Is it at all interesting that almost no regression on Linux but >large on WinXP? Yes. Although many people claim that Windows is faster, really some things are faster on Windows, some on GTK. >Are there some controls that are just known to be worse on WinXP than Linux? Not really. The bug system contains some information like this but the versions of the operating systems change and the underlying technology changes so fast that there are no generalizations. For example, new version of GTK use cairo under the covers to render the widgets. I am very interested in your findings. Please keep up the good work. It's a pitty that we aren't in the same location or I'd be hot to hack this with you.
Hopefully this will be my Friday fun project :-)
Anything?
Apologizing again...I am currently swamped with the Jazz work.
Nothing more planned for 3.2 RC2
Opportunity for improvement in 3.3
Decision has been made to remove these performance tests. With the changes to the launch configuration framework (new features added) these tests were no longer representative of changes / regressions in the Ant integration.