Community
Participate
Working Groups
RC1 RH Enterprise 3.0 I am going to attacha test where we populate a list 2 ways 1) By adding 50000 items one at a time 2) By adding 50000 items all at once On Windows the second method is about 4 times faster than the first. On GTK they are about the same amount of time. This influences our ListViewer as we switched from the first way to the second in 3.1 and have good performance improvements on Windows and a performance degradation on GTK as a result
Created attachment 22444 [details] SWT only List population test class from our suites
Here are the GTK numbers Scenario 'org.eclipse.jface.tests.performance.ListPopulationTest#testMediumAdd()' (average over 3 samples): System Time: 1.53 s Used Java Heap: -40674 Working Set: 93.33K Elapsed Process: 1.53 s Kernel time: 0 ms CPU Time: 1.51 s Hard Page Faults: 1 Soft Page Faults: 1 Text Size: 0 Data Size: 93.33K Library Size: 0 Scenario 'org.eclipse.jface.tests.performance.ListPopulationTest#testMediumSetItems()' (average over 3 samples): System Time: 1.5 s Used Java Heap: -39677 Working Set: 228K Elapsed Process: 1.5 stestMediumSetItems: tearDown... Kernel time: 0 ms CPU Time: 1.47 s Hard Page Faults: 0 Soft Page Faults: 40 Text Size: 0 Data Size: 133.33K Library Size: 94.67K Scenario 'org.eclipse.jface.tests.performance.ListPopulationTest#testLargeAdd()' (average over 3 samples): System Time: 15.01 s Used Java Heap: 444.38K Working Set: 1.2M Elapsed Process: 15.01 s Kernel time: 3 ms CPU Time: 14.97 s Hard Page Faults: 0 Soft Page Faults: 297 Text Size: 0 Data Size: 66.67K Library Size: 1.14M Scenario 'org.eclipse.jface.tests.performance.ListPopulationTest#testLargeSetItems()' (average over 3 samples): System Time: 14.96 s Used Java Heap: -63669 Working Set: 1.12M Elapsed Process: 14.96 s Kernel time: 3 ms CPU Time: 14.9 s Hard Page Faults: 0 Soft Page Faults: 301 Text Size: 0 Data Size: 42.67K Library Size: 1.08M
Here are the XP numbers Scenario 'org.eclipse.jface.tests.performance.ListPopulationTest#testMediumAdd()' (average over 100 samples): Used Java Heap: 4.94K Working Set: 200.92K Committed: 201.48K Working Set Peak: 4.52K Elapsed Process: 37 ms Kernel time: 1 ms Page Faults: 156 CPU Time: 37 ms GDI Objects: 0 Scenario 'org.eclipse.jface.tests.performance.ListPopulationTest#testMediumSetItems()' (average over 100 samples): Used Java Heap: -908 Working Set: -27525 Committed: -11059 Working Set Peak: 3.36K Elapsed Process: 13 ms Kernel time: 1 ms Page Faults: 41 CPU Time: 13 ms GDI Objects: 0 Scenario 'org.eclipse.jface.tests.performance.ListPopulationTest#testLargeAdd()' (average over 4 samples): Used Java Heap: 89.09K Working Set: 9.53M testLargeSetItems: setUp... Committed: 9.33M Working Set Peak: 5.69M Elapsed Process: 1.19 s Kernel time: 324 ms Page Faults: 52.07K CPU Time: 1.18 s GDI Objects: 0 Scenario 'org.eclipse.jface.tests.performance.ListPopulationTest#testLargeSetItems()' (average over 45 samples): Used Java Heap: 2.48K Working Set: 1.53M Committed: 1.52M Working Set Peak: 0 Elapsed Process: 88 ms Kernel time: 4 ms Page Faults: 401 CPU Time: 88 ms GDI Objects: 0
How is this a performance degradation? It looks like they take the same amount of time on GTK+.
Right but on Windows setItems() is much faster. We made a change for Bug 70856 to switch to setItems. This gave us some really good improvments on Windows but ended up causing a slowdown on GTK.
The slowdown is not demonstrated by your test application here. Is there a case which is slower? I have some theories about why add() would be significantly slower than setItems() on Windows, for example, work is done on every call to add() to fix the horizontal scrollbar but is only done once in a call to setItems(). This does not apply in GtkTreeView. There is no reason why setItems() should be slower, but it shouldn't be faster either. I am not sure there is any bug here.
I'm not sure there's a bug here either. There is no guarantee that setItems() is faster than add() for the same number of items on every platform. The main thing is that is cannot be slower. BB to close.
Please re-open if you have a case where setItems() is slower than looping and calling add().
