Re: [platform-core-dev] Not all performance test failures are bad

[Pascal Rapicault <Pascal_Rapicault@xxxxxxxxxx>]

 I remember that we restricted
ourselves to 50% for 3.1, so that we can have space to improve performance
in 3.2M7 <g>

John Arthorne/Ottawa/IBM@IBMCA
Sent by: platform-core-dev-bounces@xxxxxxxxxxx

04/06/2005 04:48 PM

Please respond to
"Eclipse Platform Core component developers list."

To	platform-core-dev@xxxxxxxxxxx
cc
Subject	[platform-core-dev] Not all performance test failures are bad

With our focus on performance bugs in M7, it's sometimes nice to discover
good performance news. 

In Eclipse 3.1 M6, our headless startup performance test failed on Windows
(StartupTest.testApplicationStartup).
 This test measures the elapsed time to start a trivial Eclipse application
with no GUI. In this week's build, it now fails on both Windows and Linux.
 After some investigation, we found that this startup test was asserting
that headless application startup is between 5% slower and 50% faster than
the elapsed time for the same test on Eclipse 3.0.  It turns out that
the reason for the failure is that this test is now MORE than 50% faster
than 3.0.  In this week's build we clocked in at 61% faster on Windows
(1.5 seconds), and 65% faster on Linux (1.6 seconds).  

I have fixed the test to assert performance in the range from 5% to 100%.
 So, if you see red X's on the performance results page, take a closer
look to make sure the result is not actually too fast rather than too slow.
 Hats off to the Runtime/OSGi developers who introduced major performance
improvements in the startup sequence in 3.1 to achieve this result._______________________________________________
platform-core-dev mailing list
platform-core-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/platform-core-dev