Some time ago we set up a slave called
hudson-perf1-tests. This slave had a dedicated CPU core and only
one thread executor, and in our own tests, over a 24h period of
mixed activity levels, our performance results have been
consistent.
https://hudson.eclipse.org/hudson/computer/hudson-perf1-tests/
We (webmasters) have not been able to confirm
if the slave produces consistent performance indicators when run
in Hudson, however, so please, feel free to try it out and report
back.
Please see bug 333594 for more background.
Denis
On 04/11/2011 08:22 AM, fgiquel@xxxxxxxxxxxxxxxx wrote:
Hi,
the simultaneous release page, http://eclipse.org/indigo/planning/EclipseSimultaneousRelease.php,
describes the following requirement :
"Performance. Project
should
have measurable performance criteria that are regularly tested
against.
Projects should devote at least one milestone to performance and
scalability
improvements."
We have some non reg scalability
tests
dealing with memory usage for some of our components (MoDisco
project)
and we are thinking about additional tests for execution time
non regression.
We are asking ourselves how a relevant time measurement can be
done on
continuous integration on Hudson ... Hudson CPU availability is
inconstant
depending on use made from the different eclipse project teams,
so one
"basic" execution time measurement (using
System.currentTimeMillis())
has no meaning.
Has anyone met this kind of
problem
? Does the eclipse architecture council have some
recommendations about
that (did not find in http://wiki.eclipse.org/Architecture_Council/Top_Ten_Project_Development_Practices)
?
Fabien Giquel.
|
