Well, I wouldn't want to tell you what to do, but how I have gone about
it in the past is to
[i] Concoct a nice long test, so start-up costs don't get measured
as the main cost in the output
[ii] Fire up the profiler
[iii] Run it under different scenarios
This profiler has several modes of gathering data: sampling and
instrumenting:
sampling is less intrusive but will take longer to get good covereage,
whereas instrumenting has limits on the methods it can profile and makes
a huge impact on overall performance,
(which therefore invalidates the results a little bit).
Most profilers have some way of sorting the results for time spent by
Class/method/% of total/package
Just browsing around that should give a clue:
there are normally a couple of hot spots;
in Axis these tend to be in the Http client code and the Xml Parser
(I didn't use the commons http client for those test, and
perhaps I should have).
I don't think it's unreasonable that a SOAP library using http should
spend time in these areas,
so I didn't get too fussed, though there might be room for improvement.
If you sort by class, it will be obvious which Axis classes contribute
significantly to the total.
If there is a performance hit there, it should be easy to spot.
Since you have before and after Axis jars, it would be a great idea to
see if there was a significant impact made to the time spent in the
patched class before and after the patch.
Just some thoughts,
Patrick
-----Original Message-----
From: Vy Ho [mailto:[EMAIL PROTECTED]
Sent: 10 January 2005 21:55
To: [EMAIL PROTECTED]
Subject: Re: axis performance in multithreading env
This is a problem for me. I use Netbeans and the JFluid supports 1.4,
not Jdk1.5. My apps run under 1.5. It would take some effort on my
part to make these apps run elsewhere. I am currently tied up at a
moment, so I have to put this off for now. I am sorry about this. If
time permit, I'll work on this. Just to clear it a bit, did you want to
see the number of objects created, or the function calls count, or time
slice of each function call, or all of them? Because my test does not
based on any special type of webservice app, I think it can easily be
replicated somewhere else.
Patrick Martin wrote:
>Is there any chance you can use a profiler on these tests? This should
>makes things a lot clearer.
>
>If you're using Eclipse, there is an eclipse profiler that is free,
>easy to set up and very informative.
>
>http://sourceforge.net/projects/eclipsecolorer/
>
>
>
