extreme performance analysis
Tim Panton - Westhawk Ltd
tpanton at ibm.net
Sat Nov 12 19:29:10 PST 2005
I'm not at my desk now, so this
is from memory:-(
Jdk 1.3 has such a tool built in.
In earlier versions it was called
Hat, but I think the name changed
for 1.3. It was announced at '98
javaone.
I've used optimizeit for this sort of
work (admittedly not for 2 days,
but overnight.) in the latest version,
you can run the interface on a
second system, which helps performance.
Tim.
John Farrell <jfarrell at mincom.com> wrote:
__________
>I had better clarify why I don't think JProbe is good enough in our
>situation.
>
>JProbe doesn't provide information until the run has finished. The
>latest version seems to run at about 1/6 real speed, so to analyse
>why the VM size blows out after 2 days, we would need to run the
>profiled version for 12 days. Furthermore, as our VM size is between
>10M and 80M, with average about 50M, the JProbe VM size would be who
>knows what over that, particularly after accumulating information for
>12 days. And if after 12 days we haven't actually observed the problem,
>what do we do? JProbe is OK for things that take a short time to run,
>but I don't trust it enough to allocate a whole machine to it for 2 weeks,
>and I certainly wouldn't suggest that that's a reliable way to solve
>our performance problems.
>
>What would be nice would be an instrumented VM which opened a socket
>to a data-gathering process on another machine, and reported every time
>an object was created or gced, and periodically reported the stack
>traces of various threads. On the data-gathering machine, we could do
>DB queries against the gathered data to see what sort of objects still
>existed, and which methods had been called recently. By putting the
>data-gathering on a separate machine, a lot of strain is taken off the
>machine doing the profiling. By splitting the profiling into two
>processes, you can analyse the data as it's being produced.
>
>Has anyone written such a VM? I suspect we would pay good money for
>it! (Oh yeah, has to work with JDK 1.3 classes.)
>
>John
---
URL http://www.westhawk.co.uk/
---
To unsubscribe, mail advanced-java-unsubscribe at xcf.berkeley.edu
To get help, mail advanced-java-help at xcf.berkeley.edu
More information about the Advanced-java
mailing list