Profiling Performance

This is my concise reference on profiling the execution of computer applications.

Topics this page:

Types

Profilers

Mercury Profiler

Your comments???

Site Map
About this site

Sound: Galloping horse

Types of Profiling

Profilers usually insert themselves at the byte-code level where machines execute instructions.

But the location of execution needs to be presented in a way that is comprehensible to mere mortal coders.

So profilers usually work with source code IDEs (Integrated Development Environments) such as Eclipse and Microsoft Visual Studio.

Profilers

There are several:

The screen shot at the right is from the MS.NET Visual Studio 2005 Profiler team blog

Java with Java Virtual Machine Profiler Interface (JVMPI) C JNI interface where profilers may access the state of the JVM via an in-process native agent, and can be notified of interesting events like object allocations and method invocations.

VMGEAR Optimizeit tools pinpoints performance and reliability issues throughout the development process.

The Sitraka JProbe Suite contains a Profiler with Memory Debugger, Thread Analyzer, and Coverage analyzer on BEA WebLogic.

Compuware

Mercury Profiler

Before being acquired by HP, Mercury used to offer a free download of the Diagnostics Profiler. The merc.diagnostics-profiler.zip file is 405,672 KB. When executed, it creates by default a "JavaProbe" folder within C:\MercuryDiagnostics.
To configure it, cd \MERCUR~1\JAVAPR~1\bin and run jreinstrumenter.cmd for this screen
In the "Mercury Java JRE Instrumation tool" pop-up, click "Copy Parameter" for
-Xmanagement:class=com.mercury.opal.capture.proxy.JRockitManagement "-Xbootclasspath/p:C:\MercuryDiagnostics\JAVAProbe\classes\boot"

Also add -Xrunheapdump to enable heap dumps within the profiler.
Click Exit and Paste the text in the JVM startup script file.

Run profiler.cmd
This screen should appear after it sets these environment variables:

PROBE_HOME and
JAVA_EXE=java and
PROBE_CP=
JAVA_EXE=%PROBE_HOME%\_jvm\bin\java
Closing the command window shuts down the service.

Instead of highlighting the screen (which disappears), copy from here:

-javaagent:C:\BEA\jrockit90_150_04\mercuryprofiler\lib\probagent.jar

Start the application server navigate to the "Arguments" field where JVM parameters are entered. For example, if you're using the WebLogic MedRec benchmark application click to medrec > Servers > MedRecServer > Configuration > Remote Start
Paste the string and click "Apply".

Restart the server: Navagate to "Control" > "Start/Stop" and click "Graceful shutdown of this server..." and confirm "Yes". Wait for "TASK IN PROGRESS" to turn to "STOPPED". Then click "Start" and Wait again for "TASK IN PROGRESS" to turn to "STARTED".

Click the "Connect to Mercury Diagnostics Profiler" link for this screen which graphs the Memory used, the Load, and Slowest Requests.
http://localhost:35000/profiler/ (after authenticating for default "admin" userid and password) opens an internet browser with the same (albeit slower performing) Java applet.
Mouse over the graphic for tooltip information.
Click the "Show Legend" tab for this screen
The tabs:

"Hotspots" tab graphs the "Slowest Methods" and "CPU Hotspots"
metrics
"All Methods"
"All SQL"
"Collections" require light-weight memory diagnostics instrumentation and capture to be enabled in the probe.
"Exceptions" only for instrumented methods.
"Server Requests" for each Layer
"Web Services"
"Heap Breakdown"

The profiler tracks where time is spent by each J2EE instance?
It also tracks VM heap usage, so it can rank collections by memory usage and
??? How does it identify memory leaks during GC??
It tracks hold times by sychronization lock ???
It counts exceptions and trace information ???
The J2EE Probe captures events such as method invocations, collection sites, and when business and server transactions begin and end.
Click the "Metrics" tab and expand the tree:

Profiling Strategy

The basic strategy is to identify the most time consuimg component, identify tunables, conduct tuning experiments, then move on to the next component.
So rather than immediately going to the most likely component, it's best to begin with an open mind. The value of a profiler is that it may identify issues in components where you didn't expect to cause delays.

Your rating of this page:
Low High

Your comments on this topic, please:

Publish this comment publicly

Your first name:

Your family name:

Your location (city, country):

Your Email address:

Email me updates

Top of Page

Thank you!