Page history Edit this page How do I edit this website?

Profiling Java Code

If your Java code seems to have bottlenecks that make it run slow, you need to profile the code. There are a couple of options you have for that.

JProfiler

JProfiler is a commercial program, but you can evaluate a trial version.

HPROF

Use HPROF, a profiler included in Sun’s Java since version 1.5.

VisualVM

Use VisualVM (a tutorial can be found here), which is a graphical alternative to HPROF.

See also Rejeev Divakaran’s detailed instructions on memory profiling with VisualVM.

OProfile

If you are on Linux, you can use OProfile, a low-overhead profiler which uses a kernel module to minimize its impact.

You need to recompile OProfile for Java support (at least on Ubuntu, there is no precompiled OProfile package with Java support). On Linux, you can use this shell script to compile and install it:

sudo apt-get install libpopt-dev binutils-dev checkinstall && 
git clone git://oprofile.git.sourceforge.net/gitroot/oprofile/oprofile && 
cd oprofile && 
./autogen.sh && 
./configure --with-kernel-support \
            --with-java=/path/to/fiji/java/linux-amd64/jdk1.6.0_24/ &&
    make && 
    sudo checkinstall

You should make sure that the path to the JDK is correct. (But you do not need to use checkinstall, but it will make it easy to remove the package once you do not want it anymore.)

Start the profiler

sudo opcontrol --start --no-vmlinux

Recent versions of OProfile require that a user account oprofile with default group oprofile exist before starting the OProfile daemon. On current Ubuntu systems, you can make one by calling sudo adduser oprofile (Ubuntu creates a default group of the same name for each new user account). If you do not have such an account, you will get something like anon (tgid:10014 range:0x100000-0x103000) in the report.

Start Fiji with support for profiling

fiji -agentpath:/usr/local/lib/oprofile/libjvmti_oprofile.so --

Stop the profiler

sudo opcontrol --stop

Getting at the profiling data

sudo opcontrol --dump && 
opreport -l image:\*/fiji

If you get entries like “anon (tgid:10014 range:0x100000-0x103000) you probably did not create a user account oprofile in a group oprofile before starting the OProfile daemon.

To get information about source files and line numbers, also pass the -g option to opreport.

Further reading

See the OProfile manual for more information.

Shark (for macOS)

You might find Shark 4 useful if you’re on macOS.

Memory profiling

Even if Java’s memory management prevents most memory issues (unaligned writes, access to uninitialized/released memory), there is a chance of memory leaks: constant accumulation of objects over time, most likely because there are stale references to them.

Note: heap is Java speak for memory.

Javassist-based

A quite versatile method is to use (and possibly modify) the class fiji.MemoryProfiler in fiji-compat.jar.

This memory profiler instruments all method entries and exists using javassist. At each exit, it reports the relative memory usage, the total memory usage, and the exit point of the current method. Call it like this:

./fiji --main-class fiji.MemoryProfiler fiji.Main

Since the memory profiling slows down execution dramatically due to the synchronous output to stderr, you may want to limit the classes to be instrumented by setting the environment variable MEMORY_PROFILE_ONLY to a space-delimited list of classes.

If you want to instrument any class handled by imagej-legacy, you need to use the slightly more complicated command line:

./fiji -Dpatch.ij1=false --cp jars/javassist.jar --cp jars/fiji-compat.jar \
	--cp jars/ij.jar --main-class fiji.MemoryProfiler -- ij.ImageJ

Using JVisualVM

JDK6 and newer come with a quite useful tool called jvisualvm. It is in the bin/ directory of the Java Development Kit (it does not come with the JRE, also known as Java Runtime Environment).

JVisualVM lets you attach to any local Java process and inspect the state. The most useful parts for memory profiling are the Memory section in the Monitor tab, showing the overall heap usage (where you can also make a Heap Dump which produces a double-clickable list of objects and their references), and the Memory button on the Sampler tab, which opens kind of a live version of the Heap Dump.

When looking at a Heap Dump, make sure you click on the Classes button in the top row. This shows the list of classes that currently have referenced instances in the Java Virtual Machine. Double-clicking on the class name will open a detailed list of instances of that class, and when clicking on one item in that list, the upper right panel will show that instance’s fields while the lower right panel will show the objects holding references to that particular instance.

Note: at no time will JVisualVM stop the virtual machine it is attached to, so if the issues you experience are time-sensitive, you definitely want to put stop-gaps into the code to be able to look at things closely.