【转贴】Five ways for tracing Java execution

I'm often faced with a situation where I need to dig into code that I didn't write. Most of it is poorly documented and, in some cases, I only have some of the sources, which means I need to decompile the rest. It's a tedious exploration. There are mainly two reasons for that: maintaining a code or using it. The later refers to 3rd party frameworks and class libraries I need to build upon. For example I may implement an interface of anevent handlerand I may be uncertain regarding which event fires when and what are the exact arguments I'm getting.

I usually find that it is faster to trace the code at runtime. Especially when it comes to non-trivial links between classes, likean interface with multiple implementationswhich may be picked at runtime. The smörgåsbord of design time tools is just insufficient.

In this post, I will summarize the common methods (I know of) for tracing the runtime execution. The tips and examples are assuming you're usingEclipse,but most of them can be achieved with other IDEs as well. I will provide some more detailed information regarding some of the techniques mentioned below in future posts.

The Basic: Breakpoints and Step-by-step Execution

Start with the simplest way: set upbreakpointsand start tracing your execution.

It is best when:You need a quick and simple solution, you have all the code and you know where you want to stop. You need elaborate information in a given point (arguments, local variables, etc.).

My tip:Even if you don't have all the sources, it's possible to break when entering methods or classes by setting-up method breakpoints or class loading breakpoints.

• Class loadingbreakpoints are simple to set up: from Eclipse Run menu, select"Add Class Load Breakpoint...".
• In order to set method breakpoints, you will need to decompile the class file first. Use theJadClipse plug-infor that. After you decompile, you may select the method and choose"Toggle Method Breakpoint"from the "Run" menu (or by right clicking the left bar in the file next to the method header). Note that you may set regular breakpoints in decompiled sources as well, but these will not work.

The Primal: Debug messages

We continue by setting up debug messages. The simplest way is to useSystem.out.printlnstatements to print out messages to the console. The next step would be to use a logging mechanism: the basicJDK logging,Apache Commons LoggingorLog4j.

It is best when:You own the code and you have a good idea of what you're looking for. Very good solution for event handlers. It's high performance makes it practical for understanding which event is fired when throughout the execution of complex flows.

My tip:It's easy to set up one generic method that prints out the method currently executing. The best way I'm familiar with is to get thestack trace. The last entry on the call stack will be the debug message method. The entry before that will be the method invoking it. How to obtain a stack trace:

• Since JDK 5.0: UseThread.currentThread().getStackTrace().
• Before (JDK 1.4): create a new exception instance and usegetStackTraceto obtain your call stack from the new instance (no need to "throw" it).

The Hot Shot: Dynamic Proxy

An improvement over simple debug messages.Dynamic Proxyis a special Java feature which enables the developer to introduce a proxy class, sitting in front of a given class and intercepting all the calls through a given interface. Once you intercept the calls, you may print out appropriate debug messages with all the information you require.

It is best when:It's a great solution for event handlers. You can set up a dummy event handler with a generic proxy in seconds and see the sequence of events. This is simplest and quickest method when it comes to understanding event handlers.

My tip:You can find a code for a generic dynamic proxy onSun's site. It's OK, but it has one small minus: it only implements the immediate interfaces implemented by the class of the instance being wrapped by the proxy. It does not implement any interface of its' ancestors. This can be easily fixed by traversing the list of ancestors and picking all the interfaces along the way. In one of the future posts I'll give a better version.

The Brute Force: Run-time Profiler

Profilersare powerful tools that trace all the calls in the system through special JVM hooks. However, it's like using a 10 pound hammer on a half inch nail.

It is best when:You want a complete picture for a very specific operation (i.e. very short execution flow).

My tip:The good tools cost around $500 per seat. Most free tools are very basic. IMHO, The best free open-source tool for the job isEclipse Testing and Performance Tools Platform (TPTP). I've had very good experience with the tool and I highly recommend it. It can  provide asequence diagramwhich is one of the best ways for understanding execution. TPTP used to be my favorite before I switched to the Mac and found it's incompatible (maybe I'll try to fix it if I'll have some spare time).

The New Age: Aspects

Aspect Oriented Programming (AOP)is a non-trivial idea. Without going into the concept of Aspects, I'm just looking at the bottom line: it's a quick and easy way of intercepting the execution of your code. You can selectively set hooks around methods, constructors, field access, etc., without modifying the original code. In these hooks you can print debug messages.

It is best when:You want to trace the execution of a code you can rebuild.

My tip:Use it. AOP is still far from the mainstream. It is uncertain whether it will ever get there. Without debating its' qualities and features, most developers can be up and running with AOP based logging in about an hour. I'll show the exact steps in one of the future posts. For Eclipse users, I recommend theAspect Java Development Toots (AJDT).

Conclusion

What method should you choose? I'll divide this into two different cases: tracing the complete execution of a short sequence and tracing events to/from a given class or tier.

When tracing the complete execution:

1. A profiler is the simplest way. IMHO, every developer should be able to operate a profiler and have one readily available in the arsenal.
   
2. The next choice would be AOP. It's the second choice only because setting up a profiler is simpler when it comes packaged JARs. However, AOP gives more control. For example, I can see the method arguments and print them.
3. Use debug messages.

When tracing events:

1. Use AOP. Selectively capture just the interesting calls.
2. Use a Dynamic Proxy. It's like using debug messages, but a more elegant solution. AOP may be more complex first, but in the long run I would go for AOP.
3. Use debug messages.
4. Use breakpoints.

As you can see, I'm not a big fan of breakpoints in this context. Breakpoints are great when debugging your code, but not for tracing execution. There are far more efficient methods of doing that.