Re: [Proxool-developer] JDK independence

[Christian N. <chr...@em...>]

[Hmm. SourceForge lists seems to have been down. Trying again...]


An ICQ discussion with Bill yesterday ended in these conclusions:

The Timer and ShutdownHook scenarios can be handled without bytecode
manipulation.

For the Timer we will check runtime wether the Timer class is
available or not. If not available we will use our own timing code.
The Timer will only be accessed through reflection so that our source
will compile under JDK 1.2.

The same strategy will be applied to the ShutdownHook except that we
will not offer any alternative implementation for 1.2. We'll simply
document that 1.2 users will have to take care of the shutdown
expilcitly.

The Proxy scenario is a bit more complex. I have evaluated the
following libraries in conjuction with this:

ASM (http://asm.objectweb.org/)
This proved to be a very low level library that requires a good
knowledge of the Java byte code instruction set. It is very small and
fast, but unless anybody wants to dive into the JVM spec I think its
going to be to complex to use.

Javassist (http://www.jboss.org/index.html?module=html&op=userdisplay&id=developers/projects/javassist)

Unlike ASM this library offers a very high level api for class
manipulation. It enables you to create and load classes runtime, and
can even compile and add methods to classes from strings containing
source code (!). It's a medium size lib (about 350k). It does not
come with a ready made Proxy replacement, but we can probably roll
our own with a bit of effort.

CGLIB (http://cglib.sourceforge.net/)
This uses ASM as a backend and comes with a ready made replacement
for Proxy. The lib size is (145k). I did a rough benchmark between
CGLIB and the JDK Proxy. CGLIB added a slight overhead compared to
JDK Proxy calls, but it was very tiny. A million method calls to a
simple string getter took about 550 milliseconds using the JDK Proxy,
and about 650 milliseconds using CGLIB. I'll do the CGLIB test with
1.2 as well (the test was performed on 1.4.2) to make sure that the
overheasd is not significantly larger there.

I guess the best solution would be to use a similar approach to that
mentioned above: Use Proxy for 1.3+ and CGLIB for 1.2. The problem
here is that I have not figured out how to make a class implement the
InnvocationHandler *runtime* with CGLIB. This is a simple method call
in Javassist. I'll investigate further...

Another compelling approach is to simply stop using the JDK Proxy and
use the CGLIB Proxy all over. If we go that route we can throw away
the 1.2 source tree today...

I'll do a bit more investigation and keep you updated.

CHR




> On Fri, 2003-09-05 at 14:45, Bill Horsman wrote:
> 
> > Hi,
> > 
> > We're looking at replacing our use of JDK 1.3's proxy
classes with those 
> > from http://asm.objectweb.org. The big benefit we
get, other than 
> > possible performance improvements, is independence from
using JDK 1.3. 
> > Our goal is to continue to support JDK 1.2 for a good while
yet.
> > 
> > The trouble is that we also make some other references to
JDK 1.3+. Namely:
> > 
> > 1) ShutdownHooks. These ensure that Proxool is properly
shutdown when 
> > the JVM dies. It works really well and is much more robust
than using 
> > finalize().
> > 
> > 2) We use a Timer class to collect statistical information.
We can get 
> > round this with either a thread that sleeps for a while and
keeps 
> > checking the time. Or possibly look around for a third
party solution.
> 
> I've been thinking about this. We could use the new proxy to
handle both
> the Timer and the ShutdownHook too. For instance, for the Timer
we write
> and interface that Timer is able to implement and then write our
own
> JDK1.2 implementation. At runtime, we try and load the JDK1.3
Timer and
> if that fails, use our own implementation instead. The proxy
classes are
> used to get JDK1.3's Timer to implement our interface at
runtime. I
> think that would work.
> 
> Advantages: 1) We get to use the (probably) more efficient JDK1.3
> classes if available. 2) Our code compiles with all JDKs.
> 
> Disadvantages: 1) It might be a bit slower (but this doesn't
matter
> because both ShutdownHook and Timer are only constructed at
startup).
> 
> Christian, if you are working on the proxy stuff over the next
few days
> then you might want to think about this).
> 
> Cheers,
> Bill

-- 
__________________________________________________________
Sign-up for your own personalized E-mail at Mail.com
http://www.mail.com/?sr=signup

CareerBuilder.com has over 400,000 jobs. Be smarter about your job search
http://corp.mail.com/careers