Calling java from RPG - how to increase performance

Hi,
 
I have written a RPG service program that transforms XML [using transform
stylesheets (XSLT)] into SQL insert statements. These are then executed over
the database on the iSeries.
 
To do this I utilise the Java Native Interface (JNI) to create and use java
objects. Basically, I create a transformer object. I then pass the object an
XML stream source to transform, and a stream result to hold the transformed
data.
 
For each transformation (assuming the transformer object does not change) I
have to create a new StreamSource object:
 
To create the stream source object from RPG I have to perform the following
steps:
 
1) Create a String object from the XML file path. (RPG-to-Java)
2) Create a File object from the String. (RPG-to-Java)
3) Free the String object - no longer needed. (RPG-to-Java)
4) Create a StreamSource object from the File object. (RPG-to-Java)
5) Free the File object - no longer needed. (RPG-to-Java)
 
I create a StreamResult object for each job that runs this process, so I
only perfrom the following once:
 
1) Create a String object from the result file path. (RPG-to-Java)
2) Create a File object from the String. (RPG-to-Java)
3) Free the String object - no longer needed. (RPG-to-Java)
4) Create a StreamResult object from the File object. (RPG-to-Java)
5) Free the File object - no longer needed. (RPG-to-Java)
 
Now, this all works fine - I'm just not sure it's running as fast as it
could be - approx 0.2 seconds for each transformation. After a few hundred
transformations it seems to slow down - could this be the Garbage Collector?
I start the JVM manually using JNI (with version 1.4.2 set). I also set the
classpath manually (using putenv). The XML file path is retrieved from a
data queue, and the result file path is created by the program. 
 
So, what are the best ways to speed things up???
 
I've considered the following:
A) Create a custom java class that performs the conversions from file path
to StreamSource/StreamResult and just call it with the two byte arrays.
B) Allocate more memory to the JVM - what's the best size? What's the
default size?
C) Run the whole process in java - I could execute the SQL using JDBC. (but
would that just be even slower???)
D) I've not compiled the XML jar files (Xerces.jar, Xalan.jar, etc) because
I believe the JIT compilers are better than CRTJVAPGM. Is this true?
E) I've not copied and compiled jt400.jar to put in my classpath - should I
need to?
 
Basically, does the JNI carry a large overhead? Should I minimise the number
of RPG-to-Java calls by writing a custom class? Should I optimise the java
environment by allocating memory or compiling objects? I need your wise
words.
 
The program source is available if it helps. (We're running at V5R2M0)
 
Thanks
 
Larry Ducie