Due to Eddie reproducing this problem and capturing a heapdump, we tracked this down to something quite simple: The Document Analyzer code has an instance field which holds the produced analysis engine. This particular analysis engine loads 300 MB of stuff (the language model) during its initialization, which happens when you do the produceAnalysisEngine call.

So the line which looks like:

ae = UIMAFramework.produceAnalysisEngine(aggDesc);

causes a temporary condition where the "ae" is holding on to 300MB from the old version of the analysis engine, while the new one is trying to initialize - so you run out of memory.

The differences between Javas comes from the details of how the String class is implemented - in the IBM 5 Java each string takes more storage than in the Java 6.

-Marshall

Eddie Epstein wrote:
"IBM 32-bit SDK for Windows, Java 2 Technology Edition, Version 5.0". I
found "SR3" mentioned in one file. The jar files are dated 2/1/2007.

On Feb 7, 2008 2:28 PM, Thilo Goetz <[EMAIL PROTECTED]> wrote:

Eddie,

so which JVM did you use to produce the problem?  I understand
1.6.0_02 fixed it, but which one didn't work?  Thanks.

Eddie Epstein wrote:
Hi Marisol,

After installing opennlp on Apache UIMA v2.2.1, I was able to reproduce
the
problem using OpenNLPParser with both documentAnalyzer and cpeGui.
However,
I had some problems reproducing this problem between an environment with
the
2.2.1 SDK jars vs. a development environment using the latest code in
svn.
Turns out the difference between working and not working was the JRE.
JRE=
jre1.6.0_02 continued working on repeated runs.

Hope this can help you,
Eddie

On Feb 5, 2008 1:13 PM, Marisol B. Gamboa <[EMAIL PROTECTED]> wrote:

I'm having a problem reallocating memory after
CollectionProcessingEngine
process completes.

The cpeGui seems to have the same problem.  I'm loading the OpenNLP
models
into a CPE for processing documents and if I want to setup another
process
without killing the GUI or the JVM the memory form the last CPE is
never
garbage collected.

Has anyone encountered this and is there a fix?

Marisol



Reply via email to