I've been profiling PDFBox during text extraction from some large PDF documents e.g. 2000 pages, mostly text, 20 Mb file size.
Some of these documents can take a long time to process e.g. 40s+, sometimes a lot more than that.
(I'm using a 2.5 GHz, 4 Gb, Mac OS X 10.5.8, Java(TM) SE Runtime Environment (build 1.6.0_22-b04-307-9M3263) with -Xms256m -Xmx1024m -XX:PermSize=256m)
I've begun by profiling where the code spends its time during text extraction and I see that a lot of time is spent constructing org.apache.pdfbox.util.Matrix objects.
Screenshot PDFReference_nopatch.tiff shows the most used methods in PDFBox during text extraction for a large document. When this screenshot was taken the percentages had stabilised and Matrix.<init> accounts for 40% of cpu time apparently - the largest time of any method. I was surprised.
Most of these Matrix instances are being constructed within PDFStreamEngine.prcoessEncodedText(byte)
On revision 1035639 (pre-1.4.0) this method constructs one Matrix object and then a further 7 within a loop which is called for each character in the document. So that's a lot of Matrix objects.
The attached patch refactors PDFStreamEngine.processEncodedText so that it now creates 5 reusable Matrix instances outside the loop and 2 within it.
This was achieved by adding a new method to Matrix: Matrix.multiply(Matrix, Matrix) which allows you to multiply two matrices and have the result stored in a specified Matrix object. This has the effect of reducing the number of temporary Matrix objects created during multiplication within PDFStreamEngine. This should save the garbage collector some work.
I profiled PDFBox again with this patch included and Matrix.<init> now accounts for only 30% of the cpu time.
Unfortunately, whilst less temporary objects are being created, it doesn't have an appreciable effect on the time it takes to extract text from my large documents.
The profiling continues...