Description: I'm utilizing the command "libreoffice --headless --convert-to pdf --outdir ${outputFolderPath} input.docx" to convert DOCX files to PDF format. However, I've noticed that the conversion process significantly slows down for files exceeding 1 MB in size. Previously, on our production server, the conversion of a 20 MB file to PDF typically took around 3 to 4 minutes. However, recently, the process seems to be taking much longer than usual. Steps to Reproduce: 1. Open the terminal and connect to the Linux server and past this command "libreoffice --headless --convert-to pdf --outdir resultFolderPath input.docx" 2. For input.docx provide any docx file path. 3. Run the final command and check for results. Actual Results: size time pages 1.2mb 1ms 9 4.5 mb 230 ms 11 4 mb 227 secs 1195 10mb 452 secs 392 <1mb 1ms 1 Expected Results: The conversion time should be less Reproducible: Always User Profile Reset: No Additional Info: Please resolve this ASAP as I have production deployment planned.
Hello, I think you should include DOCX examples in your issue. And you could add some precision in the title of your issue, at least add the keyword "Performance issue" or something like that :) I am also interested in solving this problem. I will share my research in the next message. Thank you.
For me, LibreOffice could be 10x times faster :) I spent a few days last year researching this subject. Here are my conclusions: Most of the time, especially with long documents containing very large tables, the main bottleneck is here: - File: dev/core/svl/source/items/stylepool.cxx - Function: Node* Node::findChildNode From my understanding, this function is called each time a new "style" is parsed from a part of the document (a paragraph, a table cell, etc.). The goal of this function is to find an existing style among already parsed styles. If not found, it creates a new style node. Problem, this search operation is extremely costly because: - Comparing styles is slow, often requiring the comparison of complex heterogeneous structures. - It performs an O(n) search using a simple for-loop. This loop is executed for each parsed style in the document, making it effectively O(n^2). The longer the document, the greater the search cost. I don't know the LibreOffice code well enough to optimize this myself. Ideally, I would like to create a short hash of each style node, with a hash index to find an existing style. I am also wondering about the impact of creating a new Node style instead of reusing an existing one to avoid the search. We are willing to pay a few thousand euros to a company that can solve this problem. Please contact me privately for this. I would also like to help solve this problem 🙏 Thank you.