Saxon

Legacy ID: #6423717 Legacy Poster: Alexandre Delarge (adelarge)

Hello Mickael, After I upgraded Saxon from v8.7.3 to v9.1.0.5 in order to solve a problem on the concurrent usage of compiled XQuery objects, I noticed a severe performance degradation. To get a measure of that, I wrote a test class that executes the following set of operations: - Load 2 XML documents into DOM document instances using JDK embedded Xerces version. - Use Saxon to execute an identity transformation from DOM to String onto each DOM. - Use Saxon to apply a different XQuery transformation for each XML file, and output both to String and DOM. The reason why I use 2 XML files and 2 separate XQuery files is because I wanted to determine how big was the degradation on: - a "small job" (i.e. apply simple XQuery on a small XML file). - a "big job" (i.e. complex XQuery over a big XML file). So I measured the performances of Saxon using the following strategy: - 100 iterations of each operation are performed to "warm up" the runtime execution with class loading and HotSpot optimizations. - 1000 iterations of the same exact operation are then executed, and the total execution time is recorded. - The total execution time is the difference between 2 calls to the System.nanotime() method made before and after the 1000 iterations. At first, I executed my tests other 3 versions of Saxon: - v8.7.3 (initial version) - v9.1.0.5 (the one I upgraded to) - v9.0.0.1 (intermediate version taken from Sourceforge) While v9.0.0.1 recorded slightly better performances than v8.7.3, version 9.1.0.5 performed a lot worse... After that I ran my test again over the last v9.0.x I could find, and v9.1.0.5, hoping I could confirm the difference is clearly between the v9.0.x and v9.1.x series. Here are the results of my test, executed using the same test class and these 2 Saxon versions (I used JDK v1.6.0_10 to both compile and execute my tests): ================================================= Testing Saxon version 9.1.0.5 XQuery transformation: warm=100 ; measure=1000 Identity transformation: warm=100 ; measure=1000 XQuery 1 transform: XQuery to string total = 13383238093 XQuery to string averg = 13383238 XQuery to DOM total = 14810014402 XQuery to DOM averg = 14810014 XQuery 2 transform: XQuery to string total = 1232768131 XQuery to string averg = 1232768 XQuery to DOM total = 1228801705 XQuery to DOM averg = 1228801 Identity 1 transform: Total = 9143027675 Averg = 9143027 Identity 2 transform: Total = 461751982 Averg = 461751 ================================================= Testing Saxon version 9.0.0.6 XQuery transformation: warm=100 ; measure=1000 Identity transformation: warm=100 ; measure=1000 XQuery 1 transform: XQuery to string total = 6498402730 XQuery to string averg = 6498402 XQuery to DOM total = 6665119729 XQuery to DOM averg = 6665119 XQuery 2 transform: XQuery to string total = 115673082 XQuery to string averg = 115673 XQuery to DOM total = 95874197 XQuery to DOM averg = 95874 Identity 1 transform: Total = 9155856681 Averg = 9155856 Identity 2 transform: Total = 451998508 Averg = 451998 ================================================= Few notes about this test: - "XQuery 1 transform" and "Identity 1 transform" are performed on the "big" XQuery and XML files. - "XQuery 2 transform" and "Identity 2 transform" are performed on the "small" XQuery and XML files. Then the observation: - Both Saxon versions do produce the same exact output in term of resulting XML. - Identity transformation is slightly faster in v9.0.0.6 than in v9.1.0.5. - "XQuery 1 transform" is roughly 2 times slower on v9.1.0.5. - "XQuery 2 transform" is roughly 12 times (!) slower on v9.1.0.5. So do you have any idea why I observe such degradations between both versions ? I attach the eclipse project with the source code of my test with this post to enable reproduces. The archive also contains a copy of the XML/XQuery files I used. Thank you by advance for any feedback on that issue. Regards, Alex.