opt. week

Happy new year (maybe in Chinese context).

About a week and more ago Miguel told me that there is summary page for performance comparison for Java, Mono and .NET by Jeswin.P.

It is always nice to see that someone is kind enough to provide such information :-) The numbers however are not good for Mono, so I kinda started attempt to enhance mono XML performance.

Miguel told me that I can't know how those performance improvement record is impressive to some users (well, I totally agree; I was not surprised to see the perf. results). So I started to record that XMLmark results, now using NPlot (for Gtk# (0.99)). You can see sort-of-daily graph, like:

As you see, this measuring box has some errors, but it won't matter so much. Anyways, not bad. The most effective patch was from Paolo.

The next item I want(ed) to optimize was Encoding. I have already done some tiny sort of optimization in UTF8Encoding (the latest svn should be about 1.2x-1.3x faster than 1.1.13), but it needs more love. It is ongoing.

The best improvements I have made was not any of the above. I had been optimizing my RelaxngValidatingReader by feeding OpenDocument grammar which is about 530 KB and some documents (I extracted it and converted to .odt using OO.o 2.0 and extracted content.xml).

At first I couldn't start from that massive schema. I had to fix several bugs. I spent most of hacking time on it last winter vacations. It became pretty solid. Those new year days I started to validate some OpenDocument content.

I started from one random pickup of OASIS docs, maybe this one, which was about 70KB, 1750 lines of instance. At svn r54896 it almost stopped around line 30: too bad start. I started to read James Clark's derivative algorithm again, and started several attempt to optimize it. After some step-by-step improvements, at r55497 the validator finally reached to the end for the first time. I still needed further optimization since it cost me about 80 seconds, but one hour later it went nearly 20x faster, at r55498.

After having 70KB document validated within 5 seconds, I thought it was time to try the biggest one - OpenDocument 1.0 specification. I added indentation to this 4MB of content.xml and it became 5MB. After fixing XML Schema datatype bug that was found at line 4626, I had to make further optimization, since it was eating up the memory, looked almost infinite. I dumped the list of total memory consumption for each Read(), but the fix (r56434) was stupidly simple. Anyways finally I could validate 5MB of an instance in 6 seconds.

So, now it is possible enough to validate OpenDocument xml using RelaxngValidatingReader, for example with AODL, An OpenDocument Library.

Maybe it's time to fix remaining 8 bugs found by James Clark's test suite - 7 are from XLink 5.4 violation, and 1 is in unique name analysis.

(sorry but no plot for RELAX NG stuff, it is kind of non-drawable line ;-)