[PATCH] XmlMini - Fixed bugs and improved speed of LibXML and Nokogiri backend

When I finally set up a working Rails 3 environment, I found out that most of these bugs have already been fixed in Rails 3. However, these fixes have reduced the performance of the backend quite a bit. Moreover, I did find a new incompatibility with the REXML backend: Whitespace in the content of a tag is not preserved correctly when the tag has attributes, e.g.:

<product type="file">    </product>

I found that the Nokogiri backend has the same issue. I have attached a new patch, which fixes this issue in both backends. It actually is pretty much a rewrite of both backends. Except for some small API differences between LibXML and Nokogiri, both backends now use the exact same code to build the hash. This makes it easier to fix future bugs in both backends at the same time. Moreover, in both cases, the code is cleaner than the previous versions, and for both backends the performance is significantly improved.

Performance

Running the performance tests on the master branch:

               user     system      total        real
REXML     17.200000   0.080000  17.280000 ( 17.327603)
LibXML     2.100000   0.110000   2.210000 (  2.218409)
Nokogiri   5.290000   0.030000   5.320000 (  5.343453)

Now, with my patch applied:

                user     system      total        real
rexml      24.170000   0.110000  24.280000 ( 24.334927)
libxml      0.670000   0.000000   0.670000 (  0.674913)
nokogiri    2.390000   0.020000   2.410000 (  2.417661)

Note that I also wrote a SAX-based backend using Nokogiri, which is even faster. See my other ticket at https://rails.lighthouseapp.com/projects/8994/tickets/3636. But let's first fix the current backends before adding new ones :-)

I see I ran the benchmark with my patch applied on my less powerful notebook (see the difference in REXML performance). So the performance is even more impressive with the patch applied, compared to the master branch :-)

               user     system      total        real
REXML     17.420000   0.080000  17.500000 ( 17.617890)
LibXML     0.520000   0.000000   0.520000 (  0.535704)
Nokogiri   1.840000   0.020000   1.860000 (  1.871491)

Awesome. Plus One

Thanks Willem! To backport quickly, you can checkout 2-3-stable then cherry-pick the revision you made to master.

(from [12f6fd0f2687f083bc23ad63fdc82c7e65cb8984]) Bugfixes, speed improvements and code cleanup for Nokogiri's and LibXML's XmlMini backend

[#3641]

Signed-off-by: Jeremy Kemper jeremy@bitsweat.net
http://github.com/rails/rails/commit/12f6fd0f2687f083bc23ad63fdc82c...

I updated my benchmark and testing suite here for Rails 3 master: http://github.com/stepheneb/rails_hash_from_xml.

All my tests comparing REXML equality pass on both the Rails 3 master and 2-3-stable branches.

Running in MRI I compare Nokogiri and libxml-ruby to REXML.

In JRuby I compare the JDOM implementation to REXML.

I also tested the FFI version of Nokogiri in JRuby but all of the tests fail.

It's very nice to see the Nokogiri speedups on MRI. Here are my results:

# Ruby 1.8.6 (2008-08-11 rev 287), platform: universal-darwin9.0

$ ruby -I$RAILS_SOURCE/activesupport/lib bench_hash_from_xml.rb

              user     system      total        real
REXML     10.960000   0.270000  11.230000 ( 11.464308)
Nokogiri   1.230000   0.020000   1.250000 (  1.256476)
LibXML     0.430000   0.000000   0.430000 (  0.434689)

There are some nice speedups in JRuby also for both REXML and JDOM -- Nokogiri (FFI) is quite slow however (one reason is that Nokogiri in JRuby requires objectspace to be turned on).

$ jruby --server -I$RAILS_SOURCE/activesupport/lib bench_hash_from_xml.rb 2

# OpenJDK Server VM 1.7.0-internal:

               user     system      total        real
REXML      3.448000   0.000000   3.448000 (  3.448000)
JDOM       0.761000   0.000000   0.761000 (  0.761000)
Nokogiri   6.083000   0.000000   6.083000 (  6.082000)

For more details see: http://github.com/stepheneb/rails_hash_from_xml/blob/master/readme.txt