I needed an XML parser for my tests with Rust (parsing a gigantic 2Go Wikivoyage XML dump) and there is no native one. So I would need a wrapper around a C implementation. At first, I used libxml2, because it has a very appealing name it has a very convenient XmlTextReader API where you are controlling the parsing loop (no callback like in SAX).

I wondered how it compared to the StAX API of Java, and made a simplistic test printing (to /dev/null) the names of all the nodes. And surprise, StAX was way faster:

• the StAX (openjdk-7, Ubuntu 14.04, amd64)
  • 35 seconds
  • Memory usage: 30 Mo
• XmlTextReader API of libxml2 (2.9.1), g++ 4.8.2
  • 50 seconds
  • 3.5 Mo

Then I made tests with SAX, comparing the default Java implementation with libxml2 and finally also libexpat 2.1:

• Java: 30s (39Mo)
• libxml2: 34s (2Mo)
• libexpat: 21s (180ko)

Conclusions:

• the Java implementation seems quite good (at least in this particular scenario)
• libexpat is quite fast, but it’s SAX and not pull-parsing (which is very convenient)
• if you are processing huge XML files, forget about libxml2