Christian Simms's Software R&D http://simmsland.com/blog/2003/May/31/xml Generative Programming and other R&D. en In my spare time, I've been developing a J2EE code generator using python. http://simmsland.com/blog/2003/May/31/xml/2003/May/31#perf-test However, I noticed that as I added new features, the performance was degrading considerably, so I decided to do some benchmarks. This is what I found. <p/> Below, each column name is a link to the source code for the given language. Each program is written in the style for that language, so for instance the Python version uses the built-in dictionary for hash tables, whereas the Java version uses HashMap class. Notice how these tests don't do something simple like test using hashes of int's, because I haven't had to do that kind of thing since college. <p/> <table border="1"> <tr> <td> Test </td> <td> <a href="/blog/perf.py">Python</a> </td> <td> <a href="/blog/perf.java">Java (classic) </a> </td> <td> <a href="/blog/perf.java">Java (hotspot) </a> </td> <td> <a href="/blog/perf.cs">C# </a> </td> <td> <a href="/blog/perf.cpp">g++ </a> </td> <td> <a href="/blog/perf.cpp">g++ (-O3) </a> </td> </tr> <tr> <td> Dynamic Object Allocation </td> <td> 141,000/sec </td> <td> 168,000/sec </td> <td> 1,400,000/sec </td> <td> 52,300,000/sec </td> <td> 1,060,000/sec </td> <td> 1,450,000/sec </td> </tr> <tr> <td> Dynamic Object Allocation plus hash table get/set </td> <td> 9,880/sec </td> <td> 4,670/sec </td> <td> 56,300/sec </td> <td> 53,800/sec </td> <td> 19,100/sec </td> <td> 30,200/sec </td> </tr> </table> <p/> So what can we learn from this? <ul> <li> Java with hotspot compares favorably to C# </li> <li> Java with hotspot is something like 10x faster than without </li> <li> Python's about 5x slower than Java/C# </li> <li> And most surprising (to me): Java and C# were both about twice as fast as C++! I had read that could be true, but seeing is believing. </li> </ul> I tried speeding up the Python code using the <a href="http://psyco.sourceforge.net/index.html">Psycho</a> JIT compiler, but without much luck. Actually, it was slightly slower (maybe 10% slower), and I think I know why. The Python versions of these tests primarily test the memory allocator and built-in dictionary type, neither of which can be JIT'ed because they're already native in the Python language, so I just saw the overhead of the JIT).