Remarkably slow

I finally got around to writing a simple benchmark to compare Java implementations. Collatz measures how quickly a Java implemenation (more specifically, an MIDP implementation) can perform basic arithmetic and function calls, by computing the first 30,000 Collatz numbers and measuring how long it takes using System.currentTimeMillis().

I also implemented the same benchmark in C for comparison, and ran a standalone version on Java2SE. Host-based benchmarks were run on a Pentium M 1.4 GHz, using Java version “1.4.2_14-b05″ and midp2.0fcs.

C	33289 iter/ms
Java SE	7990 iter/ms
MIDP 2.0	1595 iter/ms
BlackBerry 8800	240 iter/ms

It’s interesting that the hosted MIDP implementation runs equivalent code slower than JavaSE, but that could be down to a variety of factors (not least, it now occurs to me, that the benchmark inner loop has a function call which is a member of a class which extends and implements in the MIDP version, but not in the JavaSE version; now granted, it’s private static final so in theory it shouldn’t matter).
What’s really interesting, though, is that the Blackberry is six times slower than the hosted MIDP emulator, and 33 times slower than the JavaSE VM. (Never mind the 130x slowdown versus native code.) At speeds like this, I think Javascript (on host) is going to be faster than on-device Java…