Overall Performance

 

Table 1: Execution time, in seconds, of the benchmarks using each JVM. The ratios on the right compare the JVMs pairwise, showing the slower JVM's execution time relative to the faster one's.

	Benchmark	JDK	Jupiter	Kaffe	Jupiter/JDK	Kaffe/Jupiter
1	209_db	178s	282s	836s	1.59:1	2.96:1
2	228_jack	112s	213s	567s	1.91:1	2.66:1
3	201_compress	333s	700s	2314s	2.10:1	3.31:1
4	222_mpegaudio	276s	649s	1561s	2.35:1	2.40:1
5	213_javac	114s	313s	733s	2.74:1	2.35:1
6	202_jess	93s	257s	608s	2.76:1	2.36:1
Geometric Mean					2.20:1	2.65:1

 

To test Jupiter's functionality and performance, we used it to run the single-threaded applications² from SPECjvm98 benchmark suite [18]. In this section, we present the execution times consumed by these benchmarks running on Jupiter, and compare them with results from Kaffe 1.0.6, and from the Sun Microsystems JDK v1.2.2-L. We find that Jupiter is faster than Kaffe and slower than JDK.

Table 1 compares the execution times of each benchmark run on the three JVMs. All times were measured on a 533 MHz Pentium III with 512 MB of RAM running Linux, kernel version 2.2.19. Jupiter was compiled with gcc version 2.95.2 at optimization level -O3, with all source code combined into a single compilation unit to facilitate function inlining [16]. The times were reported by the UNIX ``time'' program, and therefore include all JVM initialization. All benchmarks were run with verification and JIT compilation disabled, since Jupiter does not yet possess either of these features. Averaged across all benchmarks (using the geometric mean), Jupiter was 2.20 times slower than JDK, and 2.65 times faster than Kaffe.