Using instruction level parallelism to characterize Java benchmarks
The efficiency of a Java program under different Java runtime
environments (JREs) can be different on different superscalar machines. We examined selected SPEC JVM98 benchmarks running under three different JREs, namely, interpretation, Just-In-Time (JIT) compilation, and native compilation. We used the Shade toolkit to simulate several multiple issue machine models. Our focus was on performance degradation from mispredicted conditional branches and indirect jumps; we found that the degradation can vary significantly across different JREs and benchmarks. These results could help to determine the hardware components that we should emphasize when designing a processor to support a specific JREs.
Bill Hsu is an associate professor of computer science at San Francisco State University. His research interests include computer music, computer architecture and performance evaluation.