WWW http://www.math.utah.edu/~beebe

Just-in-time compilation improvement on the Linpack 100x100 benchmark

Last update: [25-Jun-1998]

Original version: Tue Jun 30 12:36:21 1998
Last update: Fri Nov 12 15:52:51 2004

The following tables were computed using the Java version of the Linpack 100x100 benchmark program. For each machine tested (all at the author's site), debugging vs. no-debugging, and just-in-time compilation vs. byte-code interpretation. This gives four combinations: each was run 16 times, and the results were averaged. In some cases, the Mflop ratings varied by as much as a factor of two in these runs, so the averaging is essential.

The results were collected in a spreadsheet, where the last two columns are computed from the ratios of earlier column pairs, and the result then saved and converted to HTML for this file.

Local copies of the Java program, a UNIX Makefile for running the benchmarks, and the spreadsheets in a variety of exchange formats, are available here. Contributions of results for additional machines are welcome, and can be sent to beebe@math.utah.edu. If you run this benchmark on other Sun models running Solaris 2.x, you can use the /opt/SUNWspro/bin/fpversion program to find out the chip name and clock rate.

The table rows are sorted first by O/S, and then in descending order by the JIT Mflops values.

The Java implementation in Sun Solaris 2.5 does not offer just-in-time compilation, but Solaris 2.6 does, and the effect is dramatic. Notice the SPARC 20/512 results, which are provided for both versions of the operating system.

Compilation for debugging

Normal compilation