Benchmarks

For purposes of this paper--and to understand what directions are most profitable for future development--the simplest and most practical kinds of optimization strategies have been applied ``manually'' (but we believe honestly--i.e. it seems straightforward to apply these strategies automatically, with a modest amount of extra work on the translator) to the HPJava matrix multiplication programs. The matrix sizes are 100 by 100. The results use the IBM Developer Kit 1.3 (JIT) with -O flag on Pentium4 1.5GHz Red Hat 7.2 Linux machines. We also compared the sequential Java, C++, and Fortran version of the HPJava program, all with -O or -O5 (i.e. maximun optimization) flag when compiling.

Table 1: Matrix multiplication performance. HPJava code for single processor. Other codes sequential. All speeds in MFLOPS.

		Strength	Loop
HPJava	Naive	Reduction	unrolling
HPJava	125.0	166.7	333.4
Java	251.3	403.2	388.3
C++	533.3	436.7	552.5
F77	536.2	531.9	327.9

Table 1 shows that the strength reduction optimization helps in the Java case (for C++ and Fortran, the compilers can be expected to implement similar optimizations, and we don't see any improvement from the source-level ``optimizations''). The main message is that, with a little help in the form of source-to-source transformations, Java can get 70-75% of the performance of C++ or Fortran. After similar optimizations, HPJava with more complex subscripting expressions has similar but marginally slower performance. Of course we expect that the HPJava results will scale on suitable parallel platforms, so a modest penalty in node performance is considered acceptable.