Benchmarking HPJava, Part II:
Performance on Parallel Machines

In this chapter we will benchmark the HPJava programs introduced in chapter 6 on parallel machines. In chapter 6 we have seen that HPJava produces quite efficient node code, and also that our HPJOPT2 optimization scheme dramatically improves performance on a single processor. With successful benchmarks on one processor, we expect the HPJava system to achieve high-performance on parallel machines. The main issue of this chapter is to observe performance of HPJava programs on the parallel machines. It won't be necessary to critically analyze behaviour and performance of HPJava programs on these machines. We are reasonably confident that good results for node code will carry over to multi-processors because of the way that translation and optimization works. In addition to HPJava and sequential Java programs, we also benchmark C and Fortran programs on each machine to compare to the performance. For now the C and Fortran programs are sequential programs, not parallel programs. To assess performance on multi-processors, the benchmarks will be performed on the following machines:

• Shared Memory Machine - Sun Solaris 9 with 8 UltraSPARC III Cu 900 MHz Processors and 16 GB of main memory.
• Distributed Memory Machine - IBM SP3 running with four Power3 375 MHz CPUs and 2 GB of memory on each node.

The table 7.1 lists the compilers and optimization options used on each machine. This chapter is based on our recent publications such as [31,30]. The benchmarks on SP3 are from [32]. It experimented with only naively translated HPJava programs.

Table 7.1: Compilers and optimization options lists used on parallel machines.

	Shared	Distributed
	Memory	Memory
	Sun JDK 1.4.1 (JIT)	IBM Developer kit
HPJava	with -server -O	1.3.1 (JIT) with -O
	Sun JDK 1.4.1 (JIT)	IBM Developer kit
Java	with -server -O	1.3.1 (JIT) with -O
C	gcc -O5
Fortran		F90 -O5