Capability Comparison between Japanese HPF compilers using Scientific Applications

T. Mizuno and H. Sakagami
Computer Engineering, Himeji Institute of Technology
takao@comp.eng.himeji-tech.ac.jp

H. Murai
1st Computer Sowtware Devision, NEC Solutions
murai@hpc.bs1.fc.nec.co.jp

Abstract

The HPF was defined as a standard language that described parallel paradigms, but there are many incompatibilities depending on the vender implementations. This situation forces difficulty on developing portable programs and prevents the spread of parallel computing.

JAHPF has been aiming at preserving the compatibility between Japanese HPF compilers and makes it possible for users to program parallel codes easily and portably. JAHPF also defined a set of HPF extensions as HPF/JA to enhance capability of HPF and is focusing on developing benchmark programs that reflect real-world scientific applications.

In this paper, we checked Japanese HPF compilers that were implemented at NEC SX-4, Fujitsu VPP-800 and Hitachi SR-8000. The architectures of above parallel computers are different, namely SX-4 consists of nodes which contain multiple vector-parallel processors and an internode communication is much slower than an intranode transfer, VPP-800 is also a vector-parallel type but all processors are connected with crossbar switches uniformly, SR-8000 has RISC CPUs and an internode communication is also much slower than an intranode transfer. Users, however, want to run the same code on different machines as fast as possible. We have performed the same source codes for scientific applications on there machines and compared capabilities of Japanese HPF compilers .