Towards an automatic portation of programs from nCUBE2 to KSR1

Abstract

Several programs have been parallelized and implemented both on an nCUBE2 and on a KSR1 system, two parallel computers of different architecture, the first being of message-passing type, the latter a global-shared-memory machine. Our experience with the two parallel systems and a comparison of the implemented test programs have led to the conclusion that the work which must be invested to map a parallel algorithm onto the parallel constructs of the respective machine is of similar amount and even of similar kind for both systems. Our findings induced us to derive a strategy for the portation of programs from nCUBE2 to KSR1, which could moreover be automated and is likely to work for other message-passing and global-shared-memory systems, too.