Parallel Programming Systems For Workstation Clusters

. In this paper, we describe experiments comparing the communication times for a number of different network programming environments on isolated 2 and 4 node workstation networks. In addition to simplified benchmarks, a real application is used in one of these experiments. From our results, it is clear that the cost of buffer management at either end of the communication is more important than originally expected. Furthermore, as communication patterns become more complex, the performance differences between these environments decreased substantially. When we compared timings for an actual application program, the differences essentially disappeared. This shows the danger of relying solely on simplified benchmarks. Key words. Parallel computing, Communication, C--Linda, P4, POSYBL, PVM, TCGMSG AMS(MOS) subject classifications. 65Y05 1. Introduction. A number of programming environments exist that make distributed computing available to the application programmer. While it is natural..