This work presents a performance evaluation of the communication system supported by "Parallel Virtual Machine" (PVM) over Linux. This study purports evaluation to demonstrate the viability of distributed parallel processing implementation through PVM in low cost computing environments. Distributed parallel processing may provide a better use of the computational features in these environments and, as a consequence, an increase of computational power. Thus, tasks that demand great computational power, long processing time or that are impossible to be executed in only one machine, could be scheduled to be executed when the computers are not being used.

The evaluation of PVM communication system consists of measurement of throughput in the memory system, communication via sockets using the TCP and UDP transport protocols, and in the communication via PVM in the studied environments.

A description of the inter-processes communication mechanism for distributed applications and Ethernet technology is presented. A revision of computer architecture classification, with emphasis in the conceptualization of parallel machines, is presented. Some basic properties of parallel programming are discussed. PVM components and the way communication is implemented in PVM are also described.

Measurements made on throughput in the memory systems, in the sockets communication, and in the PVM communication are shown and compared in detail. The data from the experiments show that communication via PVM is aproximately 28% less efficient than using TCP or UDP sockets with 1Kbyte messages. For 32 Kbytes messages the performance loss with PVM is close to 16%. The results also indicate what set of parameters is most influential in achieving the best possible performance. A set of recommendations is presented so that a programmer might tune his/her application code to be able to achieve high performance in communication via PVM.