Collective Communications for Parallel Processing in Networks of Workstations

摘要

This paper has two main objectives: a) introduce a natural implementation of the collective communication routines such as multicast and broadcast on bus (Ethernet) based networks of workstations, and b) discuss the implementation of other well defined collective communications routines such as scatter and gather in the same kind of environment. The (at least apparent) lack of such considerations on networks/ clusters of workstations used for parallel processing motivates this work. Moreover, the main objective of the discussion and implementation is to obtain the best possible performance from the interconnection network for parallel processes communicating in order to solve a common task.