Matrix multiplication is taken as a test bed for parallel processing on heterogeneous networks of workstations (local area networks) used as parallel machines. Two algorithms are proposed taking into account the specific kind of parallel hardware provided by local area networks, and experimentation is used to drive the evaluation and identification of possible performance loss. A specific broadcast communication between processes of a parallel application is also proposed, taking advantage of the Ethernet interconnection network to achieve optimized performance. A special emphasis is place on already installed networks of workstations, which provide a hardware zero cost parallel computer; but a homogeneous Beowulf-class system is used to show how the algorithms are also useful on current classical high performance parallel computing with clusters.Eje: Lenguajes