| Abstract: | The tuple space abstraction is a powerful and general coordination model for parallel and distributed programming. However, this model is based on the abstraction of a global space, difficult to implement in distributed memory parallel systems with high-performance and scalability. The paper discusses the possible distribution strategies to implement a tuple space server and proposes a new replication policy suitable for massively parallel systems. The proposed strategy achieves scalability by organising the system in a hierarchical way and by integrating coherence protocols of limited overhead. Moreover, the chosen organisation encourages the presence of multiple tuple spaces each with a constrained scope. The paper describes and analyses a transputer-based implementation: the hierarchical organisation of the tuple space makes the access time proportional to the locality of the references and bounded by the logarithm of the system size. © 1998 John Wiley & Sons, Ltd. |
