PVM computation of the transitive closure: The dependency graph approach

Abstract

We investigate experimentally a dependency graph approach to the distributed parallel computation of the generic transitive closure problem. A parallel coarse-grained algorithm is derived from a finegrained algorithm. Its advantage is that approximately half of the work is organised as totally independent computation sequences of several processes. We consider conceptual description of dependencies between operations as partial order graphs of events. Such graphs can be split into disjoint subgraphs corresponding to different phases of the computation. This approach helps to design a parallel algorithm in a way which guarantees large independence of actions.We also show that a transformation of the fine-grained algorithm into the coarse-grained is rather nontrivial, and that the straightforward approach does not work.