Evaluation of Graph Representations with Active Nodes

In recent years, the reduction scheme is widely recognized as a good implementing method of functional languages. In most parallel reduction methods, it is required to inspect and update a program graph in a common store. However, this results in memory conflictions and makes a trouble in constructing a correct reducer. This paper presents a method for reducing a graph by each active node. In our method, the graph is divided and distributed to each local memory of a processing unit, and is reduced by a process assigned to each node. Since reduction is performed in each node, parallel reduction is realized without a common store.