Macro-Dataflow using Software Distributed Shared Memory

Abstract

Macro-dataflow processing, which exploits the parallelism among coarse-grain tasks (macrotasks) such as loops and subroutines, is considered promising to break the performance limits of loop parallelism. To realize macro-dataflow processing on distributed memory systems, "data reaching conditions", a method to make the sender-receiver pair of a data transfer determined at runtime, has previously been proposed. However, irregular data accesses induce extra data transfers, which lead to performance deterioration. This paper proposes an implementation method using software distributed shared memory, which enables on-demand data fetching. This paper describes the implementation using two well-accepted, page-based software distributed shared memory systems, TreadMarks and JI-AJIA. Evaluation results on a PC cluster show the software distributed memory approach is as much as 25% faster than the data reaching conditions