A Data Re-Distribution Library for Multi-Processor Task Programming

Abstract

Multiprocessor task (M-task) programming is a suitable parallel programming model for coding application problems with an inherent modular structure. An M-task can be executed on a group of processors of arbitrary size, concurrently to other M-tasks of the same application program. The data of a multiprocessor task program usually include composed data structures, like vectors or arrays. For distributed memory machines or cluster platforms, those composed data structures are distributed within one or more processor groups. Thus, a concise parallel programming model for M-tasks requires a standardized distributed data format for composed data structures. Additionally, functions for data re-distribution with respect to different data distributions and different processor group layouts are needed to glue program parts together. In this paper, we present a data re-distribution library which extends the M-task programming with Tlib, a library providing operations to split processor groups and to map M-tasks to processor groups.

Parallel Programming Models for Irregular Algorithms

“…The advantage of this dynamic behavior is that arbitrary, hierarchically structured and recursive M-task programs can be coded easily, providing an easy way to express divide-and-conquer algorithms or irregular algorithms. For a data distribution and a correct cooperation of arbitrary Mtasks a specific data format is needed which fits the dynamic needs of the model [40].…”