RIA: Lattice Parallelism: A Programming Model for Dynamically Manipulating Distributed Data Structures in Localized Non-Uniform Scientific Calculations

点击翻译按钮获取中文摘要

The goal of this research is to implement a new model for programming scientific computations on distributed memory multiprocessors, called lattice parallelism (LP). LP is intended for scientific methods that locally concentrate computational effort non-uniformly and unpredictably, for example, adaptive mesh methods. It provides high-level operations to hide the low-level details of dynamically manipulating elaborate, dynamic data structures. The objectives of the research are to gain a better understanding of lattice parallelism under realistic conditions; to apply it to scientific applications, and to explore run-time optimization techniques. The implementation and evaluation of LP will be carried out on the Intel iPSC/860, running a variety of application codes at our disposal. A novel feature of lattice parallelism is that it exploits the tightly coupled local structure present in non-uniform calculations and so avoids high communication overheads. This is significant in light of the ultimate emergence of massively parallel distributed memory multiprocessors, which are especially sensitive to the effects of communication overhead. This research will result in the development of a software tool for reducing the effort required to parallelize full scale numerical software by an order of magnitude or better.//