一种可扩展异构并行核外高阶矩量法研究_左胜_陈岩_张玉_赵勋旺_林中朝.pdfVIP

2017 年 2 月 西安电子科技大学学报（自然科学版） Feb.2017 第 44 卷 第 1 期 JOURNAL OF XIDIAN UNIVERSITY Vol.44 No.1 doi：10.3969/j.issn.1001-2400.2017.01.026 一种可扩展异构并行核外高阶矩量法研究 左 胜，陈 岩，张 玉，赵勋旺，林中朝 (西安电子科技大学 天线与微波技术重点实验室，陕西 西安 710071） 摘要：为了实现异构并行矩量法可跨节点，研究了CPU/GPU 与CPU/MIC 分布式异构平台上矩量法并行 编程模型。利用GPU 通用编程标准CUDA 中提供的CONTEXT 技术、MIC 中环境变量的概念，提出了 一种适用于CPU/GPU 与CPU/MIC 的通用异构并行编程模型，满足了异构并行矩量法可跨节点对静态负 载均衡的要求。数值结果表明，基于该并行编程模型设计的异构并行矩量法程序，可获得理想的加速比 并具有良好的可扩展性。 关键词：矩量法；异构平台；并行编程模型；可扩展性 中图分类号：TN820 文献标识码：A 文章编号：1001-2400(2017)01-0160-07 Study of scalable heterogeneous parallel out-of-core higher order method of moments ZUO Sheng, CHEN Yan, ZHANG Yu, ZHAO Xunwang, LIN Zhongchao (National Key Laboratory of Antennas and Microwaves Technology, Xidian Univ., Xi’an 710071, China) Abstract: To achieve the across-nodes technology of heterogeneous parallel method of moments (MoM), a parallel MoM programming model is studied on CPU/GPU and CPU/MIC heterogeneous platforms. By utilizing the CONTEXT technology in CUDA which is the GPU common programming criterions, and the concept of environment variables in MIC, a general heterogeneous parallel programming model for CPU/GPU and CPU/MIC is proposed, which meets the static load-balancing for the across-nodes heterogeneous parallel MoM. Numerical results show that the heterogeneous parallel MoM codes based on the proposed parallel programming model can obtain ideal speedup and good scalability. Key Words: method of moments; heterogeneous platforms; parallel programming model; scalability