神威·太湖之光上排列熵算法异构并行加速  

Heterogeneous parallel acceleration of permutation entropy algorithm on Shenwei Taihu Light

作　　者：周倩 梁建国 傅游 ZHOU Qian;LIANG Jian-guo;FU You(College of Computer Science and Engineering,Shandong University of Science and Technology,Qingdao 266590,China)

机构地区：[1]山东科技大学计算机科学与工程学院,山东青岛266590

出　　处：《计算机工程与设计》2023年第2期400-406,共7页Computer Engineering and Design

基　　金：国家重点研发计划基金项目(2017YFB0202002)。

摘　　要：为增加检测突变信号的时效性,提高各种故障检测系统的效率,提出在“神威·太湖之光”上基于两级并行模式改进的排列熵(PE)算法。在节点间采用MPI(信息传递接口)并行编程模型,通过对等模式轮询调度方式解决多文件负载不均衡问题;在核组内采用Athread(加速线程库)并行编程模型,通过相空间构建重构矩阵,实现从核级数据划分;采用双缓冲技术实现从核计算与访存的重叠,减少主从通信时间;利用DMA通信和重组传输数据的方法,减少主从通信次数。使用15个LDK UER204滚动轴承全寿命周期实验数据进行测试,结果表明,单核组性能较主核版本最高可获得11.86倍加速,128核组最高实现123.73倍的性能提升。To reduce the time consumption of detecting abrupt signals and improve the efficiency of various fault detection systems, an improved permutation entropy(PE) algorithm based on the two-level parallel mode on Shenwei·Taihu Light was proposed. MPI(information transfer interface) parallel programming model was adopted between nodes to solve the problem of multi file load imbalance through peer-to-peer polling scheduling. In the core group, the Athread(accelerated thread library) parallel programming model was used to construct the reconstruction matrix through the phase space to realize the data division from the core level. The double buffer technology was used to realize the overlap of slave core computing and memory access, which reduced the master-slave communication time. DMA communication and data reorganization were used to reduce the number of master-slave communication. The test results show that the performance of single core group can be accelerated by 11.86 times compared with that of the main core version, and the performance of 128 core group can be improved by 123.73 times.

关 键 词：SW26010处理器 信息传递接口(MPI) 加速线程库(Athread) 负载均衡 双缓冲 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]