【作者】 周宇；

【导师】 何钦铭；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2019， 硕士

【摘要】 格点QCD(Lattice quantum chromodynamics)是用格点量子场论研究量子色动力学(quantum chromodynamics)的一种主流非微扰理论计算方法。它通过将场变量定义在四维网格中并进行数值模拟计算,测量由夸克与胶子构成的物理量。格点QCD数值模拟计算对网格中每个点进行迭代计算,来求解关键量:四维网格中夸克的传播振幅。格点QCD数值模拟计算属于高性能计算,其性能热点为dslash计算。四维空间中的点在dslash计算中要进行频繁的通信与计算,因此格点QCD数值模拟计算需要大量计算资源,通常在大规模超级计算机上进行。国产超算平台“神威·太湖之光”是世界上首台峰值性能超过100PFlops的超级计算机,已有许多高性能应用在神威平台上进行了移植优化并取得了很好的效果,并连续两年夺得高性能应用领域杰出奖项——戈登贝尔奖。由于神威平台片上异构设备异于常用的x86同构处理器或GPU异构加速卡,现有的格点QCD数值模拟计算程序无法直接在神威平台的从核阵列中使用。本文首次在神威平台中实现并优化了格点QCD数值模拟计算。本文重点从格点QCD数值模拟计算在“神威·太湖之光”上实现、优化两个方面展开,主要贡献有:1.针对热点计算dslash,本文提出两维展开的dslash计算模型,实现了基于神威平台的单核组方案与多核组并行方案。2.为充分利用神威平台的计算资源,在单核组方案与多核组并行方案基础上实现循环展开、计算、访存、通信优化策略。3.测试本文方案的性能,并对测试中的现象进行分析,验证本文方案的可行性与性能优势。更多还原

【Abstract】 Lattice QCD is a mainstream non-perturbation theoretical method for studying quantum chromodynamics using lattice quantum field theory.By defining the field variables in a four-dimensional grid and carrying out numerical simulation,the physical quantities consisting of quarks and gluons are measured.Lattice QCD numerical simulation calculates each point in the grid iteratively to solve the key quantity:quark propagation amplitude in the four-dimensional grid.Lattice QCD numerical simulation belongs to high performance computing,and its performance hotspot is dslash computing.Points in four-dimensional space need frequent communication and computation in dslash computing,so Lattice QCD numerical simulation needs a lot of computing resources,usually on large-scale supercomputers.Sunway TaihuLight is the first supercomputer with peak performance over 100 PFlops in the world.Many high performance applications have been transplanted and optimized on Sunway platform and achieved good results.It has won the Golden Bell Award for two consecutive years.Because the heterogeneous devices on Sunway platform are different from x86 isomorphic processors or GPU isomeric acceleration cards,the existing Lattice QCD numerical simulation program can not be used directly from the core array in Sunway platform.In this paper,the attice QCD numerical simulation is realized and optimized in Sunway platform for the first time.This paper focuses on the realization and optimization of lattice QCD numerical simulation in Sunway platform.The main contributions are as follows:1.For the calculation of dslash for hotspots,this paper proposes a two-dimensional unfolded dslash computing model,which implements a single CG scheme and a multi-CGs parallel scheme based on Sunway platform.2.In order to make full use of the computing resources of Sunway Platform,loop unrolling,computing,memory access and communication optimization strategies are implemented on the basis of a single CG scheme and a multi-CGs parallel scheme.3.Test the performance of the proposed scheme and analyze the phenomena in the test to verify the feasibility and performance advantages of the proposed scheme.更多还原