2017-c-summer

• 蘇逸鎮（中央大學數學所）

Part C. Advanced Course on Multi-Threaded Parallel Programming using OpenMP/OpenACC for Multicore/Manycore Systems

課程內容 (Schedule and Contents)

• July 4, 2017 (T)
  • Introduction
  • Finite-Volume Method (FVM)
  • Login to Reedbush System
  • OpenMP
  • Reordering (1/2)
• July 5, 2017 (W)
  • Reordering (2/2)
  • Parallel FVM by OpenMP
• July 6, 2017 (Th)
  • Introduction to GPU Programming
  • OpenACC (1/2)
• July 7, 2017 (F)
  • OpenACC (2/2)
  • Parallel FVM by OpenACC
  • Exercises

Overview

In order to make full use of modern supercomputer systems with multicore/manycore architectures, hybrid parallel programming with message-passing and multithreading is essential. While MPI is widely used for message-passing, OpenMP for CPU and OpenACC for GPU are the most popular ways for multithreading on multicore/manycore clusters. In this 4-day course, we focus on optimization of single node performance using OpenMP and OpenACC for CPU and GPU. We “parallelize” a finite-volume method (FVM) code with Krylov iterative solvers for Poisson’s equation on Reedbush supercomputer at the University of Tokyo with 1.93 PF peak performance (http://www.cc.u-tokyo.ac.jp/system/reedbush/index-e.html), which consists of the most recent CPU’s (Intel Xeon E5-2695 v4 (Broadwell-EP)) and GPU’s (NDIVIA Tesla P100 (Pascal)).

It is assumed that all the participants have already finished “邁向 Petascale 高速計算 (February 2016)” or “國家理論中心數學組「高效能計算 」短期課程 (February 2017)”, or they have equivalent knowledge and experiences in numerical analysis and parallel programming by MPI and OpenMP.

In the winter schools in February 2016 and February 2017, the target application was a 3D FVM code for Poisson’s equation by Conjugate Gradient (CG) iterative method with very simple Point Jacobi preconditioner. This time our target is same FVM code, but linear equations are solved by ICCG (CG iterative method with Incomplete Cholesky preconditioning), which is more complicated, powerful and widely used in practical applications. Because ICCG includes “data dependency”, where writing/reading data to/from memory could occur simultaneously, parallelization using OpenMP/OpenACC is not straight forward. We need certain kind of reordering in order to extract parallelism. In this 4-day course, lectures and exercise on the following issues will be provided:

• Overview of Finite-Volume Method (FVM)
• Krylov Iterative Method, Preconditioning
• Implementation of the Program
• Introduction to OpenMP/OpenACC
• Reordering/Coloring Method
• Parallel FVM by OpenMP/OpenACC

預備知識 (Prerequisites)

• Completion of one of the following two short courses:
  • 邁向 Petascale 高速計算 (February 2016) https://sites.google.com/site/school4scicomp/previous/2016-a-winter
  • 國家理論中心數學組「高效能計算 」短期課程 (February 2017) https://sites.google.com/site/school4scicomp/previous/2017-b-spring
  • (or equivalent knowledge and experiences in numerical analysis (FVM, preconditioned iterative solvers) and parallel programming using OpenMP and MPI )
  • 6月28日課程將會幫同學複習這部分的內容
• Experiences in Unix/Linux (vi or emacs)
• Fundamental numerical algorithms (Gaussian Elimination, LU Factorization, Jacobi/Gauss-Seidel/SOR Iterative Solvers, Conjugate Gradient Method (CG))
• Experiences in SSH Public Key Authentication Method

課程講義 (Course Handouts)

• http://nkl.cc.u-tokyo.ac.jp/NTU2017S/
• http://nkl.cc.u-tokyo.ac.jp/17s/ (Lectures at the University of Tokyo (OpenMP only))
• http://nkl.cc.u-tokyo.ac.jp/NTU2015/
• http://nkl.cc.u-tokyo.ac.jp/files/multicore-c.tar
• http://nkl.cc.u-tokyo.ac.jp/files/multicore-f.tar

課前準備 (Preparation)

• Windows
  • Cygwin with gcc/gfortran and OpenSSH
    • Please make sure to install gcc (C) or gfortran (Fortran) in “Devel”, and OpenSSH in “Net”
    • ParaView
• MacOS, UNIX/Linux
  • ParaView
• Cygwin: https://www.cygwin.com/
• ParaView: http://www.paraview.org/

授課教授 (Instructors)

• Professor Kengo Nakajima (中島 研吾 教授，東京大學 情報基盤中心 超級計算研究部門)

• Professor Tetsuya Hoshino (星野哲也 助教授，東京大學 情報基盤中心 超級計算研究部門)

主辦單位與聯絡人

主辦單位：國家理論中心數學組

協辦單位：東京大學情報基盤中心，臺灣大學數學系與應用數學科學研究所，台灣工業與應用數學學會 GPU與高效能計算活動學群

主持人：王偉仲 (台灣大學應數所)，林文偉 (交通大學應數系)，舒宇宸 (成功大學數學系)，黃楓南 (中央大學數學系)，黃聰明 (台灣師範大學數學系)

聯絡人：陸小姐 (02-3366-8811, risalu@ncts.ntu.edu.tw)