Advanced High Performance Computing CSCI 580

Transcription

1 Advanced High Performance Computing CSCI 580 2:00 pm - 3:15 pm Tue & Thu Marquez Hall 322 Timothy H. Kaiser, Ph.D. tkaiser@mines.edu CTLM 241A 1

2 Two Similar Classes this fall MATH 440/540 - PARALLEL SCIENTIFIC COMPUTING CSCI Advanced High Performance Computing 2

3 Alternative Class MATH 440/540 - PARALLEL SCIENTIFIC COMPUTING Class 2:00 pm - 2:50 pm MWF Meyer Hall 363 This course is designed to facilitate students learning of parallel programming techniques to efficiently simulate various complex processes modeled by mathematical equations using multiple and multi-core processors. Emphasis will be placed on the implementation of various scientific computing algorithms in FORTRAN/C/C++ using MPI and OpenMP. Prerequisite: MATH407, CSCI407, or consent of instructor. 3 hours lecture, 3 semester hours. 3

4 Description Advanced High Performance Computing CSCI 580 This course provides students with knowledge of the fundamental concepts of high performance computing as well as hands-on experience with the core technology in the field. The objective of this class is to understand how to achieve high performance on a wide range of computational platforms. Topics will include sequential computers including memory hierarchies, shared memory computers and multicore, distributed memory computers, graphical processing units (GPUs), cloud and grid computing, threads, OpenMP, message passing (MPI), CUDA (for GPUs), parallel file systems, and scientific applications. 3 hours lecture; 3 semester hours. 4

5 to Ganesh After our discussion this summer I thought about my content. If you look at the official descriptions of the two classes there is a lot of overlap. I have decided to put more emphasis on daily problem solving, that is after you have an application running, how do you make it work better, or the old line "My program worked last week why doesn't work today?" I will cover the stuff in the course description quickly then my goal will be to teach people how to do my job. The optimization and debugging portions of my class will hit BGQ and Phi very hard. 5

6 Teach you to do my job! 6

7 Who am I? Timothy H. Kaiser, Ph.D. 214A CTLM Director: Research and High Performance Computing Golden Energy Computing Organization BS Physics UMR MS EE/Applied Physics Ph.D. Computer Science UCSD/San Diego Supercomputer Center Defense Industry 7

8 How did I get here? My purpose on campus Yearly fall workshops This year FRCRC workshop New Machine SC 13 My purpose for this class 8

9 From our Text... Employing high performance computing (HPC) as a research tool demands at least a basic understanding of the hardware concepts and software issues involved. This is already true when only using turnkey application software, but it becomes essential if code development is required. The new PhD student is all too often left alone with the steep learning curve... Georg Hager;Gerhard Wellein. Introduction to High Performance Computing for Scientists and Engineers (Kindle Locations ). Kindle Edition. 9

10 Additional Topics Libraries for HPC, threaded and MPI based Profiling Optimization (Fix my dissertation code) Debugging Advanced batch scripting Scientific visualization Intel MIC Proposal Writing Proposal Review CSM's new machine architecture Utilities for HPC ssh - keys and usage Building Applications Some administration CS is Science 10

11 Expectations Grad level CS class with content that should be reasonable for all CSM grad students, including CS students Programming ability in C, C++, or Fortran Domain knowledge (non CS students may have advantage) Future assignments will be determined from the first assignment Your experience Simple serial program Profiling Debugging Parallelize serial program Program Install Scripting 11

12 First Assignment Send to from your CSM account Subject line CSCI 580 Assignment 1 1. Name 2. Department 3. Degree goal 4. Advisor 5. Programming languages experience level 6. HPC/Parallel programming experience 7. Unix experience level 8. Math: PDE? Linear Algebra? Numerical Methods? 9. Due Wed Aug 21, 5:00PM 12

13 Suggested Books $0 Various IBM RedBooks $0 MPI: The Complete Reference. Marc Snir, Steve Otto, Steven Huss- Lederman,David Walker, Jack Dongarra mpi-book/mpi-book.html $0 MPI: A Message-Passing Interface Standard Version 2.2. Message Passing Interface Forum September 4, docs.html $0 OpenMP Version 3.0 Complete Specifications openmp-specifications 13

14 Suggested Books $41-$47 Introduction to High Performance Computing for Scientists and Engineers. Georg Hager and Gerhard Wellein $45 Parallel Programming and Optimization with IntelR Xeon Phi Coprocessors 14

15 Other Books $27 Using OpenMP: Portable Shared Memory Parallel Programming. Barbara Chapman, Gabriele Jost, Ruud van van der Pas $63 Parallel Programming with MPI. Peter Pacheco $39 Using MPI - 2nd Edition: Portable Parallel Programming with the Message Passing Interface. William Gropp, Ewing L. Lusk, Anthony Skjellum $33 Using MPI-2: Advanced Features of the Message Passing Interface. William Gropp, Ewing L. Lusk, Rajeev Thakur $99 Introduction to Parallel Computing (2nd Edition). Ananth Grama, George Karypis, Vipin Kumar, Anshul Gupta $$$ Fortran 95/2003 for Scientists & Engineers, Stephen J. Chapman 15

16 Toys for the Year Hardware Mio - 34 Tflops BlueM AuN - 50 Tflops MC2-104 Tflops 16

17 Mio 122 Nodes 2 Racks 1285 cores and Growing 7 GPUs 8 Phi 34 Tflops It s All Mine

18 Mio Phi Enhanced Nodes 18

19 BlueM Mines` Supercomputer 154 Tflops 17.4 Tbytes 10,496 Cores 85 KW Five Racks (not full) dual architecture Two Distinct Compute Units idataplex Blue Gene Q Best of both worlds Shared 480 Tbyte File System Compact Low Power Consumption

20 BlueM s Compute Units - AuN AuN (Golden) idataplex Intel 8x2 core SandyBridge 144 Nodes 2,304 Cores 9,216 Gbytes Feature Latest Generation Intel Processors Large Memory / Node Common architecture Similar user environment to RA and Mio Quickly get researchers up and running 50 Tflops

21 BlueM s Compute Units - MC2 MC2 (Energy) Blue Gene Q PowerPC A2 17 Core 512 Nodes 8,192 Cores 8,192 Gbytes 104 Tflops Feature New Architecture Designed for large core count jobs Highly scaleable Multilevel parallelism - Direction of HPC Room to Grow Future looking machine

22 Mc2 - AuN Comparison Feature Mc2 AuN Gflop/Node Memory/Node 16 Gbytes 64 Gbytes Gflop/Gbytes Recommended Loading 16*4=64 16 Bandwidth Faster Fast 22