10- High Performance Compu5ng

Size: px

Start display at page:

Download "10- High Performance Compu5ng"

Julia Stanley
10 years ago
Views:

1 10- High Performance Compu5ng (Herramientas Computacionales Avanzadas para la Inves6gación Aplicada) Rafael Palacios, Fernando de Cuadra MRE

2 Contents Implemen8ng computa8onal tools 1. High Performance Compu8ng overview 2. Parallel Performance 3. Parallel Programming 4. Parallel Matlab 2

3 1 Overview

4 Defini5on of HPC High- Performance Compu8ng uses powerful computers to solve advanced computa8on problems Supercomputers Computer clusters Computer grids 4

5 Top 500 The fastest computers in the world (source Nov/2014: hop:// ) 5

6 Top 500 The fastest computers in the world (source Nov/2012: hop:// ) 6

7 Top 500 The fastest computers in the world (source Nov/2010: hop:// ) Tflops kw 7

8 Top 500 The fastest computers in the world (source Nov/2010: hop:// ) Tflops kw 2012 becomes 8th 2012 DOE 20x computing power ½ electric power 8

9 Top 500 (SPAIN) 2010 The fastest computers in the world (source: hop:// ) TFLOP/s 9

10 High performance in standard computers Moore's Law (1970) "The number of transistors doubles every 2 years" Source: wikipedia 10

11 High performance in standard computers Parallel processing capabili8es Mul8- core: 2 or 4 Mul8- processor: 1, 2 or 4(server) Graphical Processor Units (GPU) Source: nvidia.com 11

12 Parallel processing methods Classifica8on according to processing method and data management Single data Multiple data Single Instruction SISD SIMD or SPMD Multiple Instruction MIDS MIMD 12

13 Parallel processing methods Classifica8on according to processing method and data management Single data Multiple data Single Instruction SISD SIMD or SPMD Multiple Instruction MIDS MIMD Normal computer 13

data Multiple data Single Instruction SISD SIMD or

14 Parallel processing methods Classifica8on according to processing method and data management Single data Multiple data Single Instruction SISD SIMD or SPMD Multiple Instruction MIDS MIMD Today's video cards Simulation 14

Multiple data Single Instruction SISD SIMD or SPMD

15 Parallel processing methods Classifica8on according to processing method and data management Single data Multiple data Single Instruction SISD SIMD or SPMD Multiple Instruction MIDS MIMD No use 15

16 Parallel processing methods Classifica8on according to processing method and data management Single data Multiple data Single Instruction SISD SIMD or SPMD Multiple Instruction MIDS MIMD Multi-processors with distributed memory Grid systems 16

data Single Instruction SISD SIMD or SPMD Multiple

17 Distributed vs Parallel Distributed systems Resources are separated Could be local network or Internet BeOer reliability Parallel systems One piece Grid systems Cloud computing Amazon Elastic Compute Cloud (EC2) Several processors in one server, or several servers in a fast local network BeOer locality Multicore system Supercomputer 17

computing Amazon Elastic Compute Cloud (EC2) Several processors in one server, or

18 2 Parallel performance

19 Amdahl's Law (1967) If 1/s of a computa8on is sequen8al, then the maximum performance improvement by parallelism is at most a factor of s. T sequential = 1 s T sequential +(1 1 s )T sequential T parallel = 1 s T sequential + (1 1 s )T sequential # processors T sequential speedup = T parallel = s (For infinite #processors) 10 s 20 s 19

T sequential = 1 s T sequential +(1 1 s )T sequential T parallel = 1 s T sequential

20 Amdahl's Law (1967) 20

21 Real- life performance Again, It's all about locality! AMD Quad-core Core 1 Core 2 Core 3 Core 4 Latency Transfer- rate L1 L1 L1 L1 L2 L2 L2 L2 L3 0.5 ns 1 ns 5 ns 6000 GB/s 1000 GB/s DDR memory ns local remote 50 GB/s 21

22 Real- life performance Memory access is a booleneck In the old days: Read memory: 1 CPU cycle FLOP: several cycles Today's computers Read memory: up to 1000 cycles (due to DRAM latency compared with CPU speed) FLOP: one cycle 22

23 3 Parallel programming

24 Parallel Programming Typical parallel programming languages are: MPI (Message Passing Interface) Distributed memory virtual machine Evolu8on of PVM (parallel virtual machine) OpenMP (Open Mul8- Processing) Similar to MP Available for mul8ple standard programming languages CUDA (Compute Unified Device Architecture) Library for NVIDIA graphic cards Available for mul8ple standard programming languages 24

25 Cloud compu5ng Advantages of cloud compu8ng: You don t need to maintain hardware infrastructure Cost depends on CPU usage Reliable, Scalable, Easy to start, Easy to use Cloud compu5ng providers Amazon Elas8c Compute Cloud Microsoi Azure Salesforce Google IBM 25

26 4 Parallel Matlab

27 Parallel Matlab There are three main methods to implement parallel applica8ons in Matlab: Matlab official toolboxes pmatlab StarP 27

28 Parallel Matlab: Toolboxes Parallel Compu8ng Toolbox up to 8 cores in a single computer compa8ble with DC server MATLAB Distributed Compu8ng Server For clusters and grid systems Installs on the supercomputer Is accessed by clients using Parallel Compu8ng Toolbox Support for Amazon's EC2 hop:// 28

29 Parallel Matlab: Toolboxes Parallel Compu8ng Toolbox example matlabpool(4) % 4 cores (5 Matlab processes can be seen) sz= ; %16MB vector v=rand(sz,2); res=zeros(sz,2); %for memory allocation tic parfor i=1:2 res(:,i)=sort(v(:,i)); %each iteration independent of the previous end toc matlabpool close %back to normal (when you are done) Warning: Memory alloca8on and data transfer may be more 8me consuming that the computa8on itself. 29

30 Parallel Matlab: pmatlab Developed at Lincoln Laboratory MIT student Jeremy Kepner Based on MPI Freeware Example: unzip pmatlab cd pmatlab startup.m cd examples cd mandelbrot run pmandelbrot: eval(prun('pmandelbrot',1{})); %1 core eval(prun('pmandelbrot',4{})); %4 cores 30

31 Parallel Matlab: starp Set of extensions for parallel programming in Matlab Now belongs to Microsoi 31

32 Instituto de Investigación Tecnológica C/ Santa Cruz de Marcenado, nº Madrid Tel Fax [email protected]

8- Management of large data volumes

8- Management of large data volumes (Herramientas Computacionales Avanzadas para la Inves6gación Aplicada) Rafael Palacios, Jaime Boal Contents Storage systems 1. Introduc;on 2. Database descrip;on 3. Database management systems 4. SQL 5.