Introduc)on & HPC Ivan Giro3o igiro3o@ictp.it HPC Applica?ons - - PowerPoint PPT Presentation

▶
introduc on hpc
SMART_READER_LITE
LIVE PREVIEW

Introduc)on & HPC Ivan Giro3o igiro3o@ictp.it HPC Applica?ons - - PowerPoint PPT Presentation

Aug 19, 2022 241 likes •592 views

Introduc)on & HPC Ivan Giro3o igiro3o@ictp.it HPC Applica?ons Specialist Interna?onal Centre for Theore?cal Physics (ICTP) Ivan Giro*o - igiro*o@ictp.it La?n America Introductory School to Parallel Programming 1 Abacus Cinvestav, 12

slide-1
SLIDE 1

Introduc)on & HPC

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 1

Ivan Giro3o – igiro3o@ictp.it

HPC Applica?ons Specialist Interna?onal Centre for Theore?cal Physics (ICTP)

slide-2
SLIDE 2

Mission - An ins?tute run by scien?sts for scien?sts

  • Foster the growth of advanced studies and research in physical and

mathema?cal sciences, especially in support of excellence in developing countries.

  • Develop high-level scien?fic programmes keeping in mind the needs of

developing countries, and provide an interna?onal forum of scien?fic contact for scien?sts from all countries.

  • Conduct research at the highest interna?onal standards and maintain a

conducive environment of scien?fic inquiry for the en?re ICTP community.

  • Thanks to the generous funding from the Italian Government, UNESCO and the

IAEA, ICTP has been able to ini?ate and implement various schemes of support and assistance to scien?sts from developing countries.

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 2

slide-3
SLIDE 3

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 3

slide-4
SLIDE 4

around 6000 per year permanent scien?sts

temporary scien?fic staff: post-docs, staff associates, long-term visitors short-term visitors: associates, guest scien?sts par?cipa?on in conferences, schools and workshops; diploma students

research and training coordina?on

mostly research do own or collabora?ve research, or receive training through par?cipa?on receive training through schools, conferences, and exchange of informa?on

30 100 400

ICTP from Trieste to the World

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 4

slide-5
SLIDE 5

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 5

slide-6
SLIDE 6
  • More than 140,000 visits since 1970
  • 190 countries represented
  • 20% of ICTP visi?ng scien?sts are women

MCTP Mexico ICTP-SAIFR Brazil EAIFR Rwanda ICTP-AP China ICTP-ECAR Turkey ++NEW++ The ICTP Partner Ins)tutes

slide-7
SLIDE 7

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 7

slide-8
SLIDE 8

ICTP Scien?fic Calendar

  • Schools, Conferences, Workshops around the year
  • Half of them on subjects related to main research areas

(core)

  • The rest on many subjects:

medical physics, op?cs, nano physics, plasma physics, electronics, high performance compu?ng, biophysics, satellite naviga?on, science dissemina?on and e-learning, m-science, entrepreneurship, nuclear physics (IAEA), teacher training, 3-D Prin?ng, etc...

  • h*p://www.ictp.it/scien?fic-calendar.aspx

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 8

slide-9
SLIDE 9

Scien?fic Sec?ons

Earth System Physics Mathema)cs High Energy Cosmology and Astropar)cle Physics Condensed Ma3ers and Sta)s)cal Physics Applied Physics New areas

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 9

slide-10
SLIDE 10

HPC Staff and Collaborators

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 10

  • Dr. David Grellsheid

Herwig Sogware Manager @ Durham University

  • Dr. Axel Kohlmeyer

Full Professor of Research @ Temple University

  • Dr. Clement Onime

Responsible IT/HPC Infrastructure @ ICTP me HPC Applica?on Specialist @ ICTP

slide-11
SLIDE 11

High-Performance & Scien?fic Compu?ng ac?vi?es at the ICTP

  • HPC service and HPC applica?on consul?ng

– in house HPC facility (Argo) – research enablement on massively parallel systems for HPC on both na?onal service (CINECA) and EU infrastructures (PRACE)

  • Training on HPC and Scien?fic Programming

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 11

slide-12
SLIDE 12

MHPC in pills: www.mhpc.it

  • High-level educa?onal program: not an Ms.C. program!
  • Intensive training aimed to build knowledge in solving

complex problems with an HPC approach

  • Innova?ve, hands-on based training
  • Aimed to people with strong interest in:

– advanced programming for scien?fic compu?ng – sogware op?miza?on – management of compu?ng plakorms – data management and data analy?cs

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 12

slide-13
SLIDE 13

Background Requirements

  • Candidates must have some experience in

programming and a competence in at least one

  • f the languages between C, C++ and/or Fortran

– Python knowledge is a plus

  • A sound knowledge of Linux opera?ng system
  • Master level of a scien?fic degree is required
  • No prior HPC knowledge is assumed
  • Enthusiasm is a must

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 13

slide-14
SLIDE 14
  • Scien?fic Programming Environment
  • Introduc?on to Computer

Architectures for HPC

  • Object Oriented Programming
  • Parallel Programming
  • Introduc?on to Numerical Analysis
  • Advanced Computer Architectures

and Op?miza?ons

  • Parallel Data Management and Data

Exchange

  • High Performance Compu?ng

Technology

  • Best Prac?ces in Scien?fic Compu?ng

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 14

  • Data structures, sor?ng and searching

algorithms in serial and parallel

  • Lookup tables, cell lists and neighbor

lists

  • Domain decomposi?on techniques
  • Parallel FFT techniques
  • Parallel Linear Algebra
  • Mul?pole expansion, mul?-grid

methods

  • Adap?ve Meshes
  • Maximum likelihood techniques
  • Cluster or network or graph analysis
  • Monte Carlo methods
  • Agent-based models
  • Automa?c differen?a?on
  • DFT from source to code

1 year program divided in 6-8 months courses and 6 month project (some overlap)

Mandatory Op)onal Choice

slide-15
SLIDE 15
  • Interna?onal experience
  • Parallel programming for HPC
  • HPC approach to parallel codes
  • Best prac?se experiences
  • Computer architectures for HPC

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 15

La)n American Introductory School on Parallel Programming and Parallel Architecture for High Performance Compu)ng

slide-16
SLIDE 16

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 16

  • Dr. Richard Berger

Research Assistant Professor Temple University (USA)

  • Dr. Gavin Pringle

Applica?ons Consultant EPCC @ University of of Edinburgh (UK)

  • Prof. Wolfgang Bangerth

Professor Colorado State University (USA)

  • Dr. J. Manuel Solano-Altamirano

Research Assistant Professor Benemérita Universidad Autónoma de Puebla (Mexico)

  • Dr. William Fernando Oquendo Pa?ño

Assistant Professor Universidad de La Sabana (Colombia) Jimmy Aguilar Mena PhD Student BSC-CNS (Spain) Marlon Brenes Navarro PhD Student Trinity College Dublin (Ireland)

slide-17
SLIDE 17

Why use Computers in Science?

  • Use complex theories without a closed solu?on: solve

equa?ons or problems that can only be solved numerically, i.e. by inser?ng numbers into expressions and analyzing the results

  • Do “impossible” experiments: study (virtual) experiments,

where the boundary condi?ons are inaccessible or not controllable

  • Benchmark correctness of models and theories: the be*er

a model/theory reproduces known experimental results, the be*er its predic?ons

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 17

slide-18
SLIDE 18

SW in Science

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 18

slide-19
SLIDE 19

What is High-Performance Compu?ng (HPC)?

  • Not a real defini?on, depends from the prospec?ve:

– HPC is when I care how fast I get an answer – HPC is when I foresee my problem to get bigger and bigger

  • Thus HPC can happen on:

– A worksta?on, desktop, laptop, smartphone! – A supercomputer – A Linux Cluster – A grid or a cloud – Cyberinfrastructure = any combina?on of the above

  • HPC means also High-Produc?vity Compu?ng

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 19

slide-20
SLIDE 20

Why would HPC ma*er to you?

  • Scien?fic compu?ng is becoming more important in

many research disciplines

  • Problems become more complex, thus need complex

sogware and teams of researchers with diverse exper?se working together

  • HPC hardware is more complex, applica?on

performance depends on many factors

  • Technology is also for increasing compe??veness
  • HPC knowledge is an opportunity

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 20

slide-21
SLIDE 21

More & More Compu?ng ...

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 21

slide-22
SLIDE 22

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 22

HPC INFRASTRUCTURE

HW/Resource Management/File System/... Compilers/Libraries/Debugging & Profiling

Pre-processing Precondi?oning Data Acquisi?on Computer Simula?on Post-processing Data Analy?cs Publica?on Dissemina?on Data Management

SW Workflow & Parallel Applica?ons

Scien?sts/Applica?on Developers/End Users

slide-23
SLIDE 23

Parallelism - 101

  • there are two main reasons to write a parallel

program:

  • access to larger amount of memory (aggregated, going bigger)
  • reduce ?me to solu?on (going faster)

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 23

slide-24
SLIDE 24

Programming Parallel Paradigms

  • Are the tools we use to express the parallelism

for on a given architecture

  • They differ in how programmers can manage and

define key features like:

– parallel regions – concurrency – process communica?on – synchronism

La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 24 Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018

slide-25
SLIDE 25

How do we evaluate the improvement?

  • We want es?mate the amount of the

introduced overhead => To = npesTP - TS

  • But to quan?fy the improvement we use the

term Speedup:

SP = TS TP

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 25

slide-26
SLIDE 26

Speedup

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 26

slide-27
SLIDE 27

Efficiency

  • Only embarrassing parallel algorithm can obtain an

ideal Speedup

  • The Efficiency is a measure of the frac?on of ?me for

which a processing element is usefully employed:

EP = SP p

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 27

slide-28
SLIDE 28

Efficiency

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 28

slide-29
SLIDE 29

Scalability

  • When we want consider the scalability of our

problem we are interested in two main features:

– how much faster do we go increasing the number of processes for a fixed problem size (strong scaling) – how does the applica?on behave if we increase the problem size keeping the workload fixed per processors

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 29

slide-30
SLIDE 30

*courtesy of Carlo Cavazzoni

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 30

slide-31
SLIDE 31

How do we get the profiling? /1

  • Code instrumenta?on
  • Profiling tools

double cclock() /* Return the second elapsed since Epoch (00:00:00 UTC, January 1, 1970) */ { struct timeval tmp; double sec; gettimeofday( &tmp, (struct timezone *)0 ); sec = tmp.tv_sec + ((double)tmp.tv_usec)/1000000.0; return sec; }

The GNU Profiler (GPROF)

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 31

slide-32
SLIDE 32

How do we get the profiling? /2

  • Code instrumenta?on
  • Profiling tools
slide-33
SLIDE 33

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 33

slide-34
SLIDE 34

Ivan Giro*o - igiro*o@ictp.it Abacus Cinvestav, 12 Feb 2018 La?n America Introductory School to Parallel Programming and Parallel Architecture for HPC 34