First principles modeling with Octopus: massive parallelization towards petaflop computing and more A. Castro, J. Alberdi and A. Rubio.

First principles modeling with Octopus: massive parallelization towards petaflop computing and more

A. Castro, J. Alberdi and A. Rubio

2

Outline

Theoretical SpectroscopyThe octopus codeParallelization

3

Outline

Theoretical SpectroscopyThe octopus codeParallelization

4

Theoretical Spectroscopy

5

Theoretical Spectroscopy

Electronic excitations:~Optical absorption~Electron energy

loss~Inelastic X-ray

scattering

~Photoemission~Inverse

photoemission~…

6

Theoretical Spectroscopy

Goal: First principles (from electronic structure) theoretical description of the various spectroscopies (“theoretical beamlines”):

7

Theoretical Spectroscopy

Role: interpretation of (complex) experimental findings

8

Theoretical Spectroscopy

Role: interpretation of (complex) experimental findings

Theoretical atomistic structures, and corresponding TEM images.

9

Theoretical Spectroscopy

10

Theoretical Spectroscopy

11

Theoretical Spectroscopy

The European Theoretical Spectroscopy Facility (ETSF)

12

Theoretical SpectroscopyThe European Theoretical Spectroscopy

Facility (ETSF)

~ Networking~ Integration of tools (formalism,

software)~ Maintenance of tools~ Support, service, formation

13

Theoretical Spectroscopy

The octopus code is a member of a family of free software codes developed, to a large extent, within the ETSF:~abinit~octopus~dp

14

Outline

Theoretical SpectroscopyThe octopus codeParallelization

15

The octopus code

Targets:~Optical absorption spectra of molecules,

clusters, nanostructures, solids.~Response to lasers (non-perturbative response

to high-intensity fields)~Dichroic spectra, and other mixed (electric-

magnetic responses)~Adiabatic and non-adiabatic Molecular

Dynamics (for, e.g. infrared and vibrational spectra, or photochemical reactions).

~Quantum Optimal Control Theory for molecular processes.

16

The octopus code

Physical approximations and techniques:~Density-Functional Theory, Time-

Dependent Density-Functional Theory to describe the electron structure.• Comprehensive set of functionals through the

libxc library.

~Mixed quantum-classical systems.~Both real-time and frequency domain

response (“Casida” and “Sternheimer” formulations).

17

The octopus code

Numerics:~Basic representation:

real space grid.~Usually regular and

rectangular, occasionally curvilinear.

~Plane waves for some procedures (especially for periodic systems)

~Atomic orbitals for some procedures

18

The octopus code

Derivative in a point: sum over neighbor points.Cij depend on the points used: the stencil.More points -> more precision.Semi-local operation.

19

The octopus code

The key equations~Ground-state DFT: Kohn-Sham

equations.

~Time-dependent DFT: time-dependent KS eqs:

20

The octopus code

Key numerical operations:~Linear systems with sparse matrices.~Eigenvalue systems with sparse matrices.~Non-linear eigenvalue systems.~Propagation of “Schrödinger-like”

equations.

~The dimension can go up to 10 million points.

~The storage needs can go up to 10 Gb.

21

The octopus code

Use of libraries:~BLAS, LAPACK~GNU GSL mathematical library.~FFTW~NetCDF~ETSF input/output library~Libxc exchange and correlation library~Other optional libraries.

22

www.tddft.org/programs/octopus/

23

Outline

Theoretical SpectroscopyThe octopus codeParallelization

24

Objective

Reach petaflops computing, with a scientific codeSimulate photosynthesis of the light in chlorophyll

29

Multi levelparallelization

MPIKohn­Sham­states

Real­space­domains

In NodeOpenMP­threads OpenCL­

tasksVectorization

CPU GPU

30

Target systems:Massive number of execution units ~Multi core

processors with vectorial FPUs

~IBM Blue Gene architecture

~Graphical processing units

31

High Level Parallelization

MPI parallelization

32

Parallelization by states/orbitals

Assign each processor a group of statesTime propagation is independent for each stateLittle communication requiredLimited by the number of states in the system

33

Domain parallelization

Assign each processor a set of grid pointsPartition libraries: Zoltan or Metis

34

Main operations in domain parallelization

Low level paralelization and vectorization

OpenMP andGPU

36

Two approaches

OpenMPThread programming based on compiler directivesIn node parallelizationLittle memory overhead compared to MPIScaling limited by memory bandwidthMultithreaded Blas and Lapack

OpenCLHundreds of execution unitsHigh memory bandwidth but with long latencyBehaves like a vector processor (length > 16)Separated memory: copy from/to main memory

37

Supercomputers

Corvo cluster~X86_64

VARGAS (in IDRIS)~Power6~67 teraflops

MareNostrum~PowerPC 970~94 teraflops

Jugene (image)~1 petaflops

38

Test Results

39

Laplacian operator

Comparison in performance of the finitedifference Laplacian operator

CPU uses 4 threadsGPU is 4 times fasterCache effects are visible

40

Timepropagation

Comparison in performance for a timepropagation

Fullerene moleculeThe GPU is 3 times fasterLimited by copying and non GPU code

41

Multi level parallelization

Clorophyll molecule: 650 atomsJugene Blue Gene/PSustained throughput: > 6.5 teraflopsPeak throughput: 55 teraflops

Scaling

43

Scaling (II)

Comparison of two atomic system in Jugene

44

Target system

Jugene all nodes~294 912 processor cores = 73 728

nodes~Maximum theoretical performance of

1002 MFlops

5879 atoms chlorophyll system~Complete molecule of spinach

45

Tests systems

Smaller molecules~180 atoms~441 atoms~650 atoms~1365 atoms

Partition of machines~Jugene and Corvo

46

Profiling

Profiled within the codeProfiled with Paraver tool~www.bsc.es/paraver

1 TD iteration

Some “inner” iterations

One “inner” iterationIreceive Isend Iwait

Poisson solver

2­xAlltoallAllgather Allgather Scatter

51

ImprovementsMemory improvements in GS~Split the memory among the nodes~Use of ScaLAPACK

Improvements in the Poisson solver for TD~Pipeline execution • Execute Poisson while continues with an

approximation

~Use new algorithms like FFM~Use of parallel FFTs

52

Conclusions

Kohn Sham scheme is inherently parallelIt can be exploited for parallelization and vectorizationSuited to current and future computer architecturesTheoretical improvements for large system modeling