the amp backplane

Bill CochranOak Ridge National Laboratorycochranwk@ornl.gov

The AMP BackplaneDiscreet Management of Numerical Libraries and Multiphysics Data

The AMP BackplaneDiscreet Management of Numerical Libraries and Multiphysics Data

Oak Ridge National LabKevin ClarnoBobby PhilipBill CochranSrdjan SimunovicRahul SampathSrikanth AlluGokan YesilyurtJung Ho LeeJames BanfieldPallab Barai

Los Alamos National LabGary DiltsBogdan Mihaila

Idaho National LabRichard MartineauGlen HansenSamet KadiogluRay Berry

Oak Ridge National LabSreekanth PannalaPhani NukalaLarry OttJay Billings

Los AlamosNational LabMike Rogers

Argonne National LabAbdellatif Yacout

Los Alamos National LabCetin UnalSteven Lee

Oak Ridge National LabLarry OttJohn Turner

Argonne National LabMarius Stan

Developers: Collaborators: Advisors:

Epetra_Vector x;Vec y;N_Vector z;

VecAXPBYPCZ ( z , alpha , 1 , 0 , x , y );

The AMP Backplane

Vectors

Why So Many Libraries?AMP uses:

Moertel and ML

SNES and KSPIDA

ContactPreconditioningJFNKTime integration

Epetra_CrsMatrix P;Mat A;N_Vector x, y, z;

P.Multiply ( false , x , y );MatMult ( A , y , z );

Vec multiPhysicsSolution;Vec tempPellet, displacementPellet;Vec thermoMechanicsPellet;

SolveThermoMechanics ( thermoMechanicsPellet );

Epetra_Vector x;Vec y;N_Vector z;

VecAXPBYPCZ ( z , alpha , 1 , 0 , x , y );

The AMP Backplane

VectorsMatricesMeshes

MechanicsTemperature

Oxygen DiffusionBurn UpNeutronicsEtc.stk::mesh::Entity curElement;

libMesh::FE integrator;

integrator.reinit ( &curElement );

How Does It Work?Virtual methodsPolymorphismTemplatesIteratorsStandard template library

TheLessYouKnow

How Do I Use It?Master six classes

AMP::Vector

AMP::Matrix

AMP::MeshManager

AMP::MeshManager::Adapter

AMP::Variable

AMP::DOFMap

Linear combinations,Norms,Get/Setetc.

Matrix-Vector productsScalingetc.

Multiple domainsParallel managementI/OSpace allocationetc.

Entity iterationBoundary conditionsMemory managementVector indexingetc.

Mapping meshentities to indices invectors and matrices

Describe desiredmemory layoutIndex individual physics

In ParallelHow Do I Use It?

Step 1: makeConsistent() Step 2: ??? Step 3: Profit!

MulticoreMulti-multicore

AMP::Vector::shared_ptr thermalResidual;AMP::Vector::shared_ptr thermalSolution;

thermalResidual = residual->subsetVectorForVariable ( temperatureVar );thermalSolution = solution->subsetVectorForVariable ( temperatureVar );

AMP::Vector::shared_ptr epetraView; epetraView = AMP::EpetraVector::view ( vector );

Epetra_Vector &epetraVec = epetraView->castTo<AMP::EpetraVector>().getEpetra_Vector();

AMP::Vector::shared_ptr sundialsView; sundialsView = AMP::SundialsVector::view ( vector );

N_Vector sundialsVec;sundialsVec = sundialsView->castTo<AMP::SundialsVector>().getNVector();

How Discreet Is It?AMP::Vector::shared_ptr petscView; petscView = AMP::PetscVector::view ( vector );

Vec petscVec;petscVec = petscView->castTo<AMP::PetscVector>().getVec();

-Most vector functionality-Enough matrix functionality-Works with SNES and KSP

-Most vector functionality-Works with IDA-Single domain/single physics-Default linear algebra engine-Hopefully, limitation eased by Tpetra

-Variables describe- Memory layout- Physics- Discretization

What About Performance?1) C++2) Virtual methods Clever compiler

optimizations

Iterative access: L2Norm(), dot(), min(), axpy(), scale(), …

Non-iterative access:for ( i = 0 ; i != numElems ; i++ ) for ( j = 0 ; j != 8 ; j++ ) vector->addValue ( elem[8*i+j] , phi );

FORTRAN-esque speed

DigressionTime to perform dot product 2 vectors: Virtual method penalty: Time to perform tight loop virtual methoddot product: Dot product # floating point ops: Dot product FLOPS (FORTRAN style): Similar sized matvec w.r.t. FLOPS:

matvec cache penalty:

0.05 secs50%

0.075 secs2n-140n-2024n-1240%

What About Performance?1) C++2) Virtual methods Clever compiler

optimizations

Iterative access:

Non-iterative access:for ( i = 0 ; i != numElems ; i++ ) for ( j = 0 ; j != 8 ; j++ ) vector->addValue ( elem[8*i+j] , phi[j] );

FORTRAN-esque speed

for ( i = 0 ; i != numElems ; i++ ) vector->addValues ( 8 , elem + 8*i , phi );

Does it work?100,000+ unit tests:

AMP interfaceAMP interface vs PETScAMP interface vs EpetraPETSc wrappersSUNDIALS wrappersEpetra vs PETScVarious bugs found in

development

Single physics, single domainMultiple physics, single domainSingle physics, multiple domainsMultiple physics, multiple domainsMultiple linear algebra engines

ViewsClonesClones of viewsViews of clones

AMP vectorsPETSc viewsSundials views

SerialParallel

SUNDIALS IDA time integration

PETSc SNES JFNKquasi-static

Trilinos ML preconditioning

What Can It Do?

What Can It Do?Ti

me (s)

Number of cores

10.75k Elements/core

21.5k Elements/core

43k Elements/core

Reading Meshes

128 domains(88M Elements)

32 domains(22M Elements)

Superscaling

What Can It Do?Multiphysic

sMultidomainMulticore

What’s On The Horizon?“PMPIO” check pointing and

restartRudimentary contact searchOn-the-fly d.o.f. extraction

Hot swap linear algebra engines

Better interface for multi* data

What’s Left To Do?Performance testing and

tuningMore librariesGeneralized discretizationsBringing everything together