An Overview of the Morfeus Project Damian Rouson Sandia National - - PowerPoint PPT Presentation

▶
an overview of the morfeus project
SMART_READER_LITE
LIVE PREVIEW

An Overview of the Morfeus Project Damian Rouson Sandia National - - PowerPoint PPT Presentation

Mar 26, 2023 291 likes •589 views

An Overview of the Morfeus Project Damian Rouson Sandia National Laboratories Joel Koplik, Karla Morris, Xiaofeng Xu City University of New York Sponsors: U.S. Office of Naval Research U.S. Department of Energy (OASCR) Outline

slide-1
SLIDE 1

An Overview of the Morfeus Project

Damian Rouson

Sandia National Laboratories

Joel Koplik, Karla Morris, Xiaofeng Xu

City University of New York Sponsors: U.S. Office of Naval Research U.S. Department of Energy (OASCR)

slide-2
SLIDE 2

Outline

 Introduction  Methodology  Results  Conclusion

slide-3
SLIDE 3

Outline

 Introduction ➢Motivation ➢Objectives ➢Previous work  Methodology  Results  Conclusion

slide-4
SLIDE 4

Motivation

 Published OOD patterns in scientific

programming are rare – especially in Fortran 2003/2008.

 Conventional wisdom suggests that

the higher-level abstractions characteristic of OOP impose performance penalties.

slide-5
SLIDE 5

Objectives

 To move scientific programmers to

higher-level, platform-agnostic yet scalable abstractions.

 To demonstrate general OOD

patterns & distill new domain-specific patterns from multiphysics applications in Fortran.

 To construct an open-source

framework that encourages the use

  • f the demonstrated patterns.
slide-6
SLIDE 6

Previous Work: Patterns

Building architecture: Alexander et al.

  • Vol. III (1975)
  • Vol. II (1977)
  • Vol. I (1979)

Positive Outdoor Space Julian Street Inn Shelter for the Homeless San Jose, CA

slide-7
SLIDE 7

Previous Work: OOD Patterns

Software architecture: Gamma et al. (1995) Scientific software architecture: Gardner & Manduchi (2007)

1995

Mediator

2007

slide-8
SLIDE 8

Previous Work: Patterns in Fortran

 F95: Decyk & Gardner (2006-'07)  F03: Markus (2008)  F03: Rouson, Adalsteinsson & Xia (2010)

Scientific Software Design: The Object-Oriented Way

Damian W. I. Rouson Jim Xia Xiaofeng Xu

2011

Puppeteer: Climate

Atmosphere Ice Ocean

aggregates

slide-9
SLIDE 9

Outline

 Introduction  Methodology ➢ Language selection ➢ Pattern definition  Results  Conclusion

slide-10
SLIDE 10

Why Fortran?

 Scientific programmers rule!  Mathematical expressiveness: ➢User-defined operators:

.div., .grad., .curl., etc.

➢Multidimensional arrays ➢Array operations  Platform-agnostic parallelism (Fortran 2008)  Automatic memory management: ➢Automatic re-sizing of arrays ➢Automatic destruction of dynamically

allocated arrays/objects.

slide-11
SLIDE 11

What's a pattern?

 A common solution to a recurring

OOD problem.

 Four essential elements: ➢The name ➢The problem ➢The solution ➢The consequences  Additional elements: ➢Also known as... ➢Known uses. ➢Related patterns. ➢Sample implementation.

slide-12
SLIDE 12

Multiphysics with Morfeus

Quantum vortices (red) & classical vortices (blue) in superfluid helium. [Morr is et al., PRL 2008] Optical turbulence in the ABL. [Morris et al., JoT sub. 2010] Particles in liquid metal MHD (red=fastest, blue=slowest) [Rouson et al., PoF 2008] Lattice Boltzman blood flow with & without stent. [Xu & Lee IJNMF 2008]

X 180 190 200 210 220 70 80 90 t=0.4T
slide-13
SLIDE 13

Software Stack

ForTrilinos CTrilinos Trilinos OO Interface Morfeus Application F2003 Interface C Headers Extern “C”{} Abstract class framework. Concrete classes.

} Procedural bindings

Distributed C++ objects

slide-14
SLIDE 14

Trilinos

  • Heroux et al. (1995) “An Overview of the Trilinos Project,”

ACM TOMS.

  • Over 50 packages: linear, nonlinear, & eigensolvers;
  • ptimization; automatic differentiation; load balancing,...
  • Establishes consistent, professional software engineering

practices for over 50 packages:

  • Automated nightly multi-platform build (CMake)
  • Automated test/notification/dashboard (CDash)
  • State-of-the-art repository (Git)
  • Mailing lists (Mailman)
  • Web-based issue tracking (Bugzilla)
  • Automated documentation (Doxygen)
slide-15
SLIDE 15

Outline

 Introduction  Methodology  Results ➢Abstract Calculus ➢Abstract Factory & Factory Method ➢Object & Surrogate  Conclusion

slide-16
SLIDE 16

Abstract Calculus Pattern

Blackboard abstraction

ut=νu xx−uux

Software abstraction class(field),pointer::u,du_dt

un1=unut

nt

u = u x,t 

u x =0,t =u0

ut≡∂u/∂t

call u%boundary(x,0,u0) u%t() u = u + u%t()*dt du_dt = nu*u%xx() – u*u%x() “Software abstractions should resemble blackboard abstractions.” Kevin Long, Texas Tech. U.

slide-17
SLIDE 17

Scalability

du_dt = nu*u%xx() – u*u%x() Purely functional

  • perators and methods.

⇒ Highly asynchronous. ⇒ ⇒ Blurs the boundary

between task & data parallelism.

⇒

Synchronization

slide-18
SLIDE 18

Factory Patterns

The problem:

➢GoF Philosophy:

“Program to an interface, not an implementation.”

➢Abstract classes model interfaces. ➢Opposing force:

Abstract classes cannot be instantiated.

slide-19
SLIDE 19

Factory Patterns

 The Abstract Factory solution: ➢“Provide an interface for creating

families of related or dependent

  • bjects without specifying their

concrete classes.” GoF

 The Factory Method solution: ➢“Define an interface for creating an

  • bject, but let subclasses decide

which class to instantiate. Factory method lets a class defer instantiation to subclasses.” GoF

slide-20
SLIDE 20

Class Diagram

Field Periodic6thFactory

returns constructs

Periodic6thOrder Legend

“Implements” Public Abstract Type deferred_binding()

FieldFactory + create() + create() +

slide-21
SLIDE 21

Trilinos-Based Client

program main !... #ifdef HAVE_MPI type(Epetra_MpiComm) :: comm #else type(Epetra_SerialComm) :: comm #endif class(field), pointer :: u class(field), pointer :: u class(field_factory), allocatable :: field_creator allocate(periodic_6th_factory::field_creator)

slide-22
SLIDE 22

#ifdef HAVE_MPI call MPI_INIT(ierr) comm = Epetra_MpiComm(MPI_COMM_WORLD) #else comm = Epetra_SerialComm() #endif initial => u_initial u => field_creator%create(initial,resolution) !.. do while t < t_final dt = u%euler_step(nu,grid_resolution) u = u + (nu*u%xx()*nu - u*u%x())*dt t = t + dt end do #ifdef HAVE_MPI call MPI_FINALIZE(rc) #endif call u%force_finalize end program

Trilinos-Based Client

slide-23
SLIDE 23

Object Pattern

The problem:

 Robust design requires a degree of

uniformity:

➢Consistent desirable behavior, e.g.

leak-free memory management.

➢Universal referencing.  Opposing force: ➢Too much uniformity feels like

handcuffs.

slide-24
SLIDE 24

Object Pattern

 The solution: ➢Define a universal parent class

hierarchy from which every class in the package inherits the desired behavior/interface.

➢Use the Object class hierachy to

ensure low-level services, e.g., those likely to be considered as candidates for future inclusion in the language.

slide-25
SLIDE 25

implements

Class Model

Hermetic + cpp_delete() Universal Ref_Counter Multivector Vector Field

extends extends extends aggregates aggregates

Object Pattern Surrogate Pattern

slide-26
SLIDE 26

Outline

 Introduction  Methodology  Results  Conclusion ➢Summary ➢Future work ➢Acknowledgements

slide-27
SLIDE 27

Summary

 Abstract Calculus illuminates a path

toward highly asynchronous computing that blurs the task/data parallel distinction

 Fortran 2003 appears to have the

expressiveness to support the general GoF design patterns in multiphysics applications.

 Several domain-specific and

language-specific patterns emerge along the way.

slide-28
SLIDE 28

Future Work

 Implement an Abstract Calculus

based on Fortran 2008 coarray PGAS

 Unit testing  Performance testing  Public release: ➢http://trilinos.sandia.gov  Demonstrate a scalable OO

multiphysics solver in Fortran 2003/2008.

slide-29
SLIDE 29

Acknowledgements

 Sandia National Laboratories: ➢Mike Heroux, Nicole Lemaster  IBM: ➢Jim Xia