Parallel Libraries

1
Parallel Libraries
2
Talks Schedule

• Talks Next Week sign up for 15min speaking
  slot.
• Email me talk before speaking, Ill put it on my
  U-drive and you can use instructors station.
• Turn in report by email, no later than Friday
  April 25.

3
Talk Planning

• We will need to stay on schedule so plan for 12
  minute talk 3 minutes for questions.
• Include in slides
• Description of the problem
• Description of algorithm used
• Parallelization approach
• Results
• Any interesting issues or results
• Save details for your project report.
• Practice your talk!

4
Parallel Libraries Motivation

• Writing a large parallel scientific simulation
  code from scratch is a big undertaking.
• Mathematical Model
• Grid generation
• Discretization
• Algebraic Solvers
• linear, nonlinear, eigenvalue
• Visualization

5
From PDE simulation slides
Physical Problem
PDE Model
Grid Generation
Discrete Model
Discrete Solver
Validate Results
Visualize Results
6
Parallel Libraries Motivation

• Many of the components in a simulation code may
  be general purpose
• Solving a matrix equation Axb may be a part of a
  groundwater flow simulation and also a part of a
  astrophysics simulation. They may be able to use
  the same code.
• Grid generation code for an automotive simulation
  may also be appropriate for a wind simulation
  through city.

7
Parallel Libraries Motivation

• There are sophisticated, efficient, reliable,
  parallel software packages available to handle
  some of these commonly arrising tasks
• Resources to find libraries
• Google
• Research papers and journals
• Conferences

8
Parallel Libraries Motivation

• Advantages in using libraries
• Many man-years invested in some of these
• Many are freely available
• Often quicker than writing your own, especially
  true after the first time
• Typically written by experts, and can be more
  efficient than your own code
• Allow you to focus on those areas that youre
  interested in

9
Parallel Libraries Motivation

• Possible problems in using libraries
• Most require overcoming an initial
  learning-curve to use
• Installing package
• Passing information to library (interface)
• Much less of an issue after first use
• A general purpose code may not be as efficient as
  a code you write specifically for your problem
• Little control over internals of library
• Some packages are incomplete, some are no longer
  being supported
• Libraries can have bugs, but so can your code!

10
Domain Partitioning for Graphs Metis, Chaco and
JOSTLE

• Partition goals
• Load balancing
• Minimize communication

11
Domain Partitioning for Graphs Metis, Chaco,
JOSTLE
Processor 0
Processor 1
Processor 2

• Matrix vector product (Ax) partition goals
• Load balancing near equal number of points per
  processor
• Minimize communication few edges cut by
  partition, actually, few vertexes on cut
  boundaries

12
Simple example problem steady-state temperature
profile in a rod
U0
U2
U1
U5
U6
U4
U3

• Discretizing the problem also involves converting
  the PDE

13
Simple example problem steady-state temperature
profile in a rod
U0
U2
U1
U5
U6
U4
U3

• To a set of discrete equations

14
Libraries for solving sparse matrices

• PETSC
• HYPRE
• AZTEC

15
Libraries for dense matrices

• SuperLU
• SPOOLES
• ScaLAPACK

16
More Libraries

• PETSC includes lots beyond sparse linear solvers
• SUNDIALS for nonlinear, ODE, and DAE, sensitivity
  analysis
• Optimization TAO
• Eigenvalues ARPACK
• PDE simulation OVERTURE

17
A word about using libraries

• Most library developers want people to use their
  codes, but
• Many packages are not really like commercial
  packages in that they are research efforts
• User feedback to developers is important
• Developers may have funding sources that guide
  the direction of development