SPPEXA Doctoral Retreat 2015, Universität Passau, Christian Lengauer
Doctoral Retreat 2015
HPC Software: Synthesis, Analysis, and Optimization Hosting Projects: ExaStencils + FFMK Location: Universität Passau
14-18 September 2015
Doctoral Retreat 2015 HPC Software: Synthesis, Analysis, and - - PowerPoint PPT Presentation
Doctoral Retreat 2015 HPC Software: Synthesis, Analysis, and Optimization Hosting Projects: ExaStencils + FFMK Location: Universitt Passau 14-18 September 2015 SPPEXA Doctoral Retreat 2015, Universitt Passau, Christian Lengauer Tw o
SPPEXA Doctoral Retreat 2015, Universität Passau, Christian Lengauer
14-18 September 2015
SPPEXA Doctoral Retreat 2015, Universität Passau, Christian Lengauer
Tw o Alternative Approaches in SPPEXA
The evolutionary approach:
The revolutionary approach: Fortran MPI C OpenMP Java Threads linear transformations linear algebra stencil codes . . . HPC cluster Manycores GPGPUs FPGAs
domain-specific refinement and
SPPEXA Doctoral Retreat 2015, Universität Passau, Christian Lengauer
What Makes a Domain Suitable for the Revolutionary Approach?
Size
Considerably smaller than by contemporary expectations
Theoretical basis
Algebra
Conditional equations
Significance
Stable abstract view
Stable, sustained user community
Examples
FFTW: the fastest Fourier transform in the West
Spiral: discrete linear transforms
DBMSs: relational query optimization
cpp: Linux operating system configuration
ExaStencils: Multigrid stencil codes
SPPEXA Doctoral Retreat 2015, Universität Passau, Christian Lengauer
Our Project in the Revolutionary Tier of SPPEXA
4
Christian Lengauer, Armin Größlinger, Stefan Kronawitter Sven Apel, Alexander Grebhahn Matthias Bolten, Hannah Rittich Ulrich Rüde, Harald Köstler, Sebastian Kuckuk Jürgen Teich, Frank Hannig, Christian Schmitt
A tool-assisted, fully automated domain-specific design approach for stencil codes
http://www.exastencils.org/
SPPEXA Doctoral Retreat 2015, Universität Passau, Christian Lengauer
A: Algorithm design B: Domain-specific representation und modeling C: Domain-specific optimization and generation D: Polyhedral optimization and code generation E: Platform-specific code optimization and generation
Work Areas of ExaStencils
SPPEXA Doctoral Retreat 2015, Universität Passau, Christian Lengauer
A: Algorithm design B: Domain-specific representation and modeling C: Domain-specific optimization and generation D: Polyhedral optimization and code generation E: Platform-specific code optimization and generation
Work Areas of ExaStencils
Present scope:
Future scope:
and suitable discretizations
SPPEXA Doctoral Retreat 2015, Universität Passau, Christian Lengauer
A: Algorithm design B: Domain-specific representation and modeling C: Domain-specific optimization and generation D: Polyhedral optimization and code generation E: Platform-specific code optimization and generation
Work Areas of ExaStencils
SPPEXA Doctoral Retreat 2015, Universität Passau, Christian Lengauer
A: Algorithm design B: Domain-specific representation and modeling C: Domain-specific optimization and generation D: Polyhedral optimization and code generation E: Platform-specific code optimization and generation
Work Areas of ExaStencils Empirical prediction
performing configurations
SPPEXA Doctoral Retreat 2015, Universität Passau, Christian Lengauer
A: Algorithm design B: Domain-specific representation and modeling C: Domain-specific optimization and generation D: Polyhedral optimization and code generation E: Platform-specific code optimization and generation
Work Areas of ExaStencils
Polyhedral optimization steps:
model extraction dependence computation dead-code elimination schedule optimization tiling abstract syntax tree creation
SPPEXA Doctoral Retreat 2015, Universität Passau, Christian Lengauer
A: Algorithm design B: Domain-specific representation and modeling C: Domain-specific optimization and generation D: Polyhedral optimization and code generation E: Platform-specific code optimization and generation
Work Areas in ExaStencils
100 200 300 400 500 600 700 512 1k 2k 4k 8k 16k 32k 64k 128k 256k 448k Mean Time [ms] Number of Cores
Mean Time per V-Cycle
Pure MPI 32 MPI x 2 OMP 16 MPI x 4 OMP 8 MPI x 8 OMP 4 MPI x 16 OMP 2 MPI x 32 OMP 1 MPI x 64 OMP
SPPEXA Doctoral Retreat 2015, Universität Passau, Christian Lengauer
Goal of the Retreat: Education and Technology Transfer
Education
New w ays of abstraction
New w ays of customizing code for a problem domain
Model-oriented approaches
Multigrid technology
Technology Transfer
What of this material can you leverage in your project?
Is there potential for a cooperation in funding period 2?