Results of the summarizing process for the arXiv paper: 1911.06817v2

In this paper, the authors present the ExaHyPE engine, which facilitates the collaboration of interdisciplinary teams in developing high-performance PDE solvers. The development of such solvers requires expertise in application domain, numerical scheme and low-level optimization. The ExaHyPE engine isolates three roles - application, algorithms and optimization expert - allowing team members to focus on their own areas of expertise while integrating their contributions into an HPC production code. Inspired by web application development practices, ExaHyPE utilizes two custom code generation modules: the Toolkit and the Kernel Generator. These modules follow a Model-View-Controller architectural pattern on top of the Jinja2 template engine library. By using Jinja2's templates to abstract critical components of the engine and generated glue code, the application development is isolated from the engine itself. The template language also enables the definition and use of custom template macros that isolate low-level optimizations from the numerical scheme described in the templates. This flexibility allows for easy expansion of solver schemes to support novel demands from applications, addition of optimized algorithmic schemes (with reduced memory footprint), or provision of improved low-level SIMD vectorization support. The authors present three use cases that showcase how the design of the code generation modules enables easy expansion and refinement of solver schemes. Each use case focuses on one user role and demonstrates how new demands from applications can be supported, optimized algorithmic schemes can be added or improved low-level SIMD vectorization support can be provided. Furthermore, as part of their research endeavours progress in phases: from initial attempts to implement desired PDE models in the engine to application-oriented benchmark setups for validating numerical schemes and finally large-scale demonstrator scenarios that establish viability tackling grand challenges. The authors also highlight that successful demonstrators should be further developed into production codes or services. In addition to designing more complex applications there are challenges posed by upcoming exascale architectures; thus requiring ExaHyPE to consider architecture specific optimizations to ensure efficient performance on these architectures.

• - ExaHyPE engine facilitates collaboration of interdisciplinary teams in developing high-performance PDE solvers
• - Three roles - application, algorithms, and optimization expert - are isolated in the engine
• - Two custom code generation modules: Toolkit and Kernel Generator, following a Model-View-Controller architectural pattern
• - Jinja2 template engine library used to abstract critical components and generated glue code
• - Custom template macros enable isolation of low-level optimizations from numerical scheme
• - Flexibility allows for easy expansion of solver schemes to support novel demands, optimized algorithmic schemes, or improved low-level SIMD vectorization support
• - Three use cases demonstrate how design enables easy expansion and refinement of solver schemes for different user roles
• - Progress in phases: initial implementation attempts, benchmark setups, large-scale demonstrator scenarios
• - Successful demonstrators should be further developed into production codes or services
• - Challenges posed by upcoming exascale architectures require architecture-specific optimizations

The ExaHyPE engine helps different teams work together to make fast computer programs that solve math problems. There are three important roles in the engine: application expert, algorithms expert, and optimization expert. The engine has two special parts that help create the programs: Toolkit and Kernel Generator. These parts use a special library called Jinja2 to make important code. The engine can also do special things to make the programs faster. It can be changed easily to support new ideas or improve how it works. The engine has been tested in different situations and it can be made even better for big computers in the future." Definitions- PDE solvers: Computer programs that solve math problems called partial differential equations. - Custom code generation modules: Special parts of a program that create new code automatically. - Model-View-Controller architectural pattern: A way of organizing a program's code into separate parts for managing data, showing information, and controlling actions. - Jinja2 template engine library: A tool that helps create code by filling in missing parts from templates. - Glue code: Code that connects different parts of a program together. - Low-level optimizations: Changes made to a program's code to make it run faster on a computer. - Numerical scheme: A way of solving math problems using numbers instead of symbols. - SIMD vectorization support: A technique used by computers to do many calculations at once using special instructions. - Solver schemes: Different ways of solving math problems using computer programs.

ExaHyPE: A Code Generation Engine for High-Performance PDE Solvers

The development of high-performance solvers for partial differential equations (PDEs) requires expertise in application domain, numerical scheme and low-level optimization. To facilitate the collaboration of interdisciplinary teams in developing such solvers, the authors present ExaHyPE, a code generation engine that isolates three roles - application, algorithms and optimization expert - allowing team members to focus on their own areas of expertise while integrating their contributions into an HPC production code.

Design Overview

Inspired by web application development practices, ExaHyPE utilizes two custom code generation modules: the Toolkit and the Kernel Generator. These modules follow a Model-View-Controller architectural pattern on top of the Jinja2 template engine library. By using Jinja2's templates to abstract critical components of the engine and generated glue code, the application development is isolated from the engine itself. The template language also enables definition and use of custom template macros that isolate low-level optimizations from the numerical scheme described in templates. This flexibility allows for easy expansion of solver schemes to support novel demands from applications, addition of optimized algorithmic schemes (with reduced memory footprint), or provision of improved low-level SIMD vectorization support.

Use Cases

The authors present three use cases that showcase how design of code generation modules enables easy expansion and refinement of solver schemes. Each use case focuses on one user role and demonstrates how new demands from applications can be supported, optimized algorithmic schemes can be added or improved low-level SIMD vectorization support can be provided. Furthermore, as part research endeavours progress in phases: from initial attempts to implement desired PDE models in engine to application oriented benchmark setups for validating numerical schemes and finally large scale demonstrator scenarios that establish viability tackling grand challenges. Authors also highlight successful demonstrators should be further developed into production codes or services.

Challenges Ahead

In addition designing more complex applications there are challenges posed by upcoming exascale architectures; thus requiring ExaHyPE consider architecture specific optimizations ensure efficient performance these architectures .